TY - JOUR
AU1 - Ara,, Iffat
AU2 - Harrison, Matthew, Tom
AU3 - Whitehead,, Jason
AU4 - Waldner,, François
AU5 - Bridle,, Kerry
AU6 - Gilfedder,, Louise
AU7 - Marques da Silva,, José
AU8 - Marques,, Francisco
AU9 - Rawnsley,, Richard
AB - Abstract Seasonal pasture monitoring can increase the efficiency of pasture utilization in livestock grazing enterprises. However, manual monitoring of pasture over large areas is often infeasible due to time and financial constraints. Here, we monitor changes in botanical composition in Tasmania, Australia, through application of supervised learning using satellite imagery (Sentinel-2). In the field, we measured ground cover and botanical composition over a 12-month period to develop a supervised classification approach used to identify pasture classes. Across seasons and paddocks, the approach predicted pasture classes with 75–81 % accuracy. Botanical composition varied seasonally in response to biophysical factors (primarily climate) and grazing behaviour, with seasonal highs in spring and troughs in autumn. Overall, we demonstrated that 10-m multispectral imagery can be reliably used to distinguish between pasture species as well as seasonal changes in botanical composition. Our results suggest that farmers and land managers should aim to quantify within-paddock variability rather than paddock average cover, because the extent and duration of very low ground cover puts the paddock/field at risk of adverse grazing outcomes, such as soil erosion and loss of pasture biomass, soil carbon and biodiversity. Our results indicate that satellite imagery can be used to support grazing management decisions for the benefit of pasture production and the improvement of environmental sustainability. 1. INTRODUCTION Pasture botanical composition is directly linked with overall sward productivity, with productivity varying seasonally according to pasture ecotype, resource availability and phenology. In some cases, swards with diversified botanical composition that include grasses, legumes and herbs may be more drought-tolerant and prevent weed incursions (Sanderson et al. 2005; Bell et al. 2015; Pembleton et al. 2016). Quantification and monitoring of botanical composition may thus translate to sward productivity, having implications for livestock feed supply. Increasing livestock utilization (consumption) of pastures often directly translates with increased profitability, because vegetative biomass is one of the cheapest forms of feed (Harrison et al. 2012a, b, 2016b; Ho et al. 2014; Chang-Fung-Martel et al. 2018). Conventionally, pasture quanta and botanical composition of intensively grazed pastures in southern Australia have been assessed visually, with farm managers often relying on knowledge gained through their own heuristics or professional networks to distinguish between pasture species. Although ground-based methods are useful for pasture monitoring over small scales, such methods are often laborious, time-consuming and sometimes expensive (Xu et al. 2008; Harrison et al. 2014b; Alcock et al. 2015). Reliable botanical composition classification can be further hindered by substantive pasture variation within a paddock, increasing the number of assessments required to gain an accurate measure of intra-paddock pasture variability. Remote sensing and Geographic Information System (GIS) approaches coupled with predictive algorithms may provide a viable alternative to within-field pasture monitoring and pasture species classification (Ali et al. 2016). For example, various authors have used satellite imagery time series (moderate resolution imaging spectroradiometer—MODIS; NOAA advanced very high-resolution radiometer (AVHRR) and Landsat) to obtain vegetation indices and phenological metrics to help understand pasture condition (Davidson 2008; Aguiar et al. 2010; Müller et al. 2015) and pasture biomass quanta (Wylie et al. 1991; Hanna et al. 1999; Kogan et al. 2004; Xu et al. 2008; Shuang et al. 2012; Jin et al. 2014; Barrachina et al. 2015). Zengeya et al. (2013) used WorldView-2 multispectral imagery to map pasture quality and showed that such imagery enhanced pasture quality estimation and mapping. Schaefer and Lamb (2016) used vehicle-mounted light detection and ranging (LiDAR) measurements to improve estimation of pasture biomass and found that combined LiDAR and active optical reflectance sensors contributed to better estimation of pasture biomass. However, many of these approaches do not have either the temporal (sub-weekly) or spatial (<20 m2) resolution necessary to enable pasture species cover or botanical composition quantification to provide information of sufficient resolution to facilitate grazing management decision-making at the paddock level. Sensors carried by hand, land or aerial vehicle have been used to distinguish between pasture species at the paddock level. For example, multispectral and hyper spectral sensors as well as multispectral radiometers have been used to assess pasture quality with reasonable reliability over a small area (Pullanagari 2011; Pullanagari et al. 2012; Cletah and Onisimo 2017b). These studies have shown that pasture quality can be predicted using non-destructive sampling. It remains to be seen whether multispectral imagery from satellite can be used to monitor pasture cover and botanical composition over an extended duration (several months or years) at paddock/field scales (e.g. 50–100 ha). Satellite imagery research on pastures has classically been aimed at much larger scales such as the regional, continental or global levels (Vickery et al. 1997; Crabbe et al. 2020). With the rise of cloud computing, big data analytics and machine learning (Pal and Mather 2003; Pal 2005; Mountrakis et al. 2011; Belgiu and Drăguţ 2016), vegetation classification for specific contexts has increased in popularity, including supervised classification of pixels (Toivonen and Luoto 2003; Weiers et al. 2004; Baldi et al. 2006; Miehe et al. 2011; Reiche et al. 2012; Gómez et al. 2016) and maximum likelihood classifiers (MLC). Despite this, reliable prediction of pasture and other vegetation cover at the regional scale have had limited success due to lack of ability to distinguish between land-cover attributes (Brenner et al. 2012; Rufin et al. 2015). Such results suggest a need to mask non-pasture features (e.g. trees, fences, water bodies, infrastructure, etc.) in studies that aim to distinguish between pasture species at the paddock level. Other studies using multispectral optical imagery have been conducted to distinguish between pasture functional types (C3 and C4 species) (Liu et al. 2015; Cletah and Onisimo 2017a, b); such studies have been performed using a wide array of techniques (e.g. discriminant analysis, support vector machine (SVM)) with accuracy ranging from 82 to 98 %, but in most cases have again been conducted at scales much larger than a paddock/field. Taken together, such studies highlight a dearth of information relating to the reliability and practicality of using multispectral imagery to assess pasture cover and composition at the paddock scale using simple classifiers such as MLC. The inception of the Sentinel-2 satellite constellation (ESA 2020) provides a unique opportunity to analyse changes in pasture cover and composition at spatial and temporal scales sufficient for decision-making at the paddock scale. We hypothesized that high-resolution multispectral data from Sentinel-2 could be used to model seasonal changes in pasture growth and composition at the paddock scale (i.e. 50–100 ha). Our objectives were thus to (i) characterize spatial pasture variability at the paddock level and (ii) use supervised classification to map key pasture species classes and observe how such classes change seasonally. 