Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Multi-temporal UAV Imaging-Based Mapping of Chlorophyll Content in Potato Crop

Multi-temporal UAV Imaging-Based Mapping of Chlorophyll Content in Potato Crop Spectral indices based on unmanned aerial vehicle (UAV) multispectral images combined with machine learning algorithms can more effectively assess chlorophyll content in plants, which plays a crucial role in plant nutrition diagnosis, yield estima- tion and a better understanding of plant and environment interactions. Therefore, the aim of this study was to use UAV-based spectral indices deriving from UAV-based multispectral images as inputs in different machine learning models to predict canopy chlorophyll content of potato crops. The relative chlorophyll content was obtained using a SPAD chlorophyll meter. Random Forest (RF), support vector regression (SVR), partial least squares regression (PLSR) and ridge regression (RR) were employed to predict the chlorophyll content. The results showed that RF model was the best performing algorithm with an R of 0.76, Root Mean Square Error (RMSE) of 1.97. Both RF and SVR models showed much better accuracy than PLSR and RR models. This study suggests that the best models, RF model, allow to map the spatial variation in chlorophyll content of plant canopy using the UAV multispectral images at different growth stages. Keywords Chlorophyll content · Machine learning · Multispectral images · Potato · Unmanned aerial vehicle (UAV) Zusammenfassung Multitemporale UAV-gestützte Kartierung des Chlorophyllgehalts in Kartoffelkulturen. Spektrale Indizes, die auf multi- spektralen Bildern von der Unmanned Aerial Vehicles (UAVs) basieren, können in Kombination mit Algorithmen des Maschinelles Lernen (ML) den Chlorophyllgehalt von Pflanzen effektiver bewerten. Dieser spielt eine entscheidende Rolle bei der Diagnose der Pflanzenernährung, der Ertragsschätzung und einem besseren Verständnis der Wechselwirkungen zwischen Pflanze und Umwelt. Ziel dieser Studie war es daher, UAV-basierte Spektralindizes, die aus UAV-basierten Multi - spektralbildern abgeleitet wurden, als Input für verschiedene ML-Modelle zu verwenden, um den Chlorophyllgehalt von Kartoffelpflanzen im Kronendach vorherzusagen. Der relative Chlorophyllgehalt wurde mit einem SPAD-Chlorophyllmess- gerät ermittelt. Random Forest (RF), Support Vector Regression (SVR), Partial Least Squares Regression (PLSR) und Ridge Regression (RR) wurden zur Vorhersage des Chlorophyllgehalts eingesetzt. Die Ergebnisse zeigten, dass das RF-Modell mit einem R von 0,76 und einem Root Mean Square Error (RMSE) von 1,97 der beste Algorithmus war. Sowohl das RF- als auch das SVR-Modelle zeigten eine viel bessere Genauigkeit als die PLSR- und RR-Modelle. Diese Untersuchung deutet darauf hin, dass das beste Modell, das RF-Modell, es ermöglicht, die räumliche Variation des Chlorophyllgehalts in Pflanzen- dächern mit Hilfe von multispektralen UAV-Bildern in verschiedenen Wachstumsstadien abzubilden. * Fei Li feili72@163.com 1 Introduction Inner Mongolia Key Laboratory of Soil Quality The chlorophylls are the most important pigment for photo- and Nutrient Resource, College of Grassland, Resources and Environment, Inner Mongolia Agricultural University, synthesis (Gitelson et al. 1999; Singhal et al. 2019), which Hohhot 010011, China provides energy to the biosphere (Qi et al. 2021). Since the Inner Mongolia Academy of Forestry Sciences, symptoms of plant response to stress and N deficiencies Hohhot 010010, China are often associated with plant chlorophyll content (Houles Department Life Science Engineering, School of Life et al. 2007), timely determination of plant chlorophyll con- Sciences, Technical University of Munich, 85354 Freising, tent can be important indicators of plant nutrient status and Germany Vol.:(0123456789) 1 3 PFG environmental effects, which is very useful for agricultural squares regression (PLSR), ridge regularization (RR) are field management (Argenta et al. 2004). powerful tools to assist in UAV image computation. These Conventionally, the determination of chlorophyll content algorithms performed well in current applications predicting involves outdoor sampling and indoor chemical analysis. plant conditions, such as N content (Colorado et al. 2020), These methods are not only time-consuming and labori- chlorophyll content (Qi et al. 2021), and biomass (Vilijanen ous, but also destructive and lagging. As a non-destructive et al. 2018). Using spectral indices as input variables is a technique, the portable chlorophyll meters (SPAD-502) have new approach, since these models ensure good performance been widely used to estimate chlorophyll content in agricul- in spite of only a few variables as input feature (Moghimi tural studies (Markwell et al. 1995; Martínez et al. 2004). et al. 2020). Mutanga et al. (2012). found that predicting However, the SPAD meter is a leaf-based measurement, and chlorophyll content using the RF algorithm and three NDVIs infeasible for large-scale regional monitoring (Uddling et al. calculated from the red-edge and near-infrared bands pro- 2007). Therefore, there is a need to develop the techniques to duced low prediction errors. Shah et al. (2019) used the monitor plant chlorophyll content in real time on a regional established spectral indices as an input variable combined scale. with the RF algorithm to improve the accuracy of estimating In recent years, remote sensing technology allows to chlorophyll content in wheat. However, most studies have monitor plant chlorophyll content on a large regional scale only compared a single machine learning algorithm com- (Yu et al. 2014). UAV-based multispectral imagery has been bined with spectral indices to a linear regression model and widely used to monitor the nutrient status in crops (Théau did not compare the differences between different machine et al. 2020), e.g., chlorophyll content and N (Clevers et al. learning algorithms. Therefore, the main objectives of this 2017; Verrelst et al. 2012). UAV image acquisition is not study were: (1) to evaluate the performances of published only fast, less consuming and flexible (Bendig et al. 2013), spectral indices in estimating chlorophyll content of potato but also with higher spatial and temporal resolution than the plants, (2) to compare the performance of SVR, RF, PLSR satellite and aerial images to facilitate further image analysis and RR models in predicting potato chlorophyll content, (3) (Bareth et al. 2015). Although satellite and aerial images to generate chlorophyll content maps based on the optimal are able to assess crop nutrient status on a large scale, such models, and (4) to verify whether the UAV images were images possess the low spatial and temporal resolution. able to detect the variation of potato chlorophyll content at Rafael et al. (2014) have shown that there was difficulty different growth stages. of differentiating the content of chlorophyll in forest using multispectral satellite imagery. Over the past decades, how- ever, the use of UAV images has rapidly narrowed the gap 2 Materials and Methods between satellite or aerial and ground-based sensing (Prado Osco et al. 2019).2.1 Experimental Sites However, the combination of remote sensing techniques with multispectral imagery requires more reliable assess- Experiments were carried out in Zhuozi Country, which is ment techniques due to the huge amount of data generated. located in the middle of Inner Mongolia (extending from Compared to individual spectral bands, spectral indices can 110° 51′ E, 40° 38′ N to 112° 56′ E, 41 °16′ N), China (see be applied at different scales and mitigate to a certain extent Fig. 1). The major crops are potato, corn, and beans. This the adverse effects due to anisotropic reflection, background area is characterized with a middle temperate arid and semi- shadows and soil brightness contributions (Inoue et al. 2012; arid continental monsoon, cold winters and cool summers. Kooistra and Clevers 2016). For example, Qi et al. (2021) The annual average precipitation is 544.5 mm, i.e., 90–95% found NDVI and GNDVI had a much higher degree of fit between April and October. The average temperature and precision than the other indices. Tahir et  al. (2018) is > 20 °C during the potato growing season. found MSAVI2 and TNDVI were proved to be more robust indices to estimate the chlorophyll content in the orchard 2.2 Experimental Design with the highest coefficients of determination (R ) 0.89 and 0.85, respectively. Although spectral indices can improve the Two experiments involving different potato cultivars were robustness in predicting chlorophyll content to some degree, conducted at two different sites of Zhuozi county from June the saturation effect and multi-collinearity are still a problem to September in 2020 and 2021, respectively. Experiment 1 for remotely estimating chlorophyll content in crops (Cao in 2020 was a randomized complete block design with four et al. 2017; Li et al. 2010). replications. The potato cultivar was Yingniweite. There Recently, machine learning algorithms have been induced were five N treatments: four optimization of N application in remote sensing (Wu et al. 2019; Lu et al. 2021). Support management, and conventional N rate fertilization (Con). vector machine (SVM), random forests (RF), partial least The plot size was 10 × 10 m. There were six N treatments 1 3 PFG Fig. 1 Geographical location of the experimental site (a) in 2020 and site (b) in 2021 with four replications for Experiment 2 in 2021. The potato regions: green (530–570 nm), red (640–680 nm), red-edge cultivar was Mingfeng 16. Five N treatments were same as (730–740 nm), and near-infrared (770–810 nm). Because, those in experiment 1, but control (no N was applied) was during the flight, UAV images were affected by partly added. The plot size was 10 × 12 m. For both experiments, cloudy and the changing illumination in light intensity, a the irrigation method of the study site was drip irrigation, radiometric calibration white board and an onboard irradi- and the field management was unified according to the local ance sensor were integrated in the multispectral camera to large field requirements. correct images for dash area and illumination differences the multispectral camera with (Kopačková-Strnadová et al. 2.3 Field Data Acquisition 2021; Franzini et al. 2019). Additional sensors such as GPS and laser sensor are also mounted in the UAV (see A Minolta brand chlorophyll meter (Model SPAD-502; Fig. 2). The properties of the UAV camera are presented Spectrum Technologies Inc. Plainfield, IL) was used to in Table 1. measure the chlorophyll content of potato functional leaves (the 4th compound leaf from apex) before the UAV flight was carried out. Fifteen sampling points were randomly 2.5 UAV Image Processing selected from each treatment. Each measured point repre- sented the growth of a potato plant. The average of chlo- Figure 2 illustrates the flow chart of this study. Pix4Dmap- rophyll content of each plot can be precisely acquired by per Version 4.5.6 was used to import all images taken in averaging the fifteen samples of data. the same period according to the location coordinates in its properties, and to align the overlapping images using 2.4 Acquisition of UAV Images the feature point matching algorithm. Before the UAV flight, 4 ground control points were evenly arranged in During the experimental period, the multispectral images the field to obtain accurate geographical reference posi- were taken by the Parrot Bluegrass drone with Parrot tions. The plane precision of ground control points sys- Sequoia multispectral camera on 20 July, 3 August, and tem is ± (8 + 1 × 10–6 × D) mm, and the elevation accuracy 20 August in 2020, whereas the images were acquired on is ± (15 + 1 × 10–6 × D) mm. The area marked by ground 15 July, 6 August, and 25 August in 2021. UAV flight was control points is 40 cm × 40 cm. The UAV-based images conducted between 10:00 a.m. and 2:00 p.m. at 30 m high, were first aligned using Pix4Dmapper, and the camera resulting in a 10 cm ground sample distance (GSD). Mul- parameters were estimated from the coordinates of their tispectral images were taken from the following spectral 1 3 PFG Fig. 2 Unmanned aerial vehicle (UAV)-based robotic system for chlorophyll content estimation in potato Table 1 Specifications of parrot Multi-spectral camera Flight conditions sequoia camera and flight conditions Wavelength bands Specifications Green 550 (± 40) Spectral resolution 10 bits or 10 cm Height 30 m Red 660 (± 40) HFOV 70.6° Time 01:30 P.M Red edge 735 (± 40) VFOV 52.6° Weather Partially cloudy Precipitation: 0 mm Near infrared 790 (± 40) DFOV 89.6° Wind: At 1–2 m/s photographs and GCPs. A depth sharpening filter to fur - The relationship between 12 spectral indices and the ther improve images edge detail was selected, and then chlorophyll content was calculated by Pearson correlation dense point clouds were built. Finally, meshes and tex- analysis. The correlation coefficient (r) values were used to tures based on the camera’s own parameters were created evaluate the correlations between indices and chlorophyll and the processed images were exported as TIFF images. content. Their performance was further evaluated by statis- ENVI 5.1 was used to mark the region of interest (ROI) tical comparison of the returned individual spectral indices (see Fig.  1) in the experimental plot and to extract the correlation. average reflectance values in the four bands of ROI for each images. 