Access the full text.
Sign up today, get DeepDyve free for 14 days.
Peatlands as natural carbon sinks have a major impact on the climate balance and should therefore be monitored and protected. The hydrology of the peatland serves as an indicator of the carbon storage capacity. Hence, we investigate the question how suitable different remote sensing data are for monitoring the size of open water surface and the water table depth (WTD) of a peatland ecosystem. Furthermore, we examine the potential of combining remote sensing data for this purpose. We use C-band synthetic aperture radar (SAR) data from Sentinel-1 and multi-spectral data from Sentinel-2. The radar backscat- ter , the normalized difference water index (NDWI) and the modified normalized difference water index (MNDWI) are calculated and used for consideration of the WTD and the lake size. For the measurement of the lake size, we implement and investigate the methods: random forest, adaptive thresholding and an analysis according to the Dempster–Shafer theory. Correlations between WTD and the remote sensing data as well as NDWI are investigated. When looking at the individual data sets the results of our case study show that the VH polarized data produces the clearest delineation of the peatland 0 0 lake. However the adaptive thresholding of the weighted fusion image of -VH, -VV and MNDWI, and the random for- est algorithm with all three data sets as input proves to be the most suitable for determining the lake area. The correlation coefficients between /NDWI and WTD vary greatly and lie in ranges of low to moderate correlation. Keywords Fusion · SAR · NDWI · Peatland · Water area · Water table depth Zusammenfassung Fusion von SAR- und multispektralen Zeitreihen zur Bestimmung der Tiefe des Grundwasserspiegels und der Seefläche in Moorgebieten. Da Moore als natürliche Kohlenstoffsenken agieren, spielen sie eine entscheidende Rolle in der Klimabilanz und sollten daher überwacht und geschützt werden. Das Kohlenstoffspeicherpotential ist dabei abhängig vom Wasser - haushalt des Moors. Diese Arbeit untersucht die Eignung verschiedener Fernerkundungsdaten zur Beobachtung offener Wasserflächen und Grundwasserständen (WTD) sowie das Potential der Fusion dieser Daten. Genutzt werden Sentinel-1 synthetic apertur radar (SAR)-Daten und Sentinel-2 Multispektralbilder. Der Radar-Rückstreukoeffizient , der Normalized Difference Water Index (NDWI) und der modifizierte Normalized Difference Water Index (MNDWI) werden eingesetzt, um die Grundwasserstände und die Größe des Sees zu beobachten. Zur Bestimmung der Seefläche werden der Random Forest Algorithmus, ein adaptiver Schwellwertansatz und ein Ansatz der Dempster–Shafer Theorie angewandt. Die Korrelation zwischen den Grundwasserständen und den Fernerkundungsdaten und NDWI wird untersucht. Werden nur die einzel- nen Datensätze betrachtet, lassen die Ergebnisse dieser Studie erkennen, dass sich mittels der -VH Daten die deutlichste * Katrin Krzepek email@example.com Jakob Schmidt firstname.lastname@example.org Dorota Iwaszczuk email@example.com Remote Sensing and Image Analysis, Department of Civil and Environmental Engineering Sciences, Technical University of Darmstadt, Darmstadt, Germany Vol.:(0123456789) 1 3 PFG Abgrenzung des Moorsees ergibt. Der adaptive Schwellwertansatz angewandt auf das gewichtete Fusionsbild von -VH, -VV und MNDWI und der Random Forest Algorithmus mit den drei Datensätzen als Input erweisen sich jedoch als am besten geeignet für die Bestimmung der Seefläche. Die Korrelationskoeffizienten von /NDWI und den Grundwasserständen schwanken stark und liegen in Bereichen einer geringen bis mittleren Korrelation. and not always possible. In our work we therefore explore 1 Introduction to what extent remote sensing data can replace the meas- urement of in situ data. To this end, we are investigating Wetlands are ecosystems that are permanently or seasonally the relationship between in situ and remote sensing data. flooded by water. Peatlands also belong to the generic term Measuring the lake levels by altimeter is generally possi- wetlands. Ecosystems whose organic soils are more than ble, but not suitable for small water areas like the lake in 30 cm thick and contain more than 30% organic material are our study area. So, instead of measuring the lake level, we often referred to as peatlands (Joosten and Clarke 2002). A decided to measure the lake area. We assume that the lake uniform definition does not exist. The terms peatland and area increases as the lake level increases and that there is a wetland are often used synonymously in Germany, as mainly direct correlation between these two variables. they are included in the land use category “wetlands” (UBA In the course of flood monitoring, radar data are often 2021). The peatland ecosystem is a very important natural used for water area detection with moderate to good accu- carbon sink. Carbon is sequestered from the air through plant racy: Cazals et al. (2016) indicates that open water can be growth and permanently stored through the formation of detected successfully, flooded grassland with moderate accu- peat. The rate at which atmospheric carbon is absorbed and racy. Martinis et al. (2015) reach a high user accuracy, but released from the peat is directly dependent on the hydrology in mountainous region the accuracy decreases, because of a of the ecosystem (Succow and Joosten 2012). Only a water- lot of water look alikes caused by radar shadowing effects. saturated peatland can store carbon over long periods (Succow Pulvirenti et al. (2011) also see problems with thresholding and Joosten 2012). due to roughening effects caused by wind speed. Tsygans- But many peatlands are drained and return carbon to the kaya et al. (2018) use Sentinel-1 synthetic aperture radar atmosphere in the form of greenhouse gases. In Germany, (SAR) time series data to classify temporary open water 91.9% of organic soils are drained and emit about 48 mil- and temporary flooded vegetation. One advantage of using lion tonnes of CO equivalents per year, which corresponds radar data is the