2. MATERIALS AND METHODS 2.1 Study area The study was located at Triabunna, Tasmania, Australia (Fig. 1). Triabunna has a cool maritime climate; mean monthly temperature oscillates from a minimum of 6.8 °C in July to a maximum of 16.8 °C in January; rainfall is equi-seasonal with monthly and annual average values of 56 and 665 mm, respectively (BoM 2020). Soils on the study area were primarily brown Chromosols and Eutrophic Brown Dermosols (DPIPWE 2000; Isbell 2016), while topography varied from undulating to hilly. Field samples collected on a commercial wool enterprise on four case study paddocks: Balsey Hill (67.52 ha), Cottage Hill (118.6 ha), Bouganville (58.3 ha) and Lords (73.62 ha) (Fig. 1, shaded areas). Figure 1. Open in new tabDownload slide Location of study area including individual paddock boundaries and paddock names, shaded areas represent study paddocks. The left-hand panel shows the farm location in the Australian State of Tasmania. Figure 1. Open in new tabDownload slide Location of study area including individual paddock boundaries and paddock names, shaded areas represent study paddocks. The left-hand panel shows the farm location in the Australian State of Tasmania. 2.2 Field measurement of pasture types and satellite image dates Field measurement of pasture cover following Anderson et al. (2011) was collected in the study paddocks during 2019 and 2020 (Table 1). Dates of field measurements were aligned with the nearest Sentinel-2 image date (Table 1). Locations of field measurements were undertaken with a multi-satellite handheld GPS unit (GARMIN GPS MAP 66S GPS DEVICE) with ~2 m accuracy. Pasture growth and botanical composition was monitored for 3-month periods, including summer (December–February), autumn (March–May), winter (June–August) and spring (September–November). Summer starts in December and finishes in February in the following year (e.g. ‘summer 2017/2018’ is the summer adjoining 2017 and 2018. Table 1. Seasonal alignment of field measurements with Sentinel-2 image dates. Season . Field measurement periods . Sentinel-2 image date . Winter 2019 13–29 August 2019 23 August 2019 Spring 2019 15–18 October 2019 22 October 2019 Summer 2019/2020 7–12 January 2020 30 January 2020 Autumn 2020 24–27 April 2020 24 April 2020 Winter 2020 24–29 July 2020 23 July 2020 Season . Field measurement periods . Sentinel-2 image date . Winter 2019 13–29 August 2019 23 August 2019 Spring 2019 15–18 October 2019 22 October 2019 Summer 2019/2020 7–12 January 2020 30 January 2020 Autumn 2020 24–27 April 2020 24 April 2020 Winter 2020 24–29 July 2020 23 July 2020 Open in new tab Table 1. Seasonal alignment of field measurements with Sentinel-2 image dates. Season . Field measurement periods . Sentinel-2 image date . Winter 2019 13–29 August 2019 23 August 2019 Spring 2019 15–18 October 2019 22 October 2019 Summer 2019/2020 7–12 January 2020 30 January 2020 Autumn 2020 24–27 April 2020 24 April 2020 Winter 2020 24–29 July 2020 23 July 2020 Season . Field measurement periods . Sentinel-2 image date . Winter 2019 13–29 August 2019 23 August 2019 Spring 2019 15–18 October 2019 22 October 2019 Summer 2019/2020 7–12 January 2020 30 January 2020 Autumn 2020 24–27 April 2020 24 April 2020 Winter 2020 24–29 July 2020 23 July 2020 Open in new tab To enable repeated measurements over the course of the experiment on the same points, field measurements were conducted visually following Tothill et al. (1992) and Waite (1994). Botanical composition was measured in the field using 100 mm × 100 mm quadrats placed on the same latitude and longitude for each temporal measurement shown in Table 1. Measurement points were conducted across a range of botanical compositions, topographies and soil types to gauge the full spectrum of biophysical conditions (Fig. 2). Figure 2. Open in new tabDownload slide Sentinel-2 images (true colour, red, green and blue) of study paddocks: Balsey Hill, Cottage Hill, Bouganville, and Lords on 22 October 2019. Locations of field measurements are indicated by red dots (dot size not drawn to scale). Non-pasture areas including trees, bush, wetlands and non-grazing areas were excluded from the analysis (white areas within the paddock). Image acquired in July 2020. Figure 2. Open in new tabDownload slide Sentinel-2 images (true colour, red, green and blue) of study paddocks: Balsey Hill, Cottage Hill, Bouganville, and Lords on 22 October 2019. Locations of field measurements are indicated by red dots (dot size not drawn to scale). Non-pasture areas including trees, bush, wetlands and non-grazing areas were excluded from the analysis (white areas within the paddock). Image acquired in July 2020. Pasture botanical composition consisted of introduced and native pasture ecotypes in the study area (Table 2). Improved pasture species were those sown in the study paddocks and sourced from commercial pasture seed suppliers. Botanical composition measured in the field was thus grouped into improved, native or mixed, with improved and native classes having a minimum of 20 % of the said botanical composition and ‘mixed’ having 20 % minimum of both improved and native classes. Each point was also subject to a 30 % minimum green cover threshold to ensure botanical composition spectra were sufficient to enable classification and prediction from satellite imagery. Table 2. Native and introduced pasture species used in this study. A third pasture class (mixed pasture species) was defined as having a minimum of 20 % of both improved and native species. Improved species . Native species . Phalaris (Phalaris aquatica) Kangaroo grass (Themeda triandra) Cocksfoot (Dactylis glomerata) Wallaby grass (Austrodanthonia caespitosa) Perennial ryegrass (Lolium perenne) Silver tussock grass (Poa labillardierei) Fescue (Festuca arundinacea) Subterranean clover (Trifolium subterraneum) Improved species . Native species . Phalaris (Phalaris aquatica) Kangaroo grass (Themeda triandra) Cocksfoot (Dactylis glomerata) Wallaby grass (Austrodanthonia caespitosa) Perennial ryegrass (Lolium perenne) Silver tussock grass (Poa labillardierei) Fescue (Festuca arundinacea) Subterranean clover (Trifolium subterraneum) Open in new tab Table 2. Native and introduced pasture species used in this study. A third pasture class (mixed pasture species) was defined as having a minimum of 20 % of both improved and native species. Improved species . Native species . Phalaris (Phalaris aquatica) Kangaroo grass (Themeda triandra) Cocksfoot (Dactylis glomerata) Wallaby grass (Austrodanthonia caespitosa) Perennial ryegrass (Lolium perenne) Silver tussock grass (Poa labillardierei) Fescue (Festuca arundinacea) Subterranean clover (Trifolium subterraneum) Improved species . Native species . Phalaris (Phalaris aquatica) Kangaroo grass (Themeda triandra) Cocksfoot (Dactylis glomerata) Wallaby grass (Austrodanthonia caespitosa) Perennial ryegrass (Lolium perenne) Silver tussock grass (Poa labillardierei) Fescue (Festuca arundinacea) Subterranean clover (Trifolium subterraneum) Open in new tab 2.3 Sentinel-2 image analysis Sentinel imagery was downloaded from the European Space Agency (ESA) Scientific Hub in May–July 2020 (ESA 2020). Sentinel 2A (S2A) was launched on 23 June 2015 and Sentinel 2B (S2B) on 7 March 2017, with images generated every 10 days from 2015 to present. We considered Level 1C Top of Atmosphere (TOA) images from December 2017 (summer 2017/2018) to monitor seasonal pasture growth [see Supporting Information—Table S1]. Sentinel-2 S2A and S2B used in the present analysis were based on proximity of the image acquisition date with the corresponding field measurement date (Table 1) to classify botanical composition. Only cloud-free images at the paddock scale were used: TOA images were ortho-rectified, radiometrically and geometrically corrected in the ESA hub. The Sentinel-2 Multispectral Instrument (MSI) comprises 12 spectral bands, ranging from Visible and Near-Infrared (VNIR) to Shortwave Infrared (SWIR) wavelengths along a 290-km orbital swath, of which there are four VNIR bands (B2, B3, B4 and B8) at 10-m, four red edge bands (B5, B6, B7 and B8a) and two SWIR bands (B11 and B12) at 20-m, and three other bands (aerosol, water vapour, cirrus SWIR B1, B9 and B10, respectively) at 60-m resolution [see Supporting Information—Table S2]. The study workflow is shown in Fig. 3. Input data are described above. In the processing stage, we stacked all bands and resized images at 10-m resolution using the nearest neighbour algorithm. We masked images at the paddock level (Fig. 2) to calculate normalized difference vegetation index (NDVI); these data were collected from January 2018 to February 2020 using the AgroInsider platform (AgroInsider 2020). We defined field data using botanical composition and percentage cover of pasture species as improved (I), native (N) and mixed (M) (Table 2). In the analysis stage, we developed training data based on TOA reflectance for each pasture species class