2.7 Machine Learning Algorithms 2.6 Spectral Indices 2.7.1 Partial Least Squares Regression and Ridge Four canopy reflectance values (Green, Red, Red edge, Regression (PLSR and RR) Near-infrared) were then used to computer 12 spectral indi- ces that were evaluated with their correlations with chloro- Before the estimation model was constructed, Pearson cor- relation analyses between chlorophyll content measurements phyll content (see Table 2). The indices were enhanced by the contribution of vegetation optical properties according and 12 spectral indices were conducted. The indices with higher correlation coefficients (|r |> 0.4, and p values less to the spectral response of the canopy by combining several bands. And thus, spectral indices were able to reduce the than 0.01) were selected for machine learning algorithms. Compared to simple linear regression, PLSR and RR models effects of disturbing factors, such as soil background and atmosphere, particularly at low canopy coverage. 1 3 PFG Table 2 Spectral indices used in the study Spectral indices Abbreviations Equation References Normalized Difference Vegetation Index NDVI (Rλ -Rλ )/(Rλ + Rλ ) Rouse et al. (1974) nir red nir red Ratio Vegetation Index RVI Rλ /Rλ Jordan (1969) nir red Difference Vegetation Index DVI Rλ -Rλ Jordan (1969) nir red Green Normalized Difference Vegetation Index GNDVI (Rλ -Rλ )/(Rλ + Rλ ) Gitelson and Merzlyak (1994) nir green nir green Green Ratio Vegetation Index GRVI (Rλ -Rλ )/(Rλ + Rλ ) Sripada et al. (2006) green red green red Green Difference Vegetation Index GDVI Rλ -Rλ Tucker et al. (1979) nir green Simple Ratio 695/760 Carter2 Ctr2 Rλ /Rλ Song (2013) rededge nir Chlorophyll Vegetation Index CVI Rλ *(Rλ /Rλ ) Zheng et al. (2018a) nir red green (R −R )∕(R +R ) Canopy Chlorophyll Content Index CCCI nir rededge nir rededge Wolanin et al. (2019) (R −R )∕(R +R ) nir red nir red −1 Chlorophyll Green CG (Rλ /Rλ ) Wu et al. (2008) nir green Chlorophyll Index Green CI (Rλ /Rλ )-1 Gitelson et al. (2003) green nir green Chlorophyll Index Red Edge CI (Rλ /Rλ )-1 Gitelson et al. (2003) rededge nir rededge are able to improve model accuracy and simplify the com- The importance of each variable is expressed by the mean plexity of the model (Prado Osco et al. 2019). For the PLSR squared error (MSE) at the time of model calculation (Tan and RR methods, the grid-search method was used to find et al. 2019). Subsequently, the predictors are ranked accord- the optimal number of factors and regularization factor (α), ing to the strength of the relationship between the input and and found that the optimal number of factors was 2 at tuber response variables. formation, tuber bulking and starch accumulation, 9 at the Spectral indices extracted from UAV images were used as combined stages and (α) were 10 in all growth stages. input of variables for RF regression. To determine the opti- mal number of trees (ntree) for die ff rent stages of potato ntree values were tested from 100 to 400 with an increment of 100, 2.7.2 Support Vector Regression (SVR) and the value of 200 at tuber formation, 100 at tuber bulking, 400 at starch accumulation, and 400 at the combined stages The support vector machine (SVM) was used to avoid hav- were selected due to stable lower RMSE. The number of vari- ing the problems of classification and regression (Wang et al. ables (mtry) was set to 4 at tuber formation, 9 at tuber bulking, 2020). Since the loss function gamma (g), the error penalty 4 at starch accumulation, and 1 at the combined stages, as it factor (C), and the choice of kernel function affect the perfor - yielded lower RMSE. In this study, the RF algorithms were mance of the SVM, the “Support Vactor Regression” func- implemented in the Python 3.10 environment, and the ‘Ran- tion in “sklearn” package was sued to implement the SVR dom Forest Regression” function in “sklearn” package was algorithms in the Python 3.10 environment in this study. The used. Other parameters were set as default values. commonly used radial basis kernel function (RBF) was applied in this study. Finally, a grid search was used to retrieve differ -2.8 Data Analysis ent combinations of g and C parameters. The best combination of g and C was 0.001 and 10.001 at tuber formation and tuber In this study, there were 132 sample points. Around 70% of bulking, and 0.001 and 20.001 at starch accumulation and the them were selected to train the four machine learning algo- combined stages for our data set, respectively. rithms, and 30% data were for validation. The performances of the different models of machine learning were evaluated by 2.7.3 Random Forest Regression Algorithm (RF) comparing the coefficients of determination (R ), root square error (RMSE) and mean square error (MSE) in predictions. The RF is an ensemble learning, that allows high accuracy and The higher the R and the lower the RMSE and MSE, the bet- generalization performance from the overall model (Shi et al. ter the precision and accuracy of the models. The R , RMSE 2020). Furthermore, the RF reduces overfitting by repeatedly and MSE were calculated according to Eqs. (1, 2, 3): putting back samples. In addition, RF takes into account for the � � � � ∑ ∑ 2 2 R =  y − y ∕ y − y (1) influence of input variables (name: mtry) to attenuate the auto- i i correlation between variables effectively (Peng et al. 2021). 1 3 PFG Fig. 4). However, the results showed that there was the sig- � � RMSE = y − y (2) nificant effect of the growth stages on the performances of i i i=1 the selected spectral indices (see Fig.  4). This study also found that the best spectral indices were GNDVI and CI , green � � ∑ especially at the growth stages of tuber formation and tuber MSE = y − y (3) i i bulking. At the starch accumulation and the combination i=1 of three stages, Ctr2 and CI showed significant cor - red-edge where y ,  y , and y are the observed, predicted, and observed i i relations with potato chlorophyll content, suggesting that mean values, respectively, and n is the sample size. chlorophyll content can be effectively estimated using these spectral indices. 3 Results 3.3 Estimation of Potato Chlorophyll Content Using Machine Learning Algorithms 3.1 Variations in Plant Chlorophyll Content in Different N Treatments and Growth Stages In this study, the chlorophyll content was predicted from highly correlated spectral indices at different growth stages The results in this study showed that the chlorophyll con- using four machine learning algorithms (SVR, RF, PLSR, tent increased from tuber formation to starch accumulation and RR) (see Fig. 5). The results showed that RF performed (see Fig. 3). The chlorophyll content in the training data set the highest accuracy in estimating chlorophyll content with ranged from 38.5 to 53.9 with a mean value of 45 and a CV R ranging from 0.96 to 0.97, RMSE ranging from 0.62 to value of 8.8%, while this in the testing data set ranged from 0.81, and MSE ranging from 0.38 to 0.65 for all growth 39.5 to 52.7 with a mean value of 45.5 and a CV value of stages. Except for the PLSR, all other machine learning 8.8%. The training and testing data sets showed a similar sta- algorithms showed the highest estimation accuracy at the tistical distribution, indicating that there were no potentially starch accumulation stage and followed by the tuber forma- biased estimations for model calibration and validation. tion stage. The results in Fig.  6 showed that RF was the optimal 3.2 Relationships Between Spectral Indices model for estimating potato chlorophyll content with R and Chlorophyll Content of Potato Plants of 0.61–0.83 that was greater than other models regard- less of growth stage, whereas RMSE (1.75–2.31) and MSE The results showed that the most selected spectral indi- (3.06–5.37) were smaller than other machine learning algo- ces were significantly correlated with the plant chloro- rithms. Compared to the RF, other machine learning algo- phyll content regardless of growth stage (see Table 2 and rithms showed the least performance, especially PLSR and RR algorithms due to the growth stage effect. Fig. 3 Variation of SPAD value 60 60 of potato at tuber formation, Training dataset Testing dataset tuber bulking and starch accu- 55 55 mulation stages of tuber forma- tion (T1), tuber bulking (T2) and starch accumulation (T3) 50 50 45 45 40 40 35 35 30 30 T1 T2 T3 T1 T2 T3 Growth stage Growth stage 1 3 SPAD SPAD PFG Fig. 4 Correlation analysis of spectral indices with SPAD value at different stages 1 3 PFG Fig. 4 (continued) 1 3 PFG 56 56 56 56 T1 T2 T3 T1+T2+T3 54 2 54 2 54 2 54 2 R =0.79 R =0.86 R =0.81 R =0.73 RMSE=1.56 RMSE=1.57 RMSE=1.58 RMSE=2.02 52 52 52 52 MSE=2.52 MSE=4.11 MSE=2.46 MSE=2.47 50 50 50 50 48 48 48 46 46 46 46 44 44 44 44 42 42 42 42 40 40 40 40 SVR SVR SVR SVR 38 38 38 38 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 56 56 56 56 T1 T2 T3 T1+T2+T3 54 2 54 2 54 2 54 2 R =0.96 R =0.97 R =0.96 R =0.96 RMSE=0.65 RMSE=0.62 RMSE=0.71 RMSE=0.81 52 52 52 52 MSE=0.42 MSE=0.38 MSE=0.51 MSE=0.65 50 50 50 50 48 48 48 48 46 46 46 44 44 44 42 42 42 42 40 40 40 40 RF RF RF RF 38 38 38 38 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 56 56 56 T1 T2 T3 T1+T2+T3 54 54 54 2 2 2 2 R =0.67 R =0.71 R =0.62 R =0.62 52 52 52 RMSE=2.29 RMSE=2.42 RMSE=1.96 RMSE=2.23 52 MSE=3.85 MSE=4.98 MSE=5.25 MSE=5.90 50 50 50 50 48 48 48 48 46 46 46 44 44 44 42 42 40 40 40 PLSR PLSR 40 PLSR PLSR 38 38 38 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 56 56 56 56 T1 T2 T3 T1+T2+T3 54 54 54 2 2 54 2 R =0.67 R =0.69 R =0.72 R =0.67 52 RMSE=1.98 52 RMSE=2.37 RMSE=1.98 RMSE=2.27 52 52 MSE=3.94 MSE=3.93 MSE=5.65 MSE=5.18 50 50 50 48 48 48 46 46 46 46 44 44 44 44 42 42 42 42 40 40 40 40 RR RR RR RR 38 38 38 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 Predicted SPAD value Predicted SPAD value Predicted SPAD value Predicted SPAD value Fig. 5 Relationships between the predicted value and observed value of training data set using four models at tuber formation stage (T1), tuber bulking stage (T2), starch accumulation stage (T3), and the combination of growth stages (T1 + T2 + T3) 3.4 Potato Chlorophyll Content Predictive Map increasing plant growth. During the tuber bulking stage, the number of pixels of medium chlorophyll content (see Fig. 7b, The RF model was used to map the predicted chlorophyll con- e) increased due to the rapid growth over potato tuber stage. During late reproduction, the nutrient delivery to the crop is tent of plants based on UAV multispectral images at the dif- ferent growth stages (see Fig. 7). Figure 7 shows that the esti- at its highest level, leading to the decrease in number of pixels due to low plant chlorophyll content in Fig. 7c, f. Meanwhile, mated value of chlorophyll content was relatively high in the tuber formation stage, whereas this gradually decreased with 1 3 Observed SPAD value Observed SPAD value Observed SPAD value Observed SPAD value PFG 54 56 Fig. 6 Validation of the per- SVR b RF formance of SVR (a), RF (b), R =0.54 R =0.61 RMSE=2.45 PLSR (c), and RR (d) at tuber RMSE=2.31 MSE=6.01 MSE=5.37 formation stage, tuber bulking 50 2 R =0.78 stage, starch accumulation stage R =0.79 RMSE=2.30 RMSE=2.14 and the combination of