and used these data to perform maximum likelihood classification (MLC). Refined field data were then used for accuracy assessment of the predictive MLC outcomes. Seasonal changes in botanical composition were examined by running several MLC and developing a spectral library (based on the training data developed in the analysis stage). Seasonal botanical composition changes are presented over five seasons (winter 2019 to winter 2020) (Table 1). Figure 3. Open in new tabDownload slide Schematic of study workflow. Input data includes 12 TOA bands from Sentinel-2 and field samples. Processing included stacking of bands, paddock-level masking and refining the field sampling. Analysis included model calibration, supervised classification and accuracy assessment, while change detection included development of the spectral library and identification of botanical composition changes. Figure 3. Open in new tabDownload slide Schematic of study workflow. Input data includes 12 TOA bands from Sentinel-2 and field samples. Processing included stacking of bands, paddock-level masking and refining the field sampling. Analysis included model calibration, supervised classification and accuracy assessment, while change detection included development of the spectral library and identification of botanical composition changes. We used Arc Geographic Information System (ArcGIS) 10.6 software for construction and manipulation of paddock shapefiles, including non-pasture region exclusion, resampling, layer stacking, masking, re-sizing images, layout and mapping classified images, NDVIs and the Environment for Visualizing Images (ENVI) 5.4 software for other processing, enhancement, classification and accuracy assessment. 2.4 Spatio-temporal phenological variation Intra-paddock spatial and temporal phenological variation in pasture cover was analysed using NDVI (Tucker 1979) generated from 10-m resolution red (B4) and near-infrared (B8) spectral bands of Sentinel-2. Normalized difference vegetation index (Equation 1) is a common vegetation index for studying pasture phenology (Bella et al. 2004; Boschetti et al. 2007; Flynn et al. 2008; Asher et al. 2018; Serrano et al. 2018). NDVI=(NIR−red)(NIR+red)(1) where red corresponds to band 4 (30 nm) and NIR corresponds to band 8 (842 nm) of Sentinel-2. Normalized difference vegetation index ranges from −1 to +1, with cloud, bare soil and sand showing negative/low NDVI (−1 to <0.1), shrubs and grasslands moderate NDVI (0.2 to 0.5) and dense vegetation/grassland high NDVI (0.6 to 0.9). Based on this classification, we divided the NDVI range 0.1–0.9 into five class, very low (<0.20), low (0.20–0.31), medium (0.31–0.35), high (0.35–0.40) and very high (>4.0) to represent distinctive NDVI variation according to these classes over study paddocks. 2.5 Botanical composition training data Botanical composition was classified using training data approaches outlined by Roth et al. (2012), Veganzones and Grana (2008), and Xu et al. (2015). Training sites were selected using a range of geolocated regions of improved, native and mixed composition in the study paddocks. Spectra for training samples were developed for pasture classes in autumn 2020, then merged (union) to three regions of interest (ROIs) representing improved, native and mixed classes (Fig. 4). The same process was applied to develop ROI for winter in 2020. Eighteen and 20 training samples were developed for autumn 2020 and winter 2020, respectively. In each case, two sets of ROI pairs were used to discriminate spectra for individual pasture classes using a separability threshold (Csendes and Mucsi 2016) of at least 1.95 for each ROI pair. Separability was computed using the Jeffries-Matusita (JM) distance and the transformed divergence algorithms computed following Richards and Richards (1999) using pasture class means and covariance matrices; a separability value of 2.0 corresponds to complete separability. Mean ROI for autumn and winter were used as a reference to develop the spectral library. Reflectance of each pixel was assumed to be linearly proportional to botanical composition (or ROI) (Roberts et al. 2002; Quintano et al. 2012). This approach was then used for pasture classification in other seasons (Table 1). Figure 4. Open in new tabDownload slide Training spectra for botanical composition from Sentinel-2 (image acquisition date: 24 April 2020). Twelve Sentinel-2 bands were used to collect the TOA reflectance spectra. Figure 4. Open in new tabDownload slide Training spectra for botanical composition from Sentinel-2 (image acquisition date: 24 April 2020). Twelve Sentinel-2 bands were used to collect the TOA reflectance spectra. 2.6 Maximum likelihood classification Supervised classification was conducted using the MLC in ENVI software (Russell and Norvig 2002; Mohri et al. 2018). Maximum likelihood classification computes the probability that a pixel belongs to a given class (input), assuming that spectra for each class within each band are normally distributed to produce a classification map (output) (Richards and Richards 1999). Each pixel was allocated to the class having the maximum probability (i.e. the maximum likelihood). 2.7 Accuracy assessment Akin to model validation (Harrison et al. 2012b, c; Bell et al. 2015; Pembleton et al. 2016), the MLC modelling approach was evaluated in ENVI using a confusion (or error) matrix accuracy assessment (Lillesand et al. 2015) by comparing modelled MLC classes to those measured on ground. Overall, 24 and 21 field measurements were used in the accuracy assessment for autumn 2020 and winter 2020, respectively. The confusion matrix included omission errors, commission errors, producer’s accuracy and user accuracy and contingency tables are presented in the Supporting Information related to each season. Following Congalton (1991), the overall accuracy (Oc) was calculated for each classification using Equation (2). Oc=∑irxijN×100(2) where r represents number of rows in the non-square matrix, xij is the total number of correctly classified pixels in row i and column j. 3. RESULTS 3.1 Seasonal paddock-level pasture cover Paddock average pasture cover was generally greatest in spring and least in autumn (Fig. 5) in line with growth rates typically seen for temperate pastures (Harrison et al. 2014b, 2017; Alcock et al. 2015). This pattern was consistent across all study paddocks in 2018. Persistently dry conditions in 2019 (total rainfall in 2019 was 457 mm compared with the long-term average of 613 mm) caused greater pasture senescence and low growth rates, resulting in low ground cover in 2019 until summer 2019–20. However, the situation improved in 2020, particularly for autumn, 2020 and winter 2020 for all study paddocks. Figure 5. Open in new tabDownload slide Temporal variation of seasonal pasture cover based on NDVI from Sentinel-2 images from January 2018 to August 2020 (x-axis represents delineates each season) for paddocks: Balsey Hill (A), Cottage Hill (B), Bouganville (C) and Lords (D). Figure 5. Open in new tabDownload slide Temporal variation of seasonal pasture cover based on NDVI from Sentinel-2 images from January 2018 to August 2020 (x-axis represents delineates each season) for paddocks: Balsey Hill (A), Cottage Hill (B), Bouganville (C) and Lords (D). Seasonal and spatial variation in pasture cover during 2019 and 2020 are shown in Fig. 6. All paddocks exhibited a high degree of intra-paddock pasture cover variation, indicating variability in growing conditions due to soil type, hill aspect and long-term grazing preferences, such as location with which animals camped. Figures 7 and 8 also underscore the need to quantify within-paddock pasture variability in grazing management contexts, rather than paddock averages. For example, in winter and spring of 2019, pasture cover varied from less than 0.20 to more than 0.4 in all paddocks, suggesting a high degree of spatial variability in pasture cover. In general, the lowest and highest NDVI occurred in summer 2019/2020 and autumn 2020, respectively, because many areas in autumn had NDVI > 0.4. Figure 6. Open in new tabDownload slide Spatial variation of seasonal pasture cover based on NDVI, derived from Sentinel-2 images for Balsey Hill (A), Cottage Hill (B), Bouganville (C) and Lords (D) paddocks. Figure 6. Open in new tabDownload slide Spatial variation of seasonal pasture cover based on NDVI, derived from Sentinel-2 images for Balsey Hill (A), Cottage Hill (B), Bouganville (C) and Lords (D) paddocks. Figure 7. Open in new tabDownload slide Predictions (green, blue and yellow represent improved, native and mixed, respectively) of pasture classes from the supervised classification compared with locations of measured pasture species shown by red points (I, N and M represent improved, native and mixed, respectively) for autumn 2020 on Balsey Hill (A), Cottage Hill (B), Bouganville (C) and Lords (D). Figure 7. Open in new tabDownload slide Predictions (green, blue and yellow represent improved, native and mixed, respectively) of pasture classes from the supervised classification compared with locations of measured pasture species shown by red points (I, N and M represent improved, native and mixed, respectively) for autumn 2020 on Balsey Hill (A), Cottage Hill (B), Bouganville (C) and Lords (D). Figure 8. Open in new tabDownload slide Predictions (green, blue and yellow represent improved, native and mixed, respectively) of pasture classes from supervised classification compared with the locations of measured pasture species data shown by red points (I, N and M represent improved, native and mixed, respectively) for winter 2020 on Balsey Hill (A), Cottage Hill (B), Bouganville (C) and Lords (D). Figure 8. Open in new tabDownload slide Predictions (green, blue and yellow represent improved, native and mixed, respectively) of pasture classes from supervised classification compared with the locations of measured pasture species data shown by red points (I, N and M represent improved, native and mixed, respectively) for winter 2020 on Balsey Hill (A), Cottage Hill (B), Bouganville (C) and Lords (D). 3.2 Using supervised classification to distinguish between pasture classes The distribution of native, introduced and mixed pasture classes were assessed at a pixel resolution of 10 m2 using supervised classification (MLC) in autumn 2020 (Fig. 7) and winter 2020 (Fig. 8). The majority of native pasture classes occurred in the east of Balsey Hill (Fig. 7A), in the north and centre of Cottage Hill and Lords (Fig. 7B) and on the north-eastern flats of Bouganville (Fig. 7C). Native pasture species tended to be more prolific around trees, shelter belts or on paddock hills. In winter, there were more introduced pastures, similar mixed and fewer native species relative to the distribution in autumn (Fig. 7), indicating seasonal differences in phenology of the key pasture classes: introduced species being more dominant in winter and native species being relatively more abundant in autumn. Accuracy of the supervised classification was assessed using the percentage of correctly classified pixels in a confusion matrix (locations of field measurements are shown in Fig. 2). The supervised classification had an accuracy of 75 % and 81 % in autumn and winter, respectively (contingency table presented in Supporting Information—Tables S3 and S4), indicating adequate performance of the supervised classification approach in predicting pasture classes. Many of the predicted native and mixed pasture classes aligned well with the points measured on the ground, although some of the modelled improved pasture classes did not align with those measured on ground. This may be because these points were located in areas of high pasture spatial heterogeneity, making it difficult to obtain a reliable prediction of pasture classes. To examine how pasture species distribution changed relative to previous season(s), we used the spectral library (developed based on the mean spectra for each pasture class in autumn and winter 2020) to simulate pasture species distribution winter 2019, spring 2019 and summer 2019/2020 (data shown in Fig. 9 and Supporting Information—Figs S1–S3). Figure 9. Open in new tabDownload slide Cumulative pasture area of all study paddocks from winter 2019 to autumn 2020. Error bars show one standard error of the mean. Figure 9. Open in new tabDownload slide Cumulative pasture area of all study paddocks from winter 2019 to autumn 2020. Error bars show one standard error of the mean. Overall, the supervised classification approach used here had significant ability to distinguish between native and improved ROI pairs, and between native and mixed ROI pairs, with a maximum separability threshold of 1.99 being achieved in both cases. However, the classification between improved and mixed species could be improved notwithstanding that considerable separability (1.97) between ROI pairs of these pasture classes was achieved. These results suggest that a large part of the overall accuracy obtained in the assessment (75–81 %) was due to reliable classification of native and mixed pasture. 3.3 Seasonal changes in pasture class distributions Seasonal changes in the cumulative areas of pasture classes across all paddocks are shown in Fig. 9. In autumn 2020, the improved and native pasture classes were at their highest cumulative areas of 38.2 and 44.7 ha, respectively. The lowest cumulative area for the improved species was observed in summer 2019/2020 (21.3 ha) and for the native class in winter (24.1 ha). The total area for mixed pasture classes varied little, from a maximum in winter 2019 of 156.4 ha to a minimum in autumn 2020 of 132.1 ha. 4. DISCUSSION 4.1 Quantifying intra-paddock pasture variability for sustainable grazing outcomes Here, we used high-resolution multispectral imagery to examine within-paddock pasture cover variability and pasture species composition. Our results show that grazing managers need to measure paddock-level pasture variability (Fig. 6) rather than paddock average cover, because the extent of very low ground cover (i.e. the area of ground below the paddock average) puts the paddock at risk of adverse grazing outcomes, such as soil erosion, loss of pasture biomass, soil carbon and biodiversity. Overgrazing has become an environmental concern, having adverse effects on soil erosion (Rowntree et al. 2004), soil organic matter (Conant and Paustian 2002), increasing land degradation (Sonneveld et al. 2010) and loss of valuable biodiversity (Fedrigo et al. 2018). Budd and Thorpe (2009) indicate that appropriate grazing management leads to reproduction and regeneration of vegetation, provision of wildlife habitat and potentially greater economic benefit through natural capital. Grazing managers should thus pay attention to paddock areas with the least ground cover and manage grazing according to these areas for long-term sustainability (Chang-Fung-Martel et al. 2018). Pasture growth rates vary due to abiotic (soil, climate, elevation) and biotic conditions (pasture genetic potential, insects and other wildlife, soil microbiota) as well as grazing management which together highlight the need for regular paddock-level pasture cover assessment (Harrison et al. 2016b). In the cool temperate climates of Tasmania, pasture cover peaks in spring due to increasing seasonal warmth, high rainfall and good soil moisture conditions, and troughs in late summer/autumn due to lack of rainfall and soil moisture (Harrison et al. 2016a). Such trends concur with those shown in Fig. 5. However, rapid decline of pasture cover that can even occur under conditions of high initial ground cover (e.g. winter 2018 in Bouganville) demonstrates the susceptibility of pasture cover to overgrazing and again underscores the need for regular monitoring of paddock conditions to ensure long-term sustainable grazing outcomes. Our study also showed high intra-paddock spatial variation of pasture cover (Fig. 6). This may be partly attributed to grazing regimes, grazing preferences including sheep camping locations, variation in pasture phenology, soil type and land elevation. Indeed, the majority of native pasture species in this study were concentrated in areas close to trees, on steep slopes or on low-lying coastal areas. In general, native pasture classes were associated with lower pasture cover while improved pasture classes were associated with higher NDVI (compare Fig. 6 with Figs 7 and 8). Such differences indicate that native pasture species generally had less ground cover at the time this study was conducted. Pastures with greater dry matter digestibility (DMD) per unit area will lead to greater liveweight gain and potentially profitability: higher DMD can be attained through