growth MSE=5.30 48 MSE=4.61 stages using the testing data set R =0.83 R =0.83 RMSE=2.13 RMSE=1.75 42 MSE=4.57 MSE=3.06 All R =0.74 All R =0.76 RMSE=2.35 RMSE=1.97 MSE=5.53 MSE=3.91 38 38 38 40 42 44 46 48 50 52 54 38 40 42 44 46 48 50 52 54 56 tuber formation tuber bulking starch accumulation PLSR RR R =0.49 2 R =0.50 RMSE=2.42 RMSE=2.27 MSE=5.89 MSE=5.18 50 2 R =0.74 2 R =0.71 RMSE=2.21 RMSE=2.26 48 MSE=4.89 MSE=5.13 46 46 44 44 R =0.80 R =0.71 RMSE=2.39 RMSE=2.37 42 42 MSE=5.75 MSE=5.63 All R =0.57 2 All R =0.70 40 40 RMSE=2.59 RMSE=2.19 MSE=6.72 MSE=4.80 38 38 38 40 42 44 46 48 50 52 54 38 40 42 44 46 48 50 52 54 Predicted SPAD value Predicted SPAD value chlorophyll content is significantly affected by the amount of and the accuracy of simple linear regression models was N applied, and the use of UAV images mapping allows for still poor (Lee et  al. 2020). Therefore, machine learn- real-time monitoring of the N status of the potato plant so that ing algorithms are promising as they can combine vari- nutrient management of the crop can be adjusted. ous spectral indices to improve the estimation accuracy (Qiu et  al. 2021). Numerous studies have already used RF, SVR, PLSR, and RR algorithms for estimations of 4 Discussion chlorophyll content (Tahir et  al. 2018). However, it is still unclear whether these algorithms are most suitable 4.1 Feasibility of Using UAV to Estimate Potato to fulfill the requirement (Li et al. 2020). Therefore, this Chlorophyll Content study compared four algorithms for precision and stabil- ity. The RF algorithm had higher R and lower RMSE The results in this study demonstrated that some spectral and MSE than others algorithms for chlorophyll content indices calculated from multispectral images were highly estimates in different growth stages, indicating that RF correlated with potato chlorophyll content, suggesting algorithm can provide accurate chlorophyll content esti- chlorophyll content of plants can be estimated by UAV mations. Our study also found that, since the model was multispectral image. Due to the different prediction per - trained using data set only from 2021, resulting in poor formance of spectral indices in different growth stages, models validation (see Fig.  8), there is a need to have a however, it was difficult to establish general models large amount of experimental data for training the models between different spectral indices and chlorophyll content, in future studies for the model stability. Recently, there were many research showed that using UAVs combined 1 3 Observed SPAD value Observed SPAD value PFG T1 (20 July, 2020) T2 (3 August, 2020) T3 (20 August, 2020) T1 (15 July, 2021) T2 (6 August, 2021) T3 (25 August, 2021) Fig. 7 Chlorophyll content map as indicated in SPAD values were derived by applying the RF model to the 8 spectral indices images for tuber formation (T1), tuber bulking (T2), and starch accumulation (T3) the machine learning algorithms could be monitored of 4.2 Analysis of Sensitive Bands important crop growth parameters, such as biomass, N, and leaf area index (Prado Osco et al. 2019; Abdelbaki Selection of the appropriate spectral bands from all wave- et al. 2021; Zhou et al. 2016), which further confirms the band to construct the spectral indices plays an important role feasibility our method and results. Compared with conven- in improving the model accuracy. At the tuber formation and tional wet chemical method, our method demonstrated the the tuber bulking stages, the green wavelength is closely advantages of convenience and speed, and compared with related to the potato chlorophyll content, which is in agree- the chlorophyll meters (SPAD-502), using UAVs could ment with previous study of the relationship between the achieve over a large area. Therefore, it was potential to spectral reflectance in the green region and plant chlorophyll estimate potato chlorophyll content using spectral indices content (Zhao et al. 2018). Hensen and Schjoerring (2003) calculated from UAV multispectral images. explained that chlorophyll a and b were more reflective in the green region, the green light region was often used to monitor chlorophyll content. However, this study suggests that the red-edge (680–800 nm) regions are important to 1 3 PFG 44 52 50.5 T1 T2 T3 T1+T2+T3 43 50 50.0 49.5 49.0 48.5 RR: R² = 0.53 R² = 0.35 R² = 0.38 R² = 0.35 48.0 40 RMSE = 2.38 RMSE = 7.22 RMSE = 2.44 RMSE = 6.49 40 MSE = 5.67 MSE = 52.25 MSE = 5.95 MSE = 42.23 38 38 47.5 42 44 46 48 50 52 40 42 44 46 48 50 40 42 44 46 38 40 42 44 46 48 50 52 T1 T2 T3 T1+T2+T3 46.6 43.8 46.4 43.7 46.2 46.0 43.6 42 45.8 43.5 45.6 43.4 45.4 PLSR: R² = 0.32 R² = 0.44 R² = 0.30 R² = 0.25 45.2 43.3 RMSE = 2.58 RMSE = 3.91 40 RMSE = 2.90 RMSE = 4.25 45.0 MSE = 15.34 MSE = 6.67 MSE = 8.46 MSE = 18.10 43.2 36 44.8 42 44 46 48 50 52 40 42 44 46 48 50 40 42 44 46 38 40 42 44 46 48 50 52 51 52 T1 T2 T3 T1+T2+T3 50 50 49 48 48 46 47 44 46 42 SVR: R² = 0.48 R² = 0.39 R² = 0.38 R² = 0.31 39 RMSE = 1.89 RMSE = 6.03 RMSE = 2.38 45 40 RMSE = 6.91 MSE = 3.58 MSE = 36.39 MSE = 5.69 MSE = 47.76 44 38 42 44 46 48 50 52 40 42 44 46 48 50 40 42 44 46 38 40 42 44 46 48 50 52 50 49.5 T1 T2 T3 T1+T2+T3 41.8 48 49.0 RF: 41.7 R² = 0.12 R² = 0.25 RMSE = 5.68 RMSE = 6.12 46 48.5 41.6 MSE = 32.28 MSE = 37.53 41.5 44 48.0 41.4 R² = 0.30 R² = 0.24 42 47.5 RMSE = 2.46 42 41.3 RMSE = 3.12 MSE = 6.10 MSE = 9.77 41.2 40 47.0 40 42 44 46 48 50 52 40 42 44 46 48 50 40 42 44 46 38 40 42 44 46 48 50 52 Observed SPAD value Observed SPAD value Observed SPAD value Observed SPAD value Fig. 8 Validation of the performance of RR, PLSR, SVR, and RF models at tuber formation stages (T1), tuber bulking stages (T2), starch accu- mulation (T3), and the combined stages (T1 + T2 + T3) using 2020 site data set potato chlorophyll content estimation at starch accumulation investigated at different growth stages to improve the accu- and the combination growth stages as well. Similar to the racy of estimation of chlorophyll content in potato crops. findings of Zheng et al. (2018b), the red-edge was a sensi- The spectral indices enhance the spectral characteristics tive region for potato chlorophyll content. Chang-Hua et al. sensitive to the biochemical properties of the vegetation by (2010) constructed a new red-edge spectral indices (RES) combining several bands and a specific formula format com - to estimate the chlorophyll content of sugar beet with high pared to the four bands. The wavebands of spectral indices accuracy. Therefore, the influence of growth stages on the contain some reference bands that are insensitive to chloro- selection of sensitive bands should be taken into account phyll content, and those bands are helpful to increase the sig- when spectral indices are applied. In future studies, the opti- nal-to-noise ratio (Yu et al. 2013). The spectral region used mization of the selected spectral indices should be further by the multispectral UAV is consistent with some satellite 1 3 Predicted SPAD value Predicted SPAD value Predicted SPAD value Predicted SPAD value PFG Fig. 9 Importance scores for spectral indices in the best RF model for predicting potato SPAD value at tuber formation stage (T1), tuber bulking stage (T2), starch accumulation stage (T3), and combined growth stages (T1 + T2 + T3) multispectral bands, such as the Sentinel-2. Therefore, the 5 Conclusions estimation framework constructed in this study can also be applied to satellite images for large-scale crop chlorophyll An integrated method combined with spectral indices and ran- content estimation (see Fig. 2). dom forest model was proposed to estimate chlorophyll content in potato crop. Accuracy and robustness of the RF model were 4.3 T he Eec ff t of the Number of Spectral Indices verified for each growth stage, such as tuber formation, tuber on the Model bulking, starch accumulation, and the combination of growth stages. The results of comparison of the RF model with SVR, Previous studies have also shown that predicting chlorophyll PLSR, and RR showed that the accuracy of RF outperformed content was affected by the use of variable importance plot at each stage. The correlation analysis demonstrated that in RF that eliminates spectral variable inputs. The contribu- most of spectral indices were significantly correlated with the tion of each spectral index to the RF model indicates the potato chlorophyll content, whereas the spectral indices with extent to which each of the spectral indices could predict the strongest correlation varied at different period. The CI green the chlorophyll content variations (see Table 2 and Fig. 9). and the green band highly contributed to the estimation model However, there is a need to analyze the performance of the during the tuber formation and tuber bulking stage. However, model for the reduction in the number of input variables the CI and the red edge band provided higher contribu- red edge and increase computational efficiency. By implementing the tion to the prediction model during the starch accumulation best 1 to 3 spectral indices into the RF model, only a slight and the combination of growth stages. decrease was observed in its performance (see Fig. 10). Sim- ilarly, Prado Osco et al. (2019) used the variable importance graph in RF model to eliminate spectral variable inputs to reduce the amount of spectral data computation and keep the model with better performance. 1 3 PFG 56 56 56 56 mtry: GNDVI mtry: GNDVI mtry: GDVI mtry: CI rededge T1 T3 T2 T1+T2+T3 54 54 54 54 2 2 R =0.94 R =0.97 R =0.89 RMSE=0.85 RMSE=0.69 52 52 52 RMSE=1.23 52 R =0.94 MSE=0.72 MSE=0.47 MSE=1.51 RMSE=0.97 50 50 50 50 MSE=0.95 48 48 48 2 46 46 46 R =0.43 RMSE=2.94 44 44 44 MSE=8.65 42 42 42 2 2 2 R =0.58 R =0.89 R =0.71 RMSE=2.61 RMSE=1.36 RMSE=2.35 40 training dataset 40 40 40 MSE=6.85 MSE=1.87 MSE=5.52 testing dataset 38 38 38 38 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 56 56 56 56 mtry: GNDVI, CI mtry: GDVI, CG mtry: GNDVI, CG green mtry: GNDVI, CI rededge T1 T2 T3 54 54 54 T1+T2+T3 2 2 2 R =0.94 R =0.97 R =0.89 R =0.96 52 52 52 52 RMSE=0.84 RMSE=1.22 RMSE=0.84 RMSE=0.81 MSE=0.70 MSE=1.51 MSE=0.70 50 50 50 MSE=0.66 48 48 48 46 2 46 46 R =0.43 RMSE=2.94 44 44 44 MSE=8.65 42 42 42 2 42 2 R =0.69 R =0.73 R =0.88 RMSE=2.41 RMSE=2.07 40 40 RMSE=1.43 40 training dataset 40 MSE=5.81 MSE=4.28 MSE=2.05 testing dataset 38 38 38 38 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 56 56 56 56 mtry: GNDVI, CVI, CI mtry: GDVI, DVI, CG mtry: GNDVI, CG, CI green mtry: GNDVI, CI , CI green green rededge T1 T2 T3 54 54 54 54 T1+T2+T3 2 2 R =0.95 R =0.97 R =0.89 52 52 RMSE=0.64 RMSE=1.22 R =0.96 RMSE=0.77 52 52 MSE=0.42 MSE=1.51 RMSE=0.82 MSE=0.59 50 50 50 50 MSE=0.68 48 48 48 48 46 2 46 46 46 R =0.52 RMSE=2.80 44 44 44 MSE=7.86 2 2 42 42 42 2 R =0.87 R =0.70 R =0.72 RMSE=1.51 RMSE=2.38 RMSE=2.11 40 training dataset 40 40 MSE=2.30 MSE=5.67 MSE=4.48 testing dataset 38 38 38 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 Predicted SPAD value Predicted SPAD value Predicted SPAD value Predicted SPAD value Fig. 10 Optimizing the number of predictors of spectral indices according to important scores at tuber formation stage (T1), tuber bulking stage (T2), starch accumulation stage (T3) and the combination of growth stages (T1 + T2 + T3) for training data set and testing data set Funding This research was funded by Programs for Key Science References and Technology Development of Inner Mongolia in 2019 and 2020 (2019GG248 and 2020GG0038), the National Natural Science Founda- Abdelbaki A, Schlerf M, Retzlaff R et al (2021) Comparison of crop tion of China (32160757). trait retrieval strategies using UAV-based VNIR hyperspectral imaging. Remote Sens 13(9):1748. https://doi. or g/10. 3390/ rs130 Open Access This article is licensed under a Creative Commons Attri- bution 4.0 International License, which permits use, sharing, adapta- Argenta G, Silva PRF, Sangoi L (2004) Leaf relative chlorophyll con- tion, distribution and reproduction in any medium or format, as long tent as an indicator parameter to predict nitrogen fertilization in as you give appropriate credit to the original author(s) and the source, maize. Ciência Rural 34:1379–1387. https:// doi. or g/ 10. 1590/ provide a link to the Creative Commons licence, and indicate if changes S0103- 84782 00400 05000 09 were made. The images or other third party material in this article are Bareth G, Aasen H, Bendig J, et al (2015) Low-weight and UAV-based included in the article's Creative Commons licence, unless indicated hyperspectral full-frame cameras for monitoring crops: spectral otherwise in a credit line to the material. If material is not included in comparison with portable spectroradiometer measurements. the article's Creative Commons licence and your intended use is not Unmanned aerial vehicles (UAVs) for multi-temporal crop surface permitted by statutory regulation or exceeds the permitted use, you will modelling. https:// doi. org/ 10. 1127/ pfg/ 2015/ 0256 need to obtain permission directly from the copyright holder. To view a Bendig J, Bolten A, Bareth G (2013) UAV-based imaging for multi- copy of this licence, visit http://cr eativ ecommons. or g/licen ses/ b y/4.0/ . temporal, very high resolution crop surface models to monitor crop growth variability. Unmanned aerial vehicles (UAVs) for 1 3 Observed SPAD value Observed SPAD value Observed SPAD value PFG multi-temporal crop surface modelling. https:// doi. org/ 10. 1127/ Li C, Chen P, Ma C et al (2020) Estimation of potato chlorophyll con- 1432- 8364/ 2013/ 0200 tent using composite hyperspectral index parameters collected by Cao Z, Cheng T, Ma X et al (2017) A new three-band spectral index an unmanned aerial vehicle. Int J Remote Sens 41(21):8176–8197. for mitigating the saturation in the estimation of leaf area index https:// doi. org/ 10. 1080/ 01431 161. 2020. 17577 79 in wheat. Int J Remote Sens 38(13):3865–3885. https:// doi. org/ Lu W, Lou J, Webster C et al (2021) Estimating construction waste 10. 1080/ 01431 161. 2017. 13061 41 generation in the Greater Bay Area, China using machine learn- Chang-Hua JU, Yong-Chao T, Xia YAO et al (2010) Estimating leaf ing. Waste Manage 134:78–88. https://doi. or g/10. 1016/j. w asman. chlorophyll content using red edge parameters. Pedosphere 2021. 08. 012 20(5):633–644. https:// doi. org/ 10. 1016/ S1002- 0160(10) 60053-7 Markwell J, Osterman JC, Mitchell JL (1995) Calibration of the Clevers JGPW, Kooistra L, Van den Brande MMM (2017) Using Sen- Minolta SPAD-502 leaf chlorophyll meter. Photosynth Res tinel-2 data for retrieving LAI and leaf and canopy chlorophyll 46(3):467–472. https:// doi. org/ 10. 1007/ BF000 32301 content of a potato crop. Remote Sens 9(5):405. https:// doi. org/ Martínez D, Guiamet J (2004) Distortion of the SPAD 502 chlorophyll 10. 3390/ rs905 0405 meter readings by changes in irradiance and leaf water status. Colorado JD, Calderon F, Mendez D et al (2020) A novel NIR-image Agronomie 24(1):41–46. https:// doi. org/ 10. 1051/ agro: 20030 60 segmentation method for the precise estimation of above-ground Moghimi A, Pourreza A, Zuniga-Ramirez G et  al (2020) A novel biomass in rice crops. PLoS ONE 15(10):e0239591. https:// doi. machine learning approach to estimate grapevine leaf nitrogen org/ 10. 1371/ journ al. pone. 02395 91 concentration using aerial multispectral imagery. Remote Sens Franzini M, Ronchetti G, Sona G et al (2019) Geometric and radio- 12(21):3515. https:// doi. org/ 10. 3390/ rs122 13515 metric consistency of parrot sequoia multispectral imagery for Mutanga O, Adam E, Cho MA (2012) High density biomass estimation precision agriculture applications. Appl Sci 9(24):5314. https:// for wetland vegetation using WorldView-2 imagery and random doi. org/ 10. 3390/ app92 45314 forest regression algorithm. Int J Appl Earth Obs Geoinf 18:399– Gitelson AA, Merzlyzk MN (1994) Quantitative estimation of chlo- 406. https:// doi. org/ 10. 1016/j. jag. 2012. 03. 012 rophyll-a using reflectance spectra: Experiments with autumn Peng L, Wang L, Ai XY et  al (2021) Forecasting tourist arrivals chestnut and maple leaves. J Photochem Photobiol B Biol 22:247– via random forest and long short-term memory. Cogn Comput 252. https:// doi. org/ 10. 1016/ 1011- 1344(93) 06963-4 13(1):125–138. https:// doi. org/ 10. 1007/ s12559- 020- 09747-z Gitelson AA, Buschmann C, Lichtenthaler HK (1999) The chlorophyll Prado Osco L, Marques Ramos AP, Roberto Pereira D et al (2019) fluorescence ratio F735/F700 as an accurate measure of the chlo- Predicting canopy nitrogen content in citrus-trees using random rophyll content in plants. Remote Sens Environ 69(3):296–302. forest algorithm associated to spectral vegetation indices from https:// doi. org/ 10. 1016/ S0034- 4257(99) 00023-1 UAV-imagery. Remote Sens 11(24):2925. https://doi. or g/10. 3390/ Gitelson AA, Yuri G, Merzlyark MN (2003) Relationships between rs112 42925 leaf chlorophyll content and spectral reflectance and algorithms Qi H, Wu Z, Zhang L et al (2021) Monitoring of peanut leaves chloro- for non-destructive chlorophyll assessment in higher plant phyll content based on drone-based multispectral image feature leaves. J Plant Physiol 160(3): 271–282. https:// doi. org/ 10. 1078/ extraction. Comput Electron Agric 187:106292. https:// doi. org/ 0176- 1617- 0088710. 1016/j. compag. 2021. 106292 Hansen PM, Schjoerring JK (2003) Reflectance measurement of can- Qiu Y, Garg D, Kim SM et al (2021) Machine learning algorithms opy biomass and nitrogen status in wheat crops using normalized to predict flow boiling pressure drop in mini/micro-channels difference vegetation indices and partial least squares regression. based on universal consolidated data. Int J Heat Mass Transf Remote Sens Environ 86(4):542–553. https:// doi. org/ 10. 1016/ 178:121607. https:// doi. org/ 10. 1016/j. ijhea tmass trans fer. 2021. S0034- 4257(03) 00131-7 121607 Houles V, Guerif M, Mary B (2007) Elaboration of a nitrogen nutrition Rafael MT, María SC, Sergio OM et al (2014) Environmental and indicator for winter wheat based on leaf area index and chloro- economic impact of Forest fires in Puerto Rico 2013–2014. Open phyll content for making nitrogen recommendations. Eur J Agron J for 5(04):353. https:// doi. org/ 10. 4236/ ojf. 2015. 54030 27(1):1–11. https:// doi. org/ 10. 1016/j. eja. 2006. 10. 001 Rouse JW, Haas RH, Schell JA (1974) Monitoring vegetation systems Inoue Y, Sakaiya E, Zhu Y et al (2012) Diagnostic mapping of canopy in the Great Plains with ERTS. NASA Spec Publ 351:309 nitrogen content in rice based on hyperspectral measurements. Shah SH, Angel Y, Houborg R et al (2019) A random forest machine Remote Sens Environ 126:210–221. https://doi. or g/10. 1016/j. rse. learning approach for the retrieval of leaf chlorophyll content in 2012. 08. 026 wheat. Remote Sens 11(8):920. https:// doi. or g/ 10. 3390/ rs110 Jordan CF (1969) Derivation of leaf-area index from quality of light 80920 on the forest floor. Ecology 50:663–666. https:// doi. org/ 10. 2307/ Shi L, Liu Q, Huang C et al (2020) Comparing pixel-based random 19362 56 forest and the object-based support vector machine approaches Kooistra L, Clevers JG (2016) Estimating potato leaf chlorophyll con- to map the quasi-circular vegetation patches using individual tent using ratio vegetation indices. Remote Sens Lett 7(6):611– seasonal fused GF-1 imagery. IEEE Access 8:228955–228966. 620. https:// doi. org/ 10. 1080/ 21507 04X. 2016. 11719 25https:// doi. org/ 10. 1109/ ACCESS. 2020. 30450 57 Kopačková-Strnadová V, Koucká L, Jelének J et al (2021) Canopy top, Singhal G, Bansod B, Mathew L et al (2019) Chlorophyll estimation height and photosynthetic pigment estimation using parrot sequoia using multi-spectral unmanned aerial system based on machine multispectral imagery and the unmanned aerial vehicle (UAV). learning techniques. Remote Sens Appl Soc Environ 15:100235. Remote Sens 13(4):705. https:// doi. org/ 10. 3390/ rs130 40705https:// doi. org/ 10. 1016/j. rsase. 2019. 100235 Lee H, Wang J, Leblon B (2020) Using linear regression, random for- Song C (2013) Optical remote sensing of forest leaf area index and ests, and support vector machine with unmanned aerial vehicle biomass. Prog Phys Geogr 37(1):98–113. https://doi. or g/10. 1177/ multispectral images to predict canopy nitrogen weight in corn. 03091 33312 471367 Remote Sens 12(13):2071. https:// doi. org/ 10. 3390/ rs121 32071 Sripada RP, Heiniger RW, White JG et al (2006) Aerial color infrared Li F, Miao Y, Hennig SD et al (2010) Evaluating hyperspectral veg- photography for determining early in-season nitrogen require- etation indices for estimating nitrogen concentration of winter ments in corn. Agron J 98:968–977. https://doi. or g/10. 2134/ ag ron wheat at different growth stages. Precis Agric 11(4):335–357. j2005. 0200 https:// doi. org/ 10. 1007/ s11119- 010- 9165-6 Tahir MN, Naqvi SZA, Lan Y et al (2018) Real time estimation of chlorophyll content based on vegetation indices derived from 1 3 PFG multispectral UAV in the kinnow orchard. Int J Precis Agric Aviat. Wu C, Niu Z, Tang Q et al (2008) Estimating chlorophyll content from https:// doi. org/ 10. 33440/j. ijpaa. 20180 101. 0001 hyperspectral vegetation indices: modeling and validation. Agric Tan K, Ma W, Wu F et al (2019) Random forest–based estimation of For Meteorol 148(8–9):1230–1241. https://doi. or g/10. 1016/j. ag rfo heavy metal concentration in agricultural soils with hyperspectral rmet. 2008. 03. 005 sensor data. Environ Monit Assess 191(7):1–14. https:// doi. org/ Wu J, Chen XY, Zhang H et al (2019) Hyperparameter optimization 10. 1007/ s10661- 019- 7510-4 for machine learning models based on Bayesian optimization. J Théau J, Gavelle E, Ménard P (2020) Crop scouting using UAV Electron Sci Technol 17(1):26–40. https://doi. or g/10. 11989/ JES T. imagery: a case study for potatoes. J Unman Veh Syst 8(2):99–1674- 862X. 80904 120 118. https:// doi. org/ 10. 1139/ juvs- 2019- 0009 Yu K, Li F, Gnyp ML et al (2013) Remotely detecting canopy nitrogen Tucker C, Elgin J, Mcmurtrey J (1979) Monitoring corn and soy- concentration and uptake of paddy rice in the Northeast China bean crop development with hand-held radiometer spectral Plain. ISPRS J Photogramm Remote Sens 78:102–115. https:// data. Remote Sens Environ 8:237–248. https:// doi. org/ 10. 1016/ doi. org/ 10. 1016/j. isprs jprs. 2013. 01. 008 0034- 4257(79) 90004-X Yu K, Lenz-Wiedemann V, Chen X et al (2014) Estimating leaf chlo- Uddling J, Gelang-Alfredsson J, Piikki K et al (2007) Evaluating the rophyll of barley at different growth stages using spectral indices relationship between leaf chlorophyll concentration and SPAD- to reduce soil background and canopy structure effects. ISPRS 502 chlorophyll meter readings. Photosynth Res 91(1):37–46. J Photogramm Remote Sens 97:58–77. https:// doi. org/ 10. 1016/j. https:// doi. org/ 10. 1007/ s11120- 006- 9077-5isprs jprs. 2014. 08. 005 Verrelst J, Muñoz J, Alonso L et al (2012) Machine learning regression Zhao B, Duan A, Ata-Ul-Karim ST et al (2018) Exploring new spec- algorithms for biophysical parameter retrieval: opportunities for tral bands and vegetation indices for estimating nitrogen nutrition Sentinel-2 and-3. Remote Sens Environ 118:127–139. https://d oi. index of summer maize. Eur J Agron 93:113–125. https://d oi.o rg/ org/ 10. 1016/j. rse. 2011. 11. 00210. 1016/j. eja. 2017. 12. 006 Viljanen N, Honkavaara E, Näsi R et al (2018) A novel machine learn- Zheng Q, Huang W, Cui et al (2018a) New spectral index for detecting ing method for estimating biomass of grass swards using a photo- wheat yellow rust using Sentinel-2 multispectral imagery. Sensors grammetric canopy height model, images and vegetation indices 18(3):868. https:// doi. org/ 10. 3390/ s1803 0868 captured by a drone. Agriculture 8(5):70. https://doi. or g/10. 3390/ Zheng T, Liu N, Wu L et al (2018b) Estimation of chlorophyll con- agric ultur e8050 070 tent in potato leaves based on spectral red edge position. IFAC- Wolanin A, Camps-Valls G, Gómez-Chova L et al (2019) Estimating PapersOnLine 51(17):602–606. https:// doi. org/ 10. 1016/j. ifacol. crop primary productivity with Sentinel-2 and Landsat 8 using 2018. 08. 131 machine learning methods trained with radiative transfer simu- Zhou X, Zhu X, Dong Z et al (2016) Estimation of biomass in wheat lations. Remote Sens Environ 225:441–457. https:// doi. org/ 10. using random forest regression algorithm and remote sensing data. 1016/j. rse. 2019. 03. 002 Crop J 4(3):212–219. https:// doi. org/ 10. 1016/j. cj. 2016. 01. 008 1 3 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png "PFG – Journal of Photogrammetry, Remote Sensing and Geoinformation Science" Springer Journals

Multi-temporal UAV Imaging-Based Mapping of Chlorophyll Content in Potato Crop

Loading next page...