either greater biomass, higher-quality pasture or both, so ideally farmers would cultivate pastures leading to the highest DMD per hectare (Chapman et al. 2012; Harrison et al. 2014a). However, other factors such as nutrient-use efficiency and drought resilience are also important. The native grasses in this study (primarily Kangaroo grass and Wallaby grass) require no synthetic fertilizer inputs, while the improved pastures (cocksfoot, phalaris and perennial ryegrass) require nitrogen inputs for higher productivity (Christie et al. 2020; Rawnsley et al. 2020). In this case, the farm had a focus on pasture management strategies to conserve and enhance the growth and regeneration of native species due to lower reliance on synthetic fertilizer inputs (which in Australia are costly) (Hayes et al. 2019). We suggest that future work be aimed at the derivation of metrics for coupling NDVI with botanical composition to enable better understanding of relationships between spatial pasture growth and pasture species abundance. Having such knowledge would facilitate the development of pasture management strategies (e.g. fertilizer application or strategic grazing management) that may lead to desirable shifts in botanical composition. As the majority of previous studies have focussed on assessment of pasture cover at the landscape level (Roberts et al. 2002; Shuang et al. 2012; Ali et al. 2016), the implications of grazing management decisions at the paddock level have received relatively little attention in remote sensing research. Similarly, while there are many decision-support systems to help graziers manage pastures (AgroInsider 2020; Cibolabs 2020; DataFarming 2020; Decipher 2020; Pasture.io 2020), such tools often do not allow distinction between pasture species. Optical remote sensing may provide an opportunity for pasture classification analysis (Congalton 1991; Vickery et al. 1997; Khatami et al. 2016). Employing MODIS and enhanced thematic mapper plus (ETM+), Liu et al. (2015) used time-series NDVI to classify C3 and C4 features with 86 % accuracy. Other work has used discriminant analysis to compare imagery from Landsat 8 OLI, Sentinel-2 and WorldView-2 to discriminate between C3 and C4 grass species and has found varying levels of accuracies across sensors and seasons, ranging from 86 to 100 % (Cletah and Onisimo 2017a, b). Similarly, our analysis showed that MLC with Sentinel-2 allows reasonable prediction of pasture classes at the paddock scale. Here, our analysis was performed with partial ground cover such that leaf area index was never greater than 3.0. Such leaf areas are common to many extensive grazing zones in Australia. Areas with leaf area indices greater than 3.0 (or when pasture dry matter is greater than 3000 kg DM ha−1), synthetic aperture radar imagery (SAR) may offer a way forward. Indeed, the number of studies conducted using SAR to analyse physical characteristics of pasture such as canopy height continues to grow (Hill et al. 2005; Crabbe et al. 2019, 2020; Dos Reis et al. 2020). In recent studies, such approaches have been conducted with reasonable accuracy (e.g. 68–84 % (Crabbe et al. 2019)) and have been further improved by combining SAR with optical data (87–93 % accuracy (Crabbe et al. 2020)). Using imagery from Sentinel-1, Crabbe et al. (2020) also showed that SVM (a form of machine learning) outperformed the MLC approach used here, as well as a random forest approach. Such results suggest that future studies of this type could trial the SVM as a classification method for distinguishing botanical composition. 4.2 Prospects for pasture botanical composition diagnosis using remote sensing We showed that the spectral signature of native grass was distinct compared with that of the improved and mixed species. A similar result was confirmed by Cletah and Onisimo (2017b), who found that the spectral signature (wavelength) of native Themeda grass was higher than that of Festuca, especially in summer. This indicates that native species morphology and physiology is more identifiable than (the difference between) improved and mixed species classes. This observation may be due to differences in leaf nitrogen or pigmentation between species and thus spectral reflectance (Walburg et al. 1982; Adjorlolo et al. 2014; Adjorlolo et al. 2015). Successful discrimination of botanical composition has positive ramifications for biodiversity assessment, habitat management and ecosystem services (Weibull et al. 2003; Dorrough et al. 2004; Sanderson et al. 2013). Austrheim and Eriksson (2001) showed that remote sensing information on sheep/cattle densities, wildlife densities and soil characteristics can influence species diversity and have implications for sustainable grazing management. Distinguishing botanical composition may also help manage and maintain natural capital (the economic value provided by wildlife, provision of habitat), because areas with native pasture species may have greater soil carbon stocks compared with areas with improved species (Chan et al. 2010). Ultimately the approach outlined here is conducive to greater insight into biodiversity, natural capital and habitat management, which together potentially enhance the prospects for regenerative agriculture. As more high-resolution optical satellite imagery becomes available, our approach could be further refined and implemented using imagery with greater spatial and temporal resolution (e.g. using the Planet CubeSat satellite constellation (Planet 2020)). Such information would be expected to increase analytical quality in paddock-level monitoring (Dos Reis et al. 2020). As well, the combination of SAR and optical data with different machine learning techniques (e.g. SVM) may provide further insights into pasture attributes. Combining satellite-based approaches with hyperspectral imagery collected using Unmanned Ariel Vehicle Systems (UAVS) (Psomas et al. 2011; Melville et al. 2019) may also add value, though labour complexity in such cases may also increase (indeed, a key advantage of system is the low requirement for physical input from the grazing manager). 5. CONCLUSIONS We demonstrated that the combination of Sentinel-2 imagery and supervised learning is a viable approach to quantify seasonal pasture ground cover and botanical composition. Greater performance can be anticipated when more field measurements become available for training and enable the use of more complex, data-intensive supervised classifiers. Although pasture cover varied seasonally in line with changes in climate and grazing management, paddock-level changes in pasture species distribution were able to be reliably monitored. Overall, our results could be used to enact more timely grazing management as well as tactical farming approaches, such as conservation of native pasture species when ground cover is low. Such approaches would be expected to improve ground cover, reduce soil erosion, improve soil carbon stocks and biodiversity values. Improvement of natural capital in this way would be expected to improve long-term sustainability and profitability. SUPPORTING INFORMATION The following additional information is available in the online version of this article— Table S1. Dates of cloud-free images that were considered to monitor pasture growth changes for Balsey Hill, Cottage Hill, Bouganville and Lords images over different seasons during the period of December 2017 to August 2020, presented in Fig. 5. Table S2. Radiometric and spatial resolution of Sentinel-2 bands used in this study. Table S3. Confusion matrix for autumn 2020 classification. Table S4. Confusion matrix for winter 2020 classification. Figure S1. Spatial characterization of the predictive performances of supervised classification using Sentinel-2 features for winter 2019, image date 23 August 2019. Figure S2. Spatial characterization of the predictive performances of supervised classification using Sentinel-2 features for spring 2019, image date 22 October 2019. Figure S3. Spatial characterization of the predictive performances of supervised classification using Sentinel-2 features for summer 2020, image date 30 January 2019. ACKNOWLEDGEMENTS The authors are grateful to the National Aeronautics and Space Administration (NASA) and the European Space Agency (ESA) for providing Sentinel-2 MSI images. The authors thank the farm owner, manager and staff on