 
/lp/springer-journals/multi-temporal-uav-imaging-based-mapping-of-chlorophyll-content-in-kwnZKh4UPy
Publisher
Springer Journals
Copyright
Copyright © The Author(s) 2022
ISSN
2512-2789
eISSN
2512-2819
DOI
10.1007/s41064-022-00218-8
Publisher site
See Article on Publisher Site

Abstract

Spectral indices based on unmanned aerial vehicle (UAV) multispectral images combined with machine learning algorithms can more effectively assess chlorophyll content in plants, which plays a crucial role in plant nutrition diagnosis, yield estima- tion and a better understanding of plant and environment interactions. Therefore, the aim of this study was to use UAV-based spectral indices deriving from UAV-based multispectral images as inputs in different machine learning models to predict canopy chlorophyll content of potato crops. The relative chlorophyll content was obtained using a SPAD chlorophyll meter. Random Forest (RF), support vector regression (SVR), partial least squares regression (PLSR) and ridge regression (RR) were employed to predict the chlorophyll content. The results showed that RF model was the best performing algorithm with an R of 0.76, Root Mean Square Error (RMSE) of 1.97. Both RF and SVR models showed much better accuracy than PLSR and RR models. This study suggests that the best models, RF model, allow to map the spatial variation in chlorophyll content of plant canopy using the UAV multispectral images at different growth stages. Keywords Chlorophyll content · Machine learning · Multispectral images · Potato · Unmanned aerial vehicle (UAV) Zusammenfassung Multitemporale UAV-gestützte Kartierung des Chlorophyllgehalts in Kartoffelkulturen. Spektrale Indizes, die auf multi- spektralen Bildern von der Unmanned Aerial Vehicles (UAVs) basieren, können in Kombination mit Algorithmen des Maschinelles Lernen (ML) den Chlorophyllgehalt von Pflanzen effektiver bewerten. Dieser spielt eine entscheidende Rolle bei der Diagnose der Pflanzenernährung, der Ertragsschätzung und einem besseren Verständnis der Wechselwirkungen zwischen Pflanze und Umwelt. Ziel dieser Studie war es daher, UAV-basierte Spektralindizes, die aus UAV-basierten Multi - spektralbildern abgeleitet wurden, als Input für verschiedene ML-Modelle zu verwenden, um den Chlorophyllgehalt von Kartoffelpflanzen im Kronendach vorherzusagen. Der relative Chlorophyllgehalt wurde mit einem SPAD-Chlorophyllmess- gerät ermittelt. Random Forest (RF), Support Vector Regression (SVR), Partial Least Squares Regression (PLSR) und Ridge Regression (RR) wurden zur Vorhersage des Chlorophyllgehalts eingesetzt. Die Ergebnisse zeigten, dass das RF-Modell mit einem R von 0,76 und einem Root Mean Square Error (RMSE) von 1,97 der beste Algorithmus war. Sowohl das RF- als auch das SVR-Modelle zeigten eine viel bessere Genauigkeit als die PLSR- und RR-Modelle. Diese Untersuchung deutet darauf hin, dass das beste Modell, das RF-Modell, es ermöglicht, die räumliche Variation des Chlorophyllgehalts in Pflanzen- dächern mit Hilfe von multispektralen UAV-Bildern in verschiedenen Wachstumsstadien abzubilden. * Fei Li feili72@163.com 1 Introduction Inner Mongolia Key Laboratory of Soil Quality The chlorophylls are the most important pigment for photo- and Nutrient Resource, College of Grassland, Resources and Environment, Inner Mongolia Agricultural University, synthesis (Gitelson et al. 1999; Singhal et al. 2019), which Hohhot 010011, China provides energy to the biosphere (Qi et al. 2021). Since the Inner Mongolia Academy of Forestry Sciences, symptoms of plant response to stress and N deficiencies Hohhot 010010, China are often associated with plant chlorophyll content (Houles Department Life Science Engineering, School of Life et al. 2007), timely determination of plant chlorophyll con- Sciences, Technical University of Munich, 85354 Freising, tent can be important indicators of plant nutrient status and Germany Vol.:(0123456789) 1 3 PFG environmental effects, which is very useful for agricultural squares regression (PLSR), ridge regularization (RR) are field management (Argenta et al. 2004). powerful tools to assist in UAV image computation. These Conventionally, the determination of chlorophyll content algorithms performed well in current applications predicting involves outdoor sampling and indoor chemical analysis. plant conditions, such as N content (Colorado et al. 2020), These methods are not only time-consuming and labori- chlorophyll content (Qi et al. 2021), and biomass (Vilijanen ous, but also destructive and lagging. As a non-destructive et al. 2018). Using spectral indices as input variables is a technique, the portable chlorophyll meters (SPAD-502) have new approach, since these models ensure good performance been widely used to estimate chlorophyll content in agricul- in spite of only a few variables as input feature (Moghimi tural studies (Markwell et al. 1995; Martínez et al. 2004). et al. 2020). Mutanga et al. (2012). found that predicting However, the SPAD meter is a leaf-based measurement, and chlorophyll content using the RF algorithm and three NDVIs infeasible for large-scale regional monitoring (Uddling et al. calculated from the red-edge and near-infrared bands pro- 2007). Therefore, there is a need to develop the techniques to duced low prediction errors. Shah et al. (2019) used the monitor plant chlorophyll content in real time on a regional established spectral indices as an input variable combined scale. with the RF algorithm to improve the accuracy of estimating In recent years, remote sensing technology allows to chlorophyll content in wheat. However, most studies have monitor plant chlorophyll content on a large regional scale only compared a single machine learning algorithm com- (Yu et al. 2014). UAV-based multispectral imagery has been bined with spectral indices to a linear regression model and widely used to monitor the nutrient status in crops (Théau did not compare the differences between different machine et al. 2020), e.g., chlorophyll content and N (Clevers et al. learning algorithms. Therefore, the main objectives of this 2017; Verrelst et al. 2012). UAV image acquisition is not study were: (1) to evaluate the performances of published only fast, less consuming and flexible (Bendig et al. 2013), spectral indices in estimating chlorophyll content of potato but also with higher spatial and temporal resolution than the plants, (2) to compare the performance of SVR, RF, PLSR satellite and aerial images to facilitate further image analysis and RR models in predicting potato chlorophyll content, (3) (Bareth et al. 2015). Although satellite and aerial images to generate chlorophyll content maps based on the optimal are able to assess crop nutrient status on a large scale, such models, and (4) to verify whether the UAV images were images possess the low spatial and temporal resolution. able to detect the variation of potato chlorophyll content at Rafael et al. (2014) have shown that there was difficulty different growth stages. of differentiating the content of chlorophyll in forest using multispectral satellite imagery. Over the past decades, how- ever, the use of UAV images has rapidly narrowed the gap 2 Materials and Methods between satellite or aerial and ground-based sensing (Prado Osco et al. 2019).2.1 Experimental Sites However, the combination of remote sensing techniques with multispectral imagery requires more reliable assess- Experiments were carried out in Zhuozi Country, which is ment techniques due to the huge amount of data generated. located in the middle of Inner Mongolia (extending from Compared to individual spectral bands, spectral indices can 110° 51′ E, 40° 38′ N to 112° 56′ E, 41 °16′ N), China (see be applied at different scales and mitigate to a certain extent Fig. 1). The major crops are potato, corn, and beans. This the adverse effects due to anisotropic reflection, background area is characterized with a middle temperate arid and semi- shadows and soil brightness contributions (Inoue et al. 2012; arid continental monsoon, cold winters and cool summers. Kooistra and Clevers 2016). For example, Qi et al. (2021) The annual average precipitation is 544.5 mm, i.e., 90–95% found NDVI and GNDVI had a much higher degree of fit between April and October. The average temperature and precision than the other indices. Tahir et  al. (2018) is > 20 °C during the potato growing season. found MSAVI2 and TNDVI were proved to be more robust indices to estimate the chlorophyll content in the orchard 2.2 Experimental Design with the highest coefficients of determination (R ) 0.89 and 0.85, respectively. Although spectral indices can improve the Two experiments involving different potato cultivars were robustness in predicting chlorophyll content to some degree, conducted at two different sites of Zhuozi county from June the saturation effect and multi-collinearity are still a problem to September in 2020 and 2021, respectively. Experiment 1 for remotely estimating chlorophyll content in crops (Cao in 2020 was a randomized complete block design with four et al. 2017; Li et al. 2010). replications. The potato cultivar was Yingniweite. There Recently, machine learning algorithms have been induced were five N treatments: four optimization of N application in remote sensing (Wu et al. 2019; Lu et al. 2021). Support management, and conventional N rate fertilization (Con). vector machine (SVM), random forests (RF), partial least The plot size was 10 × 10 m. There were six N treatments 1 3 PFG Fig. 1 Geographical location of the experimental site (a) in 2020 and site (b) in 2021 with four replications for Experiment 2 in 2021. The potato regions: green (530–570 nm), red (640–680 nm), red-edge cultivar was Mingfeng 16. Five N treatments were same as (730–740 nm), and near-infrared (770–810 nm). Because, those in experiment 1, but control (no N was applied) was during the flight, UAV images were affected by partly added. The plot size was 10 × 12 m. For both experiments, cloudy and the changing illumination in light intensity, a the irrigation method of the study site was drip irrigation, radiometric calibration white board and an onboard irradi- and the field management was unified according to the local ance sensor were integrated in the multispectral camera to large field requirements. correct images for dash area and illumination differences the multispectral camera with (Kopačková-Strnadová et al. 2.3 Field Data Acquisition 2021; Franzini et al. 2019). Additional sensors such as GPS and laser sensor are also mounted in the UAV (see A Minolta brand chlorophyll meter (Model SPAD-502; Fig. 2). The properties of the UAV camera are presented Spectrum Technologies Inc. Plainfield, IL) was used to in Table 1. measure the chlorophyll content of potato functional leaves (the 4th compound leaf from apex) before the UAV flight was carried out. Fifteen sampling points were randomly 2.5 UAV Image Processing selected from each treatment. Each measured point repre- sented the growth of a potato plant. The average of chlo- Figure 2 illustrates the flow chart of this study. Pix4Dmap- rophyll content of each plot can be precisely acquired by per Version 4.5.6 was used to import all images taken in averaging the fifteen samples of data. the same period according to the location coordinates in its properties, and to align the overlapping images using 2.4 Acquisition of UAV Images the feature point matching algorithm. Before the UAV flight, 4 ground control points were evenly arranged in During the experimental period, the multispectral images the field to obtain accurate geographical reference posi- were taken by the Parrot Bluegrass drone with Parrot tions. The plane precision of ground control points sys- Sequoia multispectral camera on 20 July, 3 August, and tem is ± (8 + 1 × 10–6 × D) mm, and the elevation accuracy 20 August in 2020, whereas the images were acquired on is ± (15 + 1 × 10–6 × D) mm. The area marked by ground 15 July, 6 August, and 25 August in 2021. UAV flight was control points is 40 cm × 40 cm. The UAV-based images conducted between 10:00 a.m. and 2:00 p.m. at 30 m high, were first aligned using Pix4Dmapper, and the camera resulting in a 10 cm ground sample distance (GSD). Mul- parameters were estimated from the coordinates of their tispectral images were taken from the following spectral 1 3 PFG Fig. 2 Unmanned aerial vehicle (UAV)-based robotic system for chlorophyll content estimation in potato Table 1 Specifications of parrot Multi-spectral camera Flight conditions sequoia camera and flight conditions Wavelength bands Specifications Green 550 (± 40) Spectral resolution 10 bits or 10 cm Height 30 m Red 660 (± 40) HFOV 70.6° Time 01:30 P.M Red edge 735 (± 40) VFOV 52.6° Weather Partially cloudy Precipitation: 0 mm Near infrared 790 (± 40) DFOV 89.6° Wind: At 1–2 m/s photographs and GCPs. A depth sharpening filter to fur - The relationship between 12 spectral indices and the ther improve images edge detail was selected, and then chlorophyll content was calculated by Pearson correlation dense point clouds were built. Finally, meshes and tex- analysis. The correlation coefficient (r) values were used to tures based on the camera’s own parameters were created evaluate the correlations between indices and chlorophyll and the processed images were exported as TIFF images. content. Their performance was further evaluated by statis- ENVI 5.1 was used to mark the region of interest (ROI) tical comparison of the returned individual spectral indices (see Fig.  1) in the experimental plot and to extract the correlation. average reflectance values in the four bands of ROI for each images. 2.7 Machine Learning Algorithms 2.6 Spectral Indices 2.7.1 Partial Least Squares Regression and Ridge Four canopy reflectance values (Green, Red, Red edge, Regression (PLSR and RR) Near-infrared) were then used to computer 12 spectral indi- ces that were evaluated with their correlations with chloro- Before the estimation model was constructed, Pearson cor- relation analyses between chlorophyll content measurements phyll content (see Table 2). The indices were enhanced by the contribution of vegetation optical properties according and 12 spectral indices were conducted. The indices with higher correlation coefficients (|r |> 0.4, and p values less to the spectral response of the canopy by combining several bands. And thus, spectral indices were able to reduce the than 0.01) were selected for machine learning algorithms. Compared to simple linear regression, PLSR and RR models effects of disturbing factors, such as soil background and atmosphere, particularly at low canopy coverage. 1 3 PFG Table 2 Spectral indices used in the study Spectral indices Abbreviations Equation References Normalized Difference Vegetation Index NDVI (Rλ -Rλ )/(Rλ + Rλ ) Rouse et al. (1974) nir red nir red Ratio Vegetation Index RVI Rλ /Rλ Jordan (1969) nir red Difference Vegetation Index DVI Rλ -Rλ Jordan (1969) nir red Green Normalized Difference Vegetation Index GNDVI (Rλ -Rλ )/(Rλ + Rλ ) Gitelson and Merzlyak (1994) nir green nir green Green Ratio Vegetation Index GRVI (Rλ -Rλ )/(Rλ + Rλ ) Sripada et al. (2006) green red green red Green Difference Vegetation Index GDVI Rλ -Rλ Tucker et al. (1979) nir green Simple Ratio 695/760 Carter2 Ctr2 Rλ /Rλ Song (2013) rededge nir Chlorophyll Vegetation Index CVI Rλ *(Rλ /Rλ ) Zheng et al. (2018a) nir red green (R −R )∕(R +R ) Canopy Chlorophyll Content Index CCCI nir rededge nir rededge Wolanin et al. (2019) (R −R )∕(R +R ) nir red nir red −1 Chlorophyll Green CG (Rλ /Rλ ) Wu et al. (2008) nir green Chlorophyll Index Green CI (Rλ /Rλ )-1 Gitelson et al. (2003) green nir green Chlorophyll Index Red Edge CI (Rλ /Rλ )-1 Gitelson et al. (2003) rededge nir rededge are able to improve model accuracy and simplify the com- The importance of each variable is expressed by the mean plexity of the model (Prado Osco et al. 2019). For the PLSR squared error (MSE) at the time of model calculation (Tan and RR methods, the grid-search method was used to find et al. 2019). Subsequently, the predictors are ranked accord- the optimal number of factors and regularization factor (α), ing to the strength of the relationship between the input and and found that the optimal number of factors was 2 at tuber response variables. formation, tuber bulking and starch accumulation, 9 at the Spectral indices extracted from UAV images were used as combined stages and (α) were 10 in all growth stages. input of variables for RF regression. To determine the opti- mal number of trees (ntree) for die ff rent stages of potato ntree values were tested from 100 to 400 with an increment of 100, 2.7.2 Support Vector Regression (SVR) and the value of 200 at tuber formation, 100 at tuber bulking, 400 at starch accumulation, and 400 at the combined stages The support vector machine (SVM) was used to avoid hav- were selected due to stable lower RMSE. The number of vari- ing the problems of classification and regression (Wang et al. ables (mtry) was set to 4 at tuber formation, 9 at tuber bulking, 2020). Since the loss function gamma (g), the error penalty 4 at starch accumulation, and 1 at the combined stages, as it factor (C), and the choice of kernel function affect the perfor - yielded lower RMSE. In this study, the RF algorithms were mance of the SVM, the “Support Vactor Regression” func- implemented in the Python 3.10 environment, and the ‘Ran- tion in “sklearn” package was sued to implement the SVR dom Forest Regression” function in “sklearn” package was algorithms in the Python 3.10 environment in this study. The used. Other parameters were set as default values. commonly used radial basis kernel function (RBF) was applied in this study. Finally, a grid search was used to retrieve differ -2.8 Data Analysis ent combinations of g and C parameters. The best combination of g and C was 0.001 and 10.001 at tuber formation and tuber In this study, there were 132 sample points. Around 70% of bulking, and 0.001 and 20.001 at starch accumulation and the them were selected to train the four machine learning algo- combined stages for our data set, respectively. rithms, and 30% data were for validation. The performances of the different models of machine learning were evaluated by 2.7.3 Random Forest Regression Algorithm (RF) comparing the coefficients of determination (R ), root square error (RMSE) and mean square error (MSE) in predictions. The RF is an ensemble learning, that allows high accuracy and The higher the R and the lower the RMSE and MSE, the bet- generalization performance from the overall model (Shi et al. ter the precision and accuracy of the models. The R , RMSE 2020). Furthermore, the RF reduces overfitting by repeatedly and MSE were calculated according to Eqs. (1, 2, 3): putting back samples. In addition, RF takes into account for the � � � � ∑ ∑ 2 2 R =  y − y ∕ y − y (1) influence of input variables (name: mtry) to attenuate the auto- i i correlation between variables effectively (Peng et al. 2021). 1 3 PFG Fig. 4). However, the results showed that there was the sig- � � RMSE = y − y (2) nificant effect of the growth stages on the performances of i i i=1 the selected spectral indices (see Fig.  4). This study also found that the best spectral indices were GNDVI and CI , green � � ∑ especially at the growth stages of tuber formation and tuber MSE = y − y (3) i i bulking. At the starch accumulation and the combination i=1 of three stages, Ctr2 and CI showed significant cor - red-edge where y ,  y , and y are the observed, predicted, and observed i i relations with potato chlorophyll content, suggesting that mean values, respectively, and n is the sample size. chlorophyll content can be effectively estimated using these spectral indices. 3 Results 3.3 Estimation of Potato Chlorophyll Content Using Machine Learning Algorithms 3.1 Variations in Plant Chlorophyll Content in Different N Treatments and Growth Stages In this study, the chlorophyll content was predicted from highly correlated spectral indices at different growth stages The results in this study showed that the chlorophyll con- using four machine learning algorithms (SVR, RF, PLSR, tent increased from tuber formation to starch accumulation and RR) (see Fig. 5). The results showed that RF performed (see Fig. 3). The chlorophyll content in the training data set the highest accuracy in estimating chlorophyll content with ranged from 38.5 to 53.9 with a mean value of 45 and a CV R ranging from 0.96 to 0.97, RMSE ranging from 0.62 to value of 8.8%, while this in the testing data set ranged from 0.81, and MSE ranging from 0.38 to 0.65 for all growth 39.5 to 52.7 with a mean value of 45.5 and a CV value of stages. Except for the PLSR, all other machine learning 8.8%. The training and testing data sets showed a similar sta- algorithms showed the highest estimation accuracy at the tistical distribution, indicating that there were no potentially starch accumulation stage and followed by the tuber forma- biased estimations for model calibration and validation. tion stage. The results in Fig.  6 showed that RF was the optimal 3.2 Relationships Between Spectral Indices model for estimating potato chlorophyll content with R and Chlorophyll Content of Potato Plants of 0.61–0.83 that was greater than other models regard- less of growth stage, whereas RMSE (1.75–2.31) and MSE The results showed that the most selected spectral indi- (3.06–5.37) were smaller than other machine learning algo- ces were significantly correlated with the plant chloro- rithms. Compared to the RF, other machine learning algo- phyll content regardless of growth stage (see Table 2 and rithms showed the least performance, especially PLSR and RR algorithms due to the growth stage effect. Fig. 3 Variation of SPAD value 60 60 of potato at tuber formation, Training dataset Testing dataset tuber bulking and starch accu- 55 55 mulation stages of tuber forma- tion (T1), tuber bulking (T2) and starch accumulation (T3) 50 50 45 45 40 40 35 35 30 30 T1 T2 T3 T1 T2 T3 Growth stage Growth stage 1 3 SPAD SPAD PFG Fig. 4 Correlation analysis of spectral indices with SPAD value at different stages 1 3 PFG Fig. 4 (continued) 1 3 PFG 56 56 56 56 T1 T2 T3 T1+T2+T3 54 2 54 2 54 2 54 2 R =0.79 R =0.86 R =0.81 R =0.73 RMSE=1.56 RMSE=1.57 RMSE=1.58 RMSE=2.02 52 52 52 52 MSE=2.52 MSE=4.11 MSE=2.46 MSE=2.47 50 50 50 50 48 48 48 46 46 46 46 44 44 44 44 42 42 42 42 40 40 40 40 SVR SVR SVR SVR 38 38 38 38 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 56 56 56 56 T1 T2 T3 T1+T2+T3 54 2 54 2 54 2 54 2 R =0.96 R =0.97 R =0.96 R =0.96 RMSE=0.65 RMSE=0.62 RMSE=0.71 RMSE=0.81 52 52 52 52 MSE=0.42 MSE=0.38 MSE=0.51 MSE=0.65 50 50 50 50 48 48 48 48 46 46 46 44 44 44 42 42 42 42 40 40 40 40 RF RF RF RF 38 38 38 38 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 56 56 56 T1 T2 T3 T1+T2+T3 54 54 54 2 2 2 2 R =0.67 R =0.71 R =0.62 R =0.62 52 52 52 RMSE=2.29 RMSE=2.42 RMSE=1.96 RMSE=2.23 52 MSE=3.85 MSE=4.98 MSE=5.25 MSE=5.90 50 50 50 50 48 48 48 48 46 46 46 44 44 44 42 42 40 40 40 PLSR PLSR 40 PLSR PLSR 38 38 38 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 56 56 56 56 T1 T2 T3 T1+T2+T3 54 54 54 2 2 54 2 R =0.67 R =0.69 R =0.72 R =0.67 52 RMSE=1.98 52 RMSE=2.37 RMSE=1.98 RMSE=2.27 52 52 MSE=3.94 MSE=3.93 MSE=5.65 MSE=5.18 50 50 50 48 48 48 46 46 46 46 44 44 44 44 42 42 42 42 40 40 40 40 RR RR RR RR 38 38 38 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 Predicted SPAD value Predicted SPAD value Predicted SPAD value Predicted SPAD value Fig. 5 Relationships between the predicted value and observed value of training data set using four models at tuber formation stage (T1), tuber bulking stage (T2), starch accumulation stage (T3), and the combination of growth stages (T1 + T2 + T3) 3.4 Potato Chlorophyll Content Predictive Map increasing plant growth. During the tuber bulking stage, the number of pixels of medium chlorophyll content (see Fig. 7b, The RF model was used to map the predicted chlorophyll con- e) increased due to the rapid growth over potato tuber stage. During late reproduction, the nutrient delivery to the crop is tent of plants based on UAV multispectral images at the dif- ferent growth stages (see Fig. 7). Figure 7 shows that the esti- at its highest level, leading to the decrease in number of pixels due to low plant chlorophyll content in Fig. 7c, f. Meanwhile, mated value of chlorophyll content was relatively high in the tuber formation stage, whereas this gradually decreased with 1 3 Observed SPAD value Observed SPAD value Observed SPAD value Observed SPAD value PFG 54 56 Fig. 6 Validation of the per- SVR b RF formance of SVR (a), RF (b), R =0.54 R =0.61 RMSE=2.45 PLSR (c), and RR (d) at tuber RMSE=2.31 MSE=6.01 MSE=5.37 formation stage, tuber bulking 50 2 R =0.78 stage, starch accumulation stage R =0.79 RMSE=2.30 RMSE=2.14 