the study farm for their support and assistance. Thanks to Hai Dang Le and Tom Whitehead for assistance in collecting field data, and to Rowan Smith for providing advice on pasture monitoring. SOURCE OF FUNDING This research was supported by the Department of Agriculture and Water Resources, Government of Australia, through funding from the Australian National Landcare Program (NLP)-Round 3. The research was also supported by the Tasmanian Institute of Agriculture. CONFLICT OF INTEREST None declared. CONTRIBUTIONS BY THE AUTHORS Conceptualization: I.A. and M.T.H.; method: I.A., M.T.H. and J.W.; data/image collection: I.A., J.W., K.B., L.G., J.M.D.S. and F.M.; data/image analysis: I.A.; original draft preparation: I.A.; review and editing: M.T.H., J.W., F.W., K.B. and L.G.; funding acquisition: J.W. and M.T.H. LITERATURE CITED Adjorlolo C , Mutanga O, Cho MA. 2014 . Estimation of canopy nitrogen concentration across C3 and C4 grasslands using WorldView-2 multispectral data . IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 7 : 4385 – 4392 . Google Scholar Crossref Search ADS WorldCat Adjorlolo C , Mutanga O, Cho MA. 2015 . Predicting C3 and C4 grass nutrient variability using in situ canopy reflectance and partial least squares regression . International Journal of Remote Sensing 36 : 1743 – 1761 . Google Scholar Crossref Search ADS WorldCat AgroInsider . 2020 . Agroinsider Web App . AgroInsider . https://agroinsider.com/ ( 24 August 2020 ). Aguiar A , Adami M, Silva WF, Rudorff BFT, Mello MP, Silva JdSVd. 2010 . Modis time series to assess pasture land . 2010 IEEE International Geoscience and Remote Sensing Symposium , 2123 – 2126 . Google Scholar OpenURL Placeholder Text WorldCat Alcock DJ , Harrison MT, Rawnsley RP, Eckard RJ. 2015 . Can animal genetics and flock management be used to reduce greenhouse gas emissions but also maintain productivity of wool-producing enterprises? Agricultural Systems 132 : 25 – 34 . Google Scholar Crossref Search ADS WorldCat Ali I , Cawkwell F, Dwyer E, Barrett B, Green S. 2016 . Satellite remote sensing of grasslands: from observation to management . Journal of Plant Ecology 9 : 649 – 671 . Google Scholar Crossref Search ADS WorldCat Anderson SA , Anderson WR, Hollis JJ, Botha EJ. 2011 . A simple method for field-based grassland curing assessment . International Journal of Wildland Fire 20 : 804 – 814 . Google Scholar Crossref Search ADS WorldCat Asher M , Raedts P, Harrison M, Hills J, Rawnsley R. 2018 . Estimating pasture biomass with Planet Labs CubeSats . Proceedings of the 2018 Australasian Dairy Science Symposium , 196 – 198 . Google Scholar OpenURL Placeholder Text WorldCat Austrheim G , Eriksson O. 2001 . Plant species diversity and grazing in the Scandinavian mountains-patterns and processes at different spatial scales . Ecography 24 : 683 – 695 . Google Scholar Crossref Search ADS WorldCat Baldi G , Guerschman JP, Paruelo JM. 2006 . Characterizing fragmentation in temperate South America grasslands . Agriculture, Ecosystems & Environment 116 : 197 – 208 . Google Scholar Crossref Search ADS WorldCat Barrachina M , Cristóbal J, Tulla AF. 2015 . Estimating above-ground biomass on mountain meadows and pastures through remote sensing . International Journal of Applied Earth Observation and Geoinformation 38 : 184 – 192 . Google Scholar Crossref Search ADS WorldCat Belgiu M , Drăguţ L. 2016 . Random forest in remote sensing: a review of applications and future directions . ISPRS Journal of Photogrammetry and Remote Sensing 114 : 24 – 31 . Google Scholar Crossref Search ADS WorldCat Bell LW , Harrison MT, Kirkegaard JA. 2015 . Dual-purpose cropping–capitalising on potential grain crop grazing to enhance mixed-farming profitability . Crop and Pasture Science 66 : i – iv . Google Scholar Crossref Search ADS WorldCat Bella D , Faivre R, Ruget F, Seguin B, Guérif M, Combal B, Weiss M, Rebella C. 2004 . Remote sensing capabilities to estimate pasture production in France . International Journal of Remote Sensing 25 : 5359 – 5372 . Google Scholar Crossref Search ADS WorldCat BoM . 2020 . Bureau of meteorology . Government of Asutralia . http://www.bom.gov.au/climate/averages/tables/ca_tas_names.shtml ( 25 August 2020 ). Boschetti M , Bocchi S, Brivio PA. 2007 . Assessment of pasture production in the Italian Alps using spectrometric and remote sensing information . Agriculture, Ecosystems & Environment 118 : 267 – 272 . Google Scholar Crossref Search ADS WorldCat Brenner JC , Christman Z, Rogan J. 2012 . Segmentation of Landsat Thematic Mapper imagery improves buffelgrass (Pennisetum ciliare) pasture mapping in the Sonoran Desert of Mexico . Applied Geography 34 : 569 – 575 . Google Scholar Crossref Search ADS WorldCat Budd B , Thorpe J. 2009 . Benefits of managed grazing: a manager’s perspective . Rangelands 31 : 11 – 14 . Google Scholar Crossref Search ADS WorldCat Chan KY , Oates A, Li G, Conyers M, Prangnell R, Poile G, Liu D, Barchia I. 2010 . Soil carbon stocks under different pastures and pasture management in the higher rainfall areas of south-eastern Australia . Soil Research 48 : 7 – 15 . Google Scholar Crossref Search ADS WorldCat Chang-Fung-Martel J , Harrison M, Rawnsley R, Smith A, Meinke H. 2018 . The impact of extreme climatic events on pasture-based dairy systems: a review . Crop and Pasture Science 68 : 1158 – 1169 . Google Scholar Crossref Search ADS WorldCat Chapman D , Tharmaraj J, Agnusdei M, Hill J. 2012 . Regrowth dynamics and grazing decision rules: further analysis for dairy production systems based on perennial ryegrass (Lolium perenne L.) pastures . Grass and Forage Science 67 : 77 – 95 . Google Scholar Crossref Search ADS WorldCat Christie K , Smith A, Rawnsley R, Harrison M, Eckard R. 2020 . Simulated seasonal responses of grazed dairy pastures to nitrogen fertilizer in SE Australia: N loss and recovery . Agricultural Systems 182 : 102847 . Google Scholar Crossref Search ADS WorldCat Cibolabs . 2020 . Cibolabs - building solutions for agriculture . Cibolabs . https://www.cibolabs.com.au/ ( 24 August 2020 ). Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Cletah S , Onisimo M. 2017a . Examining the strength of the newly-launched Sentinel 2 MSI sensor in detecting and discriminating subtle differences between C3 and C4 grass species . ISPRS Journal of Photogrammetry and Remote Sensing 129 : 32 – 40 . Google Scholar Crossref Search ADS WorldCat Cletah S , Onisimo M. 2017b . Seasonal discrimination of C3 and C4 grasses functional types: an evaluation of the prospects of varying spectral configurations of new generation sensors . International Journal of Applied Earth Observation and Geoinformation 62 : 47 – 55 . Google Scholar Crossref Search ADS WorldCat Conant RT , Paustian K. 2002 . Potential soil carbon sequestration in overgrazed grassland ecosystems . Global Biogeochemical Cycles 16 : 90-91 – 90-99 . Google Scholar Crossref Search ADS WorldCat Congalton RG . 1991 . A review of assessing the accuracy of classifications of remotely sensed data . Remote Sensing of Environment 37 : 35 – 46 . Google Scholar Crossref Search ADS WorldCat Crabbe RA , Lamb DW, Edwards C. 2019 . Discriminating between C3, C4, and mixed C3/C4 pasture grasses of a grazed landscape using multi-temporal Sentinel-1a data . Remote Sensing 11 : 253 . Google Scholar Crossref Search ADS WorldCat Crabbe RA , Lamb D, Edwards C. 2020 . Discrimination of species composition types of a grazed pasture landscape using Sentinel-1 and Sentinel-2 data . International Journal of Applied Earth Observation and Geoinformation 84 : 101978 . Google Scholar Crossref Search ADS WorldCat Csendes B , Mucsi L. 2016 . Identification and spectral evaluation of agricultural crops on hyperspectral airborne data . Journal of Environmental Geography 9 : 49 – 53 . Google Scholar Crossref Search ADS WorldCat DataFarming . 2020 . Data Farming-high resolution images available now . DataFarming . https://www.datafarming.com.au/ ( 24 August 2020 ). Davidson EA . 2008 . Objective indicators of pasture degradation from spectral mixture analysis of Landsat imagery spectral scalars of pasture degradation . Journal of Geophysical Research: Biogeosciences 113 . Google Scholar OpenURL Placeholder Text WorldCat Decipher. 