and the combination of growth MSE=5.30 48 MSE=4.61 stages using the testing data set R =0.83 R =0.83 RMSE=2.13 RMSE=1.75 42 MSE=4.57 MSE=3.06 All R =0.74 All R =0.76 RMSE=2.35 RMSE=1.97 MSE=5.53 MSE=3.91 38 38 38 40 42 44 46 48 50 52 54 38 40 42 44 46 48 50 52 54 56 tuber formation tuber bulking starch accumulation PLSR RR R =0.49 2 R =0.50 RMSE=2.42 RMSE=2.27 MSE=5.89 MSE=5.18 50 2 R =0.74 2 R =0.71 RMSE=2.21 RMSE=2.26 48 MSE=4.89 MSE=5.13 46 46 44 44 R =0.80 R =0.71 RMSE=2.39 RMSE=2.37 42 42 MSE=5.75 MSE=5.63 All R =0.57 2 All R =0.70 40 40 RMSE=2.59 RMSE=2.19 MSE=6.72 MSE=4.80 38 38 38 40 42 44 46 48 50 52 54 38 40 42 44 46 48 50 52 54 Predicted SPAD value Predicted SPAD value chlorophyll content is significantly affected by the amount of and the accuracy of simple linear regression models was N applied, and the use of UAV images mapping allows for still poor (Lee et  al. 2020). Therefore, machine learn- real-time monitoring of the N status of the potato plant so that ing algorithms are promising as they can combine vari- nutrient management of the crop can be adjusted. ous spectral indices to improve the estimation accuracy (Qiu et  al. 2021). Numerous studies have already used RF, SVR, PLSR, and RR algorithms for estimations of 4 Discussion chlorophyll content (Tahir et  al. 2018). However, it is still unclear whether these algorithms are most suitable 4.1 Feasibility of Using UAV to Estimate Potato to fulfill the requirement (Li et al. 2020). Therefore, this Chlorophyll Content study compared four algorithms for precision and stabil- ity. The RF algorithm had higher R and lower RMSE The results in this study demonstrated that some spectral and MSE than others algorithms for chlorophyll content indices calculated from multispectral images were highly estimates in different growth stages, indicating that RF correlated with potato chlorophyll content, suggesting algorithm can provide accurate chlorophyll content esti- chlorophyll content of plants can be estimated by UAV mations. Our study also found that, since the model was multispectral image. Due to the different prediction per - trained using data set only from 2021, resulting in poor formance of spectral indices in different growth stages, models validation (see Fig.  8), there is a need to have a however, it was difficult to establish general models large amount of experimental data for training the models between different spectral indices and chlorophyll content, in future studies for the model stability. Recently, there were many research showed that using UAVs combined 1 3 Observed SPAD value Observed SPAD value PFG T1 (20 July, 2020) T2 (3 August, 2020) T3 (20 August, 2020) T1 (15 July, 2021) T2 (6 August, 2021) T3 (25 August, 2021) Fig. 7 Chlorophyll content map as indicated in SPAD values were derived by applying the RF model to the 8 spectral indices images for tuber formation (T1), tuber bulking (T2), and starch accumulation (T3) the machine learning algorithms could be monitored of 4.2 Analysis of Sensitive Bands important crop growth parameters, such as biomass, N, and leaf area index (Prado Osco et al. 2019; Abdelbaki Selection of the appropriate spectral bands from all wave- et al. 2021; Zhou et al. 2016), which further confirms the band to construct the spectral indices plays an important role feasibility our method and results. Compared with conven- in improving the model accuracy. At the tuber formation and tional wet chemical method, our method demonstrated the the tuber bulking stages, the green wavelength is closely advantages of convenience and speed, and compared with related to the potato chlorophyll content, which is in agree- the chlorophyll meters (SPAD-502), using UAVs could ment with previous study of the relationship between the achieve over a large area. Therefore, it was potential to spectral reflectance in the green region and plant chlorophyll estimate potato chlorophyll content using spectral indices content (Zhao et al. 2018). Hensen and Schjoerring (2003) calculated from UAV multispectral images. explained that chlorophyll a and b were more reflective in the green region, the green light region was often used to monitor chlorophyll content. However, this study suggests that the red-edge (680–800 nm) regions are important to 1 3 PFG 44 52 50.5 T1 T2 T3 T1+T2+T3 43 50 50.0 49.5 49.0 48.5 RR: R² = 0.53 R² = 0.35 R² = 0.38 R² = 0.35 48.0 40 RMSE = 2.38 RMSE = 7.22 RMSE = 2.44 RMSE = 6.49 40 MSE = 5.67 MSE = 52.25 MSE = 5.95 MSE = 42.23 38 38 47.5 42 44 46 48 50 52 40 42 44 46 48 50 40 42 44 46 38 40 42 44 46 48 50 52 T1 T2 T3 T1+T2+T3 46.6 43.8 46.4 43.7 46.2 46.0 43.6 42 45.8 43.5 45.6 43.4 45.4 PLSR: R² = 0.32 R² = 0.44 R² = 0.30 R² = 0.25 45.2 43.3 RMSE = 2.58 RMSE = 3.91 40 RMSE = 2.90 RMSE = 4.25 45.0 MSE = 15.34 MSE = 6.67 MSE = 8.46 MSE = 18.10 43.2 36 44.8 42 44 46 48 50 52 40 42 44 46 48 50 40 42 44 46 38 40 42 44 46 48 50 52 51 52 T1 T2 T3 T1+T2+T3 50 50 49 48 48 46 47 44 46 42 SVR: R² = 0.48 R² = 0.39 R² = 0.38 R² = 0.31 39 RMSE = 1.89 RMSE = 6.03 RMSE = 2.38 45 40 RMSE = 6.91 MSE = 3.58 MSE = 36.39 MSE = 5.69 MSE = 47.76 44 38 42 44 46 48 50 52 40 42 44 46 48 50 40 42 44 46 38 40 42 44 46 48 50 52 50 49.5 T1 T2 T3 T1+T2+T3 41.8 48 49.0 RF: 41.7 R² = 0.12 R² = 0.25 RMSE = 5.68 RMSE = 6.12 46 48.5 41.6 MSE = 32.28 MSE = 37.53 41.5 44 48.0 41.4 R² = 0.30 R² = 0.24 42 47.5 RMSE = 2.46 42 41.3 RMSE = 3.12 MSE = 6.10 MSE = 9.77 41.2 40 47.0 40 42 44 46 48 50 52 40 42 44 46 48 50 40 42 44 46 38 40 42 44 46 48 50 52 Observed SPAD value Observed SPAD value Observed SPAD value Observed SPAD value Fig. 8 Validation of the performance of RR, PLSR, SVR, and RF models at tuber formation stages (T1), tuber bulking stages (T2), starch accu- mulation (T3), and the combined stages (T1 + T2 + T3) using 2020 site data set potato chlorophyll content estimation at starch accumulation investigated at different growth stages to improve the accu- and the combination growth stages as well. Similar to the racy of estimation of chlorophyll content in potato crops. findings of Zheng et al. (2018b), the red-edge was a sensi- The spectral indices enhance the spectral characteristics tive region for potato chlorophyll content. Chang-Hua et al. sensitive to the biochemical properties of the vegetation by (2010) constructed a new red-edge spectral indices (RES) combining several bands and a specific formula format com - to estimate the chlorophyll content of sugar beet with high pared to the four bands. The wavebands of spectral indices accuracy. Therefore, the influence of growth stages on the contain some reference bands that are insensitive to chloro- selection of sensitive bands should be taken into account phyll content, and those bands are helpful to increase the sig- when spectral indices are applied. In future studies, the opti- nal-to-noise ratio (Yu et al. 2013). The spectral region used mization of the selected spectral indices should be further by the multispectral UAV is consistent with some satellite 1 3 Predicted SPAD value Predicted SPAD value Predicted SPAD value Predicted SPAD value PFG Fig. 9 Importance scores for spectral indices in the best RF model for predicting potato SPAD value at tuber formation stage (T1), tuber bulking stage (T2), starch accumulation stage (T3), and combined growth stages (T1 + T2 + T3) multispectral bands, such as the Sentinel-2. Therefore, the 5 Conclusions estimation framework constructed in this study can also be applied to satellite images for large-scale crop chlorophyll An integrated method combined with spectral indices and ran- content estimation (see Fig. 2). dom forest model was proposed to estimate chlorophyll content in potato crop. Accuracy and robustness of the RF model were 4.3 T he Eec ff t of the Number of Spectral Indices verified for each growth stage, such as tuber formation, tuber on the Model bulking, starch accumulation, and the combination of growth stages. The results of comparison of the RF model with SVR, Previous studies have also shown that predicting chlorophyll PLSR, and RR showed that the accuracy of RF outperformed content was affected by the use of variable importance plot at each stage. The correlation analysis demonstrated that in RF that eliminates spectral variable inputs. The contribu- most of spectral indices were significantly correlated with the tion of each spectral index to the RF model indicates the potato chlorophyll content, whereas the spectral indices with extent to which each of the spectral indices could predict the strongest correlation varied at different period. The CI green the chlorophyll content variations (see Table 2 and Fig. 9). and the green band highly contributed to the estimation model However, there is a need to analyze the performance of the during the tuber formation and tuber bulking stage. However, model for the reduction in the number of input variables the CI and the red edge band provided higher contribu- red edge and increase computational efficiency. By implementing the tion to the prediction model during the starch accumulation best 1 to 3 spectral indices into the RF model, only a slight and the combination of growth stages. decrease was observed in its performance (see Fig. 10). Sim- ilarly, Prado Osco et al. (2019) used the variable importance graph in RF model to eliminate spectral variable inputs to reduce the amount of spectral data computation and keep the model with better performance. 1 3 PFG 56 56 56 56 mtry: GNDVI mtry: GNDVI mtry: GDVI mtry: CI rededge T1 T3 T2 T1+T2+T3 54 54 54 54 2 2 R =0.94 R =0.97 R =0.89 RMSE=0.85 RMSE=0.69 52 52 52 RMSE=1.23 52 R =0.94 MSE=0.72 MSE=0.47 MSE=1.51 RMSE=0.97 50 50 50 50 MSE=0.95 48 48 48 2 46 46 46 R =0.43 RMSE=2.94 44 44 44 MSE=8.65 42 42 42 2 2 2 R =0.58 R =0.89 R =0.71 RMSE=2.61 RMSE=1.36 RMSE=2.35 40 training dataset 40 40 40 MSE=6.85 MSE=1.87 MSE=5.52 testing dataset 38 38 38 38 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 56 56 56 56 mtry: GNDVI, CI mtry: GDVI, CG mtry: GNDVI, CG green mtry: GNDVI, CI rededge T1 T2 T3 54 54 54 T1+T2+T3 2 2 2 R =0.94 R =0.97 R =0.89 R =0.96 52 52 52 52 RMSE=0.84 RMSE=1.22 RMSE=0.84 RMSE=0.81 MSE=0.70 MSE=1.51 MSE=0.70 50 50 50 MSE=0.66 48 48 48 46 2 46 46 R =0.43 RMSE=2.94 44 44 44 MSE=8.65 42 42 42 2 42 2 R =0.69 R =0.73 R =0.88 RMSE=2.41 RMSE=2.07 40 40 RMSE=1.43 40 training dataset 40 MSE=5.81 MSE=4.28 MSE=2.05 testing dataset 38 38 38 38 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 56 56 56 56 mtry: GNDVI, CVI, CI mtry: GDVI, DVI, CG mtry: GNDVI, CG, CI green mtry: GNDVI, CI , CI green green rededge T1 T2 T3 54 54 54 54 T1+T2+T3 2 2 R =0.95 R =0.97 R =0.89 52 52 RMSE=0.64 RMSE=1.22 R =0.96 RMSE=0.77 52 52 MSE=0.42 MSE=1.51 RMSE=0.82 MSE=0.59 50 50 50 50 MSE=0.68 48 48 48 48 46 2 46 46 46 R =0.52 RMSE=2.80 44 44 44 MSE=7.86 2 2 42 42 42 2 R =0.87 R =0.70 R =0.72 RMSE=1.51 RMSE=2.38 RMSE=2.11 40 training dataset 40 40 MSE=2.30 MSE=5.67 MSE=4.48 testing dataset 38 38 38 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 38 40 42 44 46 48 50 52 54 56 Predicted SPAD value Predicted SPAD value Predicted SPAD value Predicted SPAD value Fig. 10 Optimizing the number of predictors of spectral indices according to important scores at tuber formation stage (T1), tuber bulking stage (T2), starch accumulation stage (T3) and the combination of growth stages (T1 + T2 + T3) for training data set and testing data set Funding This research was funded by Programs for Key Science References and Technology Development of Inner Mongolia in 2019 and 2020 (2019GG248 and 2020GG0038), the National Natural Science Founda- Abdelbaki A, Schlerf M, Retzlaff R et al (2021) Comparison of crop tion of China (32160757). trait retrieval strategies using UAV-based VNIR hyperspectral imaging. Remote Sens 13(9):1748. https://doi. or g/10. 3390/ rs130 Open Access This article is licensed under a Creative Commons Attri- bution 4.0 International License, which permits use, sharing, adapta- Argenta G, Silva PRF, Sangoi L (2004) Leaf relative chlorophyll con- tion, distribution and reproduction in any medium or format, as long tent as an indicator parameter to predict nitrogen fertilization in as you give appropriate credit to the original author(s) and the source, maize. Ciência Rural 34:1379–1387. https:// doi. or g/ 10. 