2020 . Decipher-monitor and manage your land like never before . https://www.decipher.com.au/ ( 24 August 2020 ). Dorrough J , Yen A, Turner V, Clark S, Crosthwaite J, Hirth J. 2004 . Livestock grazing management and biodiversity conservation in Australian temperate grassy landscapes . Australian Journal of Agricultural Research 55 : 279 – 295 . Google Scholar Crossref Search ADS WorldCat Dos Reis AA , Werner JPS, Silva BC, Figueiredo GKDA, Antunes JFG, Esquerdo JCDM, Coutinho AC, Lamparelli RAC, Rocha JV, Magalhães PSG. 2020 . Monitoring pasture aboveground biomass and canopy height in an integrated crop–livestock system using textural information from PlanetScope imagery . Remote Sensing 12 : 2534 . Google Scholar Crossref Search ADS WorldCat DPIPWE . 2000 . Soil map series of Tasmania . Tasmania Department of Primary Industries, Water & Environment, Tasmania . https://dpipwe.tas.gov.au/Documents/bucksoil.pdf (24 August 2020). ESA . 2020 . ESA Sentinel online . European Space Agency (ESA) . https://sentinel.esa.int/web/sentinel/missions ( 29 January 2020 ). Fedrigo JK , Ataide PF, Filho JA, Oliveira LV, Jaurena M, Laca EA, Overbeck GE, Nabinger C. 2018 . Temporary grazing exclusion promotes rapid recovery of species richness and productivity in a long-term overgrazed Campos grassland . Restoration Ecology 26 : 677 – 685 . Google Scholar Crossref Search ADS WorldCat Flynn ES , Dougherty CT, Wendroth O. 2008 . Assessment of pasture biomass with the normalized difference vegetation index from active ground-based sensors . Agronomy Journal 100 : 114 – 121 . Google Scholar Crossref Search ADS WorldCat Gómez C , White JC, Wulder MA. 2016 . Optical remotely sensed time series data for land cover classification: a review . ISPRS Journal of Photogrammetry and Remote Sensing 116 : 55 – 72 . Google Scholar Crossref Search ADS WorldCat Hanna MM , Steyn-Ross DA, Steyn-Ross M. 1999 . Estimating biomass for New Zealand pasture using optical remote sensing techniques . Geocarto International 1 : 89 – 94 . Google Scholar Crossref Search ADS WorldCat Harrison MT , Christie KM, Rawnsley RP, Eckard RJ. 2014a . Modelling pasture management and livestock genotype interventions to improve whole-farm productivity and reduce greenhouse gas emissions intensities . Animal Production Science 54 : 2018 – 2028 . Google Scholar Crossref Search ADS WorldCat Harrison MT , Cullen BR, Armstrong D. 2017 . Management options for dairy farms under climate change: effects of intensification, adaptation and simplification on pastures, milk production and profitability . Agricultural Systems 155 : 19 – 32 . Google Scholar Crossref Search ADS WorldCat Harrison MT , Cullen BR, Rawnsley RP. 2016a . Modelling the sensitivity of agricultural systems to climate change and extreme climatic events . Agricultural Systems 148 : 135 – 148 . Google Scholar Crossref Search ADS WorldCat Harrison MT , Cullen BR, Tomkins NW, McSweeney C, Cohn P, Eckard RJ. 2016b . The concordance between greenhouse gas emissions, livestock production and profitability of extensive beef farming systems . Animal Production Science 56 : 370 – 384 . Google Scholar Crossref Search ADS WorldCat Harrison MT , Evans JR, Dove H, Moore AD. 2012a . Dual-purpose cereals: can the relative influences of management and environment on crop recovery and grain yield be dissected? Crop and Pasture Science 62 : 930 – 946 . Google Scholar Crossref Search ADS WorldCat Harrison MT , Evans JR, Dove H, Moore AD. 2012b . Recovery dynamics of rainfed winter wheat after livestock grazing 1. Growth rates, grain yields, soil water use and water-use efficiency . Crop and Pasture Science 62 : 947 – 959 . Google Scholar Crossref Search ADS WorldCat Harrison MT , Evans JR, Moore AD. 2012c . Using a mathematical framework to examine physiological changes in winter wheat after livestock grazing: 1. Model derivation and coefficient calibration . Field Crops Research 136 : 116 – 126 . Google Scholar Crossref Search ADS WorldCat Harrison MT , Jackson T, Cullen BR, Rawnsley RP, Ho C, Cummins L, Eckard RJ. 2014b . Increasing ewe genetic fecundity improves whole-farm production and reduces greenhouse gas emissions intensities: 1. Sheep production and emissions intensities . Agricultural Systems 131 : 23 – 33 . Google Scholar Crossref Search ADS WorldCat Hayes RC , Ara I, Badgery WB, Culvenor RA, Haling RE, Harris CA, Li GD, Norton MR, Orgill SE, Penrose B, Smith RW. 2019 . Prospects for improving perennial legume persistence in mixed grazed pastures of south-eastern Australia, with particular reference to white clover . Crop and Pasture Science . doi:10.1071/CP19063. Google Scholar OpenURL Placeholder Text WorldCat Hill MJ , Ticehurst CJ, Lee J-S, Grunes MR, Donald GE, Henry D. 2005 . Integration of optical and radar classifications for mapping pasture type in Western Australia . IEEE Transactions on Geoscience and Remote Sensing 43 : 1665 – 1681 . Google Scholar Crossref Search ADS WorldCat Ho C , Jackson T, Harrison M, Eckard R. 2014 . Increasing ewe genetic fecundity improves whole-farm production and reduces greenhouse gas emissions intensities: 2. Economic performance . Animal Production Science 54 : 1248 – 1253 . Google Scholar Crossref Search ADS WorldCat Isbell R . 2016 . The Australian soil classification . Australia : CSIRO . Google Scholar Crossref Search ADS Google Preview WorldCat COPAC Jin Y , Yang X, Qiu J, Li J, Gao T, Wu Q, Zhao F, Ma H, Yu H, Xu B. 2014 . Remote sensing-based biomass estimation and its spatio-temporal variations in temperate Grassland, Northern China . Remote Sensing 6 : 1496 – 1513 . Google Scholar Crossref Search ADS WorldCat Khatami R , Mountrakis G, Stehman SV. 2016 . A meta-analysis of remote sensing research on supervised pixel-based land-cover image classification processes: general guidelines for practitioners and future research . Remote Sensing of Environment 177 : 89 – 100 . Google Scholar Crossref Search ADS WorldCat Kogan F , Stark R, Gitelson A, Jargalsaikhan L, Dugrajav C, Tsooj S. 2004 . Derivation of pasture biomass in Mongolia from AVHRR-based vegetation health indices . International Journal of Remote Sensing 25 : 2889 – 2896 . Google Scholar Crossref Search ADS WorldCat Lillesand T , Kiefer RW, Chipman J. 2015 . Remote sensing and image interpretation . Newyork : John Wiley & Sons . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Liu X , Bo Y, Zhang J, He Y. 2015 . Classification of C3 and C4 vegetation types using MODIS and ETM+ blended high spatio-temporal resolution data . Remote Sensing 7 : 15244 – 15268 . Google Scholar Crossref Search ADS WorldCat Melville B , Lucieer A, Aryal J. 2019 . Classification of lowland native grassland communities using hyperspectral Unmanned Aircraft System (UAS) Imagery in the Tasmanian midlands . Drones 3 : 5 . Google Scholar Crossref Search ADS WorldCat Miehe G , Miehe S, Bach K, Nölling J, Hanspach J, Reudenbach C, Kaiser K, Wesche K, Mosbrugger V, Yang Y. 2011 . Plant communities of central Tibetan pastures in the Alpine Steppe/Kobresia pygmaea ecotone . Journal of Arid Environments 75 : 711 – 723 . Google Scholar Crossref Search ADS WorldCat Mohri M , Rostamizadeh A, Talwalkar A. 2018 . Foundations of machine learning . England : MIT Press . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Mountrakis G , Im J, Ogole C. 2011 . Support vector machines in remote sensing: a review . ISPRS Journal of Photogrammetry and Remote Sensing 66 : 247 – 259 . Google Scholar Crossref Search ADS WorldCat Müller H , Rufin P, Griffiths P, Barros Siqueira AJ, Hostert P. 2015 . Mining dense Landsat time series for separating cropland and pasture in a heterogeneous Brazilian savanna landscape . Remote Sensing of Environment 156 : 490 – 499 . Google Scholar Crossref Search ADS WorldCat Pal M . 2005 . Random forest classifier for remote sensing classification . International Journal of Remote Sensing 26 : 217 – 222 . Google Scholar Crossref Search ADS WorldCat Pal M , Mather PM. 2003 . An assessment of the effectiveness of decision tree methods for land cover classification . Remote Sensing of Environment 86 : 554 – 565 . Google Scholar Crossref Search ADS WorldCat Pasture.io . 2020 . Pasture.io-pasture measurement on autopilot . Pasture.io . https://pasture.io/ ( 24 August 2020 ). Pembleton K , Cullen B, Rawnsley R, Harrison M, Ramilan T. 2016 . Modelling the resilience of forage crop production to future climate change in the dairy regions of Southeastern Australia using APSIM . Journal of Agricultural Science 154 : 1131 – 1152 . Google Scholar Crossref Search ADS WorldCat Planet . 