1590/ provide a link to the Creative Commons licence, and indicate if changes S0103- 84782 00400 05000 09 were made. The images or other third party material in this article are Bareth G, Aasen H, Bendig J, et al (2015) Low-weight and UAV-based included in the article's Creative Commons licence, unless indicated hyperspectral full-frame cameras for monitoring crops: spectral otherwise in a credit line to the material. If material is not included in comparison with portable spectroradiometer measurements. the article's Creative Commons licence and your intended use is not Unmanned aerial vehicles (UAVs) for multi-temporal crop surface permitted by statutory regulation or exceeds the permitted use, you will modelling. https:// doi. org/ 10. 1127/ pfg/ 2015/ 0256 need to obtain permission directly from the copyright holder. To view a Bendig J, Bolten A, Bareth G (2013) UAV-based imaging for multi- copy of this licence, visit http://cr eativ ecommons. or g/licen ses/ b y/4.0/ . temporal, very high resolution crop surface models to monitor crop growth variability. Unmanned aerial vehicles (UAVs) for 1 3 Observed SPAD value Observed SPAD value Observed SPAD value PFG multi-temporal crop surface modelling. https:// doi. org/ 10. 1127/ Li C, Chen P, Ma C et al (2020) Estimation of potato chlorophyll con- 1432- 8364/ 2013/ 0200 tent using composite hyperspectral index parameters collected by Cao Z, Cheng T, Ma X et al (2017) A new three-band spectral index an unmanned aerial vehicle. Int J Remote Sens 41(21):8176–8197. for mitigating the saturation in the estimation of leaf area index https:// doi. org/ 10. 1080/ 01431 161. 2020. 17577 79 in wheat. Int J Remote Sens 38(13):3865–3885. https:// doi. org/ Lu W, Lou J, Webster C et al (2021) Estimating construction waste 10. 1080/ 01431 161. 2017. 13061 41 generation in the Greater Bay Area, China using machine learn- Chang-Hua JU, Yong-Chao T, Xia YAO et al (2010) Estimating leaf ing. Waste Manage 134:78–88. https://doi. or g/10. 1016/j. w asman. chlorophyll content using red edge parameters. Pedosphere 2021. 08. 012 20(5):633–644. https:// doi. org/ 10. 1016/ S1002- 0160(10) 60053-7 Markwell J, Osterman JC, Mitchell JL (1995) Calibration of the Clevers JGPW, Kooistra L, Van den Brande MMM (2017) Using Sen- Minolta SPAD-502 leaf chlorophyll meter. Photosynth Res tinel-2 data for retrieving LAI and leaf and canopy chlorophyll 46(3):467–472. https:// doi. org/ 10. 1007/ BF000 32301 content of a potato crop. Remote Sens 9(5):405. https:// doi. org/ Martínez D, Guiamet J (2004) Distortion of the SPAD 502 chlorophyll 10. 3390/ rs905 0405 meter readings by changes in irradiance and leaf water status. Colorado JD, Calderon F, Mendez D et al (2020) A novel NIR-image Agronomie 24(1):41–46. https:// doi. org/ 10. 1051/ agro: 20030 60 segmentation method for the precise estimation of above-ground Moghimi A, Pourreza A, Zuniga-Ramirez G et  al (2020) A novel biomass in rice crops. PLoS ONE 15(10):e0239591. https:// doi. machine learning approach to estimate grapevine leaf nitrogen org/ 10. 1371/ journ al. pone. 02395 91 concentration using aerial multispectral imagery. Remote Sens Franzini M, Ronchetti G, Sona G et al (2019) Geometric and radio- 12(21):3515. https:// doi. org/ 10. 3390/ rs122 13515 metric consistency of parrot sequoia multispectral imagery for Mutanga O, Adam E, Cho MA (2012) High density biomass estimation precision agriculture applications. Appl Sci 9(24):5314. https:// for wetland vegetation using WorldView-2 imagery and random doi. org/ 10. 3390/ app92 45314 forest regression algorithm. Int J Appl Earth Obs Geoinf 18:399– Gitelson AA, Merzlyzk MN (1994) Quantitative estimation of chlo- 406. https:// doi. org/ 10. 1016/j. jag. 2012. 03. 012 rophyll-a using reflectance spectra: Experiments with autumn Peng L, Wang L, Ai XY et  al (2021) Forecasting tourist arrivals chestnut and maple leaves. J Photochem Photobiol B Biol 22:247– via random forest and long short-term memory. Cogn Comput 252. https:// doi. org/ 10. 1016/ 1011- 1344(93) 06963-4 13(1):125–138. https:// doi. org/ 10. 1007/ s12559- 020- 09747-z Gitelson AA, Buschmann C, Lichtenthaler HK (1999) The chlorophyll Prado Osco L, Marques Ramos AP, Roberto Pereira D et al (2019) fluorescence ratio F735/F700 as an accurate measure of the chlo- Predicting canopy nitrogen content in citrus-trees using random rophyll content in plants. Remote Sens Environ 69(3):296–302. forest algorithm associated to spectral vegetation indices from https:// doi. org/ 10. 1016/ S0034- 4257(99) 00023-1 UAV-imagery. Remote Sens 11(24):2925. https://doi. or g/10. 3390/ Gitelson AA, Yuri G, Merzlyark MN (2003) Relationships between rs112 42925 leaf chlorophyll content and spectral reflectance and algorithms Qi H, Wu Z, Zhang L et al (2021) Monitoring of peanut leaves chloro- for non-destructive chlorophyll assessment in higher plant phyll content based on drone-based multispectral image feature leaves. J Plant Physiol 160(3): 271–282. https:// doi. org/ 10. 1078/ extraction. Comput Electron Agric 187:106292. https:// doi. org/ 0176- 1617- 0088710. 1016/j. compag. 2021. 106292 Hansen PM, Schjoerring JK (2003) Reflectance measurement of can- Qiu Y, Garg D, Kim SM et al (2021) Machine learning algorithms opy biomass and nitrogen status in wheat crops using normalized to predict flow boiling pressure drop in mini/micro-channels difference vegetation indices and partial least squares regression. based on universal consolidated data. Int J Heat Mass Transf Remote Sens Environ 86(4):542–553. https:// doi. org/ 10. 1016/ 178:121607. https:// doi. org/ 10. 1016/j. ijhea tmass trans fer. 2021. S0034- 4257(03) 00131-7 121607 Houles V, Guerif M, Mary B (2007) Elaboration of a nitrogen nutrition Rafael MT, María SC, Sergio OM et al (2014) Environmental and indicator for winter wheat based on leaf area index and chloro- economic impact of Forest fires in Puerto Rico 2013–2014. Open phyll content for making nitrogen recommendations. Eur J Agron J for 5(04):353. https:// doi. org/ 10. 4236/ ojf. 2015. 54030 27(1):1–11. https:// doi. org/ 10. 1016/j. eja. 2006. 10. 001 Rouse JW, Haas RH, Schell JA (1974) Monitoring vegetation systems Inoue Y, Sakaiya E, Zhu Y et al (2012) Diagnostic mapping of canopy in the Great Plains with ERTS. NASA Spec Publ 351:309 nitrogen content in rice based on hyperspectral measurements. Shah SH, Angel Y, Houborg R et al (2019) A random forest machine Remote Sens Environ 126:210–221. https://doi. or g/10. 1016/j. rse. learning approach for the retrieval of leaf chlorophyll content in 2012. 08. 026 wheat. Remote Sens 11(8):920. https:// doi. or g/ 10. 3390/ rs110 Jordan CF (1969) Derivation of leaf-area index from quality of light 80920 on the forest floor. Ecology 50:663–666. https:// doi. org/ 10. 2307/ Shi L, Liu Q, Huang C et al (2020) Comparing pixel-based random 19362 56 forest and the object-based support vector machine approaches Kooistra L, Clevers JG (2016) Estimating potato leaf chlorophyll con- to map the quasi-circular vegetation patches using individual tent using ratio vegetation indices. Remote Sens Lett 7(6):611– seasonal fused GF-1 imagery. IEEE Access 8:228955–228966. 620. https:// doi. org/ 10. 1080/ 21507 04X. 2016. 11719 25https:// doi. org/ 10. 1109/ ACCESS. 2020. 30450 57 Kopačková-Strnadová V, Koucká L, Jelének J et al (2021) Canopy top, Singhal G, Bansod B, Mathew L et al (2019) Chlorophyll estimation height and photosynthetic pigment estimation using parrot sequoia using multi-spectral unmanned aerial system based on machine multispectral imagery and the unmanned aerial vehicle (UAV). learning techniques. Remote Sens Appl Soc Environ 15:100235. Remote Sens 13(4):705. https:// doi. org/ 10. 3390/ rs130 40705https:// doi. org/ 10. 1016/j. rsase. 2019. 100235 Lee H, Wang J, Leblon B (2020) Using linear regression, random for- Song C (2013) Optical remote sensing of forest leaf area index and ests, and support vector machine with unmanned aerial vehicle biomass. Prog Phys Geogr 37(1):98–113. https://doi. or g/10. 1177/ multispectral images to predict canopy nitrogen weight in corn. 03091 33312 471367 Remote Sens 12(13):2071. https:// doi. org/ 10. 3390/ rs121 32071 Sripada RP, Heiniger RW, White JG et al (2006) Aerial color infrared Li F, Miao Y, Hennig SD et al (2010) Evaluating hyperspectral veg- photography for determining early in-season nitrogen require- etation indices for estimating nitrogen concentration of winter ments in corn. Agron J 98:968–977. https://doi. or g/10. 2134/ ag ron wheat at different growth stages. Precis Agric 11(4):335–357. j2005. 0200 https:// doi. org/ 10. 1007/ s11119- 010- 9165-6 Tahir MN, Naqvi SZA, Lan Y et al (2018) Real time estimation of chlorophyll content based on vegetation indices derived from 1 3 PFG multispectral UAV in the kinnow orchard. Int J Precis Agric Aviat. Wu C, Niu Z, Tang Q et al (2008) Estimating chlorophyll content from https:// doi. org/ 10. 33440/j. ijpaa. 20180 101. 0001 hyperspectral vegetation indices: modeling and validation. Agric Tan K, Ma W, Wu F et al (2019) Random forest–based estimation of For Meteorol 148(8–9):1230–1241. https://doi. or g/10. 1016/j. ag rfo heavy metal concentration in agricultural soils with hyperspectral rmet. 2008. 03. 005 sensor data. Environ Monit Assess 191(7):1–14. https:// doi. org/ Wu J, Chen XY, Zhang H et al (2019) Hyperparameter optimization 10. 1007/ s10661- 019- 7510-4 for machine learning models based on Bayesian optimization. J Théau J, Gavelle E, Ménard P (2020) Crop scouting using UAV Electron Sci Technol 17(1):26–40. https://doi. or g/10. 11989/ JES T. imagery: a case study for potatoes. J Unman Veh Syst 8(2):99–1674- 862X. 80904 120 118. https:// doi. org/ 10. 1139/ juvs- 2019- 0009 Yu K, Li F, Gnyp ML et al (2013) Remotely detecting canopy nitrogen Tucker C, Elgin J, Mcmurtrey J (1979) Monitoring corn and soy- concentration and uptake of paddy rice in the Northeast China bean crop development with hand-held radiometer spectral Plain. ISPRS J Photogramm Remote Sens 78:102–115. https:// data. Remote Sens Environ 8:237–248. https:// doi. org/ 10. 1016/ doi. org/ 10. 1016/j. isprs jprs. 2013. 01. 008 0034- 4257(79) 90004-X Yu K, Lenz-Wiedemann V, Chen X et al (2014) Estimating leaf chlo- Uddling J, Gelang-Alfredsson J, Piikki K et al (2007) Evaluating the rophyll of barley at different growth stages using spectral indices relationship between leaf chlorophyll concentration and SPAD- to reduce soil background and canopy structure effects. ISPRS 502 chlorophyll meter readings. Photosynth Res 91(1):37–46. J Photogramm Remote Sens 97:58–77. https:// doi. org/ 10. 1016/j. https:// doi. org/ 10. 1007/ s11120- 006- 9077-5isprs jprs. 2014. 08. 005 Verrelst J, Muñoz J, Alonso L et al (2012) Machine learning regression Zhao B, Duan A, Ata-Ul-Karim ST et al (2018) Exploring new spec- algorithms for biophysical parameter retrieval: opportunities for tral bands and vegetation indices for estimating nitrogen nutrition Sentinel-2 and-3. Remote Sens Environ 118:127–139. https://d oi. index of summer maize. Eur J Agron 93:113–125. https://d oi.o rg/ org/ 10. 1016/j. rse. 2011. 11. 00210. 1016/j. eja. 2017. 12. 006 Viljanen N, Honkavaara E, Näsi R et al (2018) A novel machine learn- Zheng Q, Huang W, Cui et al (2018a) New spectral index for detecting ing method for estimating biomass of grass swards using a photo- wheat yellow rust using Sentinel-2 multispectral imagery. Sensors grammetric canopy height model, images and vegetation indices 18(3):868. https:// doi. org/ 10. 3390/ s1803 0868 captured by a drone. Agriculture 8(5):70. https://doi. or g/10. 3390/ Zheng T, Liu N, Wu L et al (2018b) Estimation of chlorophyll con- agric ultur e8050 070 tent in potato leaves based on spectral red edge position. IFAC- Wolanin A, Camps-Valls G, Gómez-Chova L et al (2019) Estimating PapersOnLine 51(17):602–606. https:// doi. org/ 10. 1016/j. ifacol. crop primary productivity with Sentinel-2 and Landsat 8 using 2018. 08. 131 machine learning methods trained with radiative transfer simu- Zhou X, Zhu X, Dong Z et al (2016) Estimation of biomass in wheat lations. Remote Sens Environ 225:441–457. https:// doi. org/ 10. using random forest regression algorithm and remote sensing data. 1016/j. rse. 2019. 03. 002 Crop J 4(3):212–219. https:// doi. org/ 10. 1016/j. cj. 2016. 01. 008 1 3

Journal

"PFG – Journal of Photogrammetry, Remote Sensing and Geoinformation Science"Springer Journals

Published: Sep 23, 2022

Keywords: Chlorophyll content; Machine learning; Multispectral images; Potato; Unmanned aerial vehicle (UAV)

References