2020 . Planet - using space to help life ion earth. Planet Labs Inc . https://www.planet.com/company/ ( 7 September 2020 ). Psomas A , Kneubühler M, Huber S, Itten K, Zimmermann N. 2011 . Hyperspectral remote sensing for estimating aboveground biomass and for exploring species richness patterns of grassland habitats . International Journal of Remote Sensing 32 : 9007 – 9031 . Google Scholar Crossref Search ADS WorldCat Pullanagari RR , Yule I, King W, Dalley D, Dynes R. 2011 . The use of optical sensors to estimate pasture quality . International Journal on Smart Sensing and Intelligent Systems 4 : 125 – 137 . Google Scholar Crossref Search ADS WorldCat Pullanagari RR , Yule IJ, Hedley MJ, Tuohy MP, Dynes RA, King WM. 2012 . Multi-spectral radiometry to estimate pasture quality components . Precision Agriculture 13 : 442 – 456 . Google Scholar Crossref Search ADS WorldCat Quintano C , Fernández-Manso A, Shimabukuro YE, Pereira G. 2012 . Spectral unmixing . International Journal of Remote Sensing 33 : 5307 – 5340 . Google Scholar Crossref Search ADS WorldCat Rawnsley R , Smith A, Christie K, Harrison M, Eckard R. 2020 . Current and future direction of nitrogen fertiliser use in Australian grazing systems . Crop and Pasture Science 70 : 1034 – 1043 . Google Scholar Crossref Search ADS WorldCat Reiche M , Funk R, Zhang Z, Hoffmann C, Reiche J, Wehrhan M, Li Y, Sommer M. 2012 . Application of satellite remote sensing for mapping wind erosion risk and dust emission-deposition in Inner Mongolia grassland, China . Grassland Science 58 : 8 – 19 . Google Scholar Crossref Search ADS WorldCat Richards JA , Richards J. 1999 . Remote sensing digital image analysis , Vol. 3 . Springer , Berlin . Google Scholar Crossref Search ADS Google Preview WorldCat COPAC Roberts D , Numata I, Holmes K, Batista G, Krug T, Monteiro A, Powell B, Chadwick O. 2002 . Large area mapping of land-cover change in Rondônia using multitemporal spectral mixture analysis and decision tree classifiers . Journal of Geophysical Research: Atmospheres 107 : LBA 40-41 – LBA 40-18 . Google Scholar Crossref Search ADS WorldCat Roth KL , Dennison PE, Roberts DA. 2012 . Comparing endmember selection techniques for accurate mapping of plant species and land cover using imaging spectrometer data . Remote Sensing of Environment 127 : 139 – 152 . Google Scholar Crossref Search ADS WorldCat Rowntree K , Duma M, Kakembo V, Thornes J. 2004 . Debunking the myth of overgrazing and soil erosion . Land Degradation & Development 15 : 203 – 214 . Google Scholar Crossref Search ADS WorldCat Rufin P , Müller H, Pflugmacher D, Hostert P. 2015 . Land use intensity trajectories on Amazonian pastures derived from Landsat time series . International Journal of Applied Earth Observation and Geoinformation 41 : 1 – 10 . Google Scholar Crossref Search ADS WorldCat Russell S , Norvig P. 2002 . Artificial intelligence: a modern approach . Pearson Education, England . Google Scholar Google Preview OpenURL Placeholder Text WorldCat COPAC Sanderson MA , Archer D, Hendrickson J, Kronberg S, Liebig M, Nichols K, Schmer M, Tanaka D, Aguilar J. 2013 . Diversification and ecosystem services for conservation agriculture: outcomes from pastures and integrated crop–livestock systems . Renewable Agriculture and Food Systems 28 : 129 – 144 . Google Scholar Crossref Search ADS WorldCat Sanderson M , Soder K, Muller L, Klement K, Skinner R, Goslee S. 2005 . Forage mixture productivity and botanical composition in pastures grazed by dairy cattle . Agronomy Journal 97 : 1465 – 1471 . Google Scholar Crossref Search ADS WorldCat Schaefer MT , Lamb DW. 2016 . A combination of plant NDVI and LiDAR measurements improve the estimation of pasture biomass in Tall fescue (Festuca arundinacea var. Fletcher) . Remote Sensing 8 : 109 . Google Scholar Crossref Search ADS WorldCat Serrano J , Shahidian S, Marques da Silva J. 2018 . Monitoring seasonal pasture quality degradation in the Mediterranean montado ecosystem: proximal versus remote sensing . Water 10 : 1422 . Google Scholar Crossref Search ADS WorldCat Shuang L , Potter C, Hiatt C. 2012 . Monitoring of net primary production in California rangelands using Landsat and MODIS satellite remote sensing . Natural Resources 3 : 56 – 65 . Google Scholar Crossref Search ADS WorldCat Sonneveld BGJS , Pande S, Georgis K, Keyzer MA, Ali AS, Takele A. 2010 . Land degradation and overgrazing in the Afar region, Ethiopia: a spatial analysis . Zdruli P, Pagliai M, Kapur S, Faz Cano A, eds. Land degradation and desertification: assessment, mitigation and remediation . Dordrecht : Springer , 97 – 109 . Google Scholar Crossref Search ADS Google Preview WorldCat COPAC Toivonen T , Luoto M. 2003 . Landsat TM images in mapping of semi-natural grasslands and analysing of habitat pattern in an agricultural landscape in south-west Finland . Fennia-International Journal of Geography 181 : 49 – 67 . Google Scholar OpenURL Placeholder Text WorldCat Tothill JC , Hargreaves JNG, Jones RM, McDonald CK. 1992 . BOTANAL–a comprehensive sampling and computing procedure for estimating pasture yield and composition. 1. Field sampling . Tropical Agronomy Technical Memorandum 78 ( 24 ). Google Scholar OpenURL Placeholder Text WorldCat Tucker CJ . 1979 . Red and photographic infrared linear combinations for monitoring vegetation . Remote sensing of Environment 8 : 127 – 150 . Google Scholar Crossref Search ADS WorldCat Veganzones MA , Grana M. 2008 . Endmember extraction methods: a short review . International Conference on Knowledge-Based and Intelligent Information and Engineering Systems . Google Scholar OpenURL Placeholder Text WorldCat Vickery PJ , Hill MJ, Donald GE. 1997 . Satellite derived maps of pasture growth status: association of classification with botanical composition . Australian Journal of Experimental Agriculture 37 : 547 – 562 . Google Scholar Crossref Search ADS WorldCat Waite R . 1994 . The application of visual estimation procedures for monitoring pasture yield and composition in exclosures and small plots . Tropical Grasslands 28 : 38 . Google Scholar OpenURL Placeholder Text WorldCat Walburg G , Bauer ME, Daughtry C, Housley T. 1982 . Effects of nitrogen nutrition on the growth, yield, and reflectance characteristics of corn canopies 1 . Agronomy Journal 74 : 677 – 683 . Google Scholar Crossref Search ADS WorldCat Weibull A-C , Östman Ö, Granqvist Å. 2003 . Species richness in agroecosystems: the effect of landscape, habitat and farm management . Biodiversity & Conservation 12 : 1335 – 1355 . Google Scholar Crossref Search ADS WorldCat Weiers S , Bock M, Wissen M, Rossner G. 2004 . Mapping and indicator approaches for the assessment of habitats at different scales using remote sensing and GIS methods . Landscape and Urban Planning 67 : 43 – 65 . Google Scholar Crossref Search ADS WorldCat Wylie BK , Harrington JA, Prince SD, Denda I. 1991 . Satellite and ground-based pasture production assessment in Niger: 1986–1988 . International Journal of Remote Sensing 12 : 1281 – 1300 . Google Scholar Crossref Search ADS WorldCat Xu Y , Shi J, Du J. 2015 . An improved endmember selection method based on vector length for MODIS reflectance channels . Remote Sensing 7 : 6280 – 6295 . Google Scholar Crossref Search ADS WorldCat Xu Y , Tao W, Qin Z, Liu H, Miao J, Bi Y. 2008 . MODIS-based remote sensing monitoring of grass production in China . International Journal of Remote Sensing 29 : 5313 – 5327 . Google Scholar Crossref Search ADS WorldCat Zengeya FM , Mutanga O, Murwira A. 2013 . Linking remotely sensed forage quality estimates from WorldView-2 multispectral data with cattle distribution in a savanna landscape . International Journal of Applied Earth Observation and Geoinformation 21 : 513 – 524 . Google Scholar Crossref Search ADS WorldCat © The Author(s) 2020. Published by Oxford University Press on behalf of the Annals of Botany Company. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium, provided the original work is properly cited.
TI - Modelling seasonal pasture growth and botanical composition at the paddock scale with satellite imagery
JF - in silico Plants
DO - 10.1093/insilicoplants/diaa013
DA - 2021-01-01
UR - https://www.deepdyve.com/lp/oxford-university-press/modelling-seasonal-pasture-growth-and-botanical-composition-at-the-Xwan72fSn0
VL - 3
IS - 1
DP - DeepDyve
ER -