independence from cloud cover, which to about 6% of total emissions in Germany (UBA 2021). often occurs simultaneously with flood events (Boni et al. The global emission from drained organic soils were nearly 2016). Water surface detection also finds application in one billion tonnes CO equivalents annually (Tubiello et al. land use land cover (LULC) detection. Bagwan and Sopan 2016). Other important ecosystem services, such as flood Gavali (2021) and Haque and Basak (2017) achieved high control and providing habitat for rare animal and plant spe- accuracy in detecting water bodies by using Landsat data. cies, are also lost through drainage. Drainages lead to a low- However, a detailed consideration of the lakeshore region ering of the water level in the area. This is necessary for and the boundaries to other areas is not presented. Huang land-use such as agriculture or pasture. Drainage measures et al. (2018) present in their review the growing need to can also lead to the drying out of permanently open water map surface water at a global scale and discuss the prob- areas in a peatland. In the course of rewetting, which is cur- lem of thresholding as the most critical issue in using water rently taking place on more and more peatlands, the drainage indices. Another issue Huang et al. (2018) and Liao et al. measures are being reversed. For deeply drained grassland, (2014) discuss is that vegetation cover tends to dominate an annual saving of 17 tonnes of CO equivalent per ha can the spectral signal, so flooding water under vegetation is be achieved through rewetting (Wilson et al. 2016). If rewet- difficult to detect. Liu et al. (2019) analyse the shadow effect ting is successful, the groundwater level rises. Usually the in optical data in urban centres, which were mostly misclas- water table depth (WTD) is measured by wells, which have sified as water, when only optical images are used. They to be checked individually and regularly. investigate a method for LULC mapping with optical and Our work looks at two sub-aspects of hydrology: perma- SAR data and reach high user accuracy in detecting water nently open water areas and the groundwater level. To moni- surfaces. In the special consideration of wetlands the focus is tor the rewetting of peatlands, the WTD and the lake level on monitoring landscape changes (Muro et al. 2016), distin- are often measured ground-based, as in our case study. It is guishing wetland areas from other land covers (Dong et al. assumed that when peatlands dry out, WTD and lake level 2014; Kaplan and Avdan 2018) and mapping of vegetation decrease, or when peatlands get wetter, WTD and lake level types (Mleczko and Mróz 2018) and water bodies (Huang increase. However, the collection of in situ data is costly 1 3 PFG and Jin 2019; Mleczko and Mróz 2018; Moser et al. 2016). size measurement, because MNDWI can obtain great water White et al. (2015) review techniques to demonstrate SAR body separation precision and, according to Bhunia (2021), capabilities for wetland monitoring and recommend SAR has more accurate open water knowledge than other spec- as the primary source of imagery, supported by other data tral indices like automated Water Extraction Index (AWEI), such as lidar, thermal, and optical imagery. Guo et al. (2017) Water Ratio Index (WRI) and New Water Index (NWI). The concluded in their review of wetland remote sensing that radar data allows us to investigate the moisture in the upper multi-source integration should be the trend of future wet- soil layers. To analyse the relationship of the remote sens- land remote sensing. Kaplan et al. (2019) show the strong ing data to the WTD, we decided to use the multi-spectral relation between Landsat 8 NDWI and Sentinel-1 data in index NDWI in addition to C-band SAR data, because it wetlands. shows strong correlations with peat moisture at 3 cm depth Previous studies are successfully showing the connection (Meingast et al. 2014), which is approximately the penetra- between SAR backscatter and groundwater levels. The radar tion depth of the radar signal. signal penetrates the upper soil layer and the backscatter is The aim of this study is to develop methods that help to influenced by soil moisture (Baghdadi and Zribi 2016). Soil monitor the renaturation of peatlands and to demonstrate moisture is related to groundwater via capillary forces so the potentials and limitations of these methods using freely there is a connection between radar backscatter and WTD available Sentinel-1 and Sentinel-2 data. In this work we (Asmuß et al. 2019). Kim et al. (2017) analysed the hydrau- explore the following research questions: lic change in a peatland by comparing averaged SAR inten- sity from Radarsat-1 and ALOS PALSAR with groundwater – Is the MNDWI or the SAR backscatter better suited for level changes. They found that the SAR backscattering is measuring the lake size and which data have the highest significantly responsive. Bechtold et al. (2018) use ENVI- correlation with the lake level? SAT ASAR C-band backscatter data and obtain moderate – Is the fusion of the data via manually weighted features correlation coefficients. They conclude that backscatter is a or via the random forest algorithm more suitable for good indicator for WTD dynamics with a strong capillary determining the lake area? connection between WTD and the C-band-sensitive top 1–2 – Can we find correlations between SAR backscatter or cm of peat soils, but the interpretation seems to be more dif- NDWI and the WTD and what are the limiting factors ficult for natural than for drained peatlands. They see poten- here? tial in the use of finer resolution SAR data. Asmuß et al. – Is there a relation between the lake size and the lake level (2019) evaluated the correlation between high spatial reso- with regard to mapping them from Sentinel-data? lution Sentinel-1 and WTD in drained organic soils. All the mentioned publications discuss the influence of vegeta- tion. Bryant and Baird (2003) and Harris and Bryant (2009) 2 Study Area show in their analysis of sphagnum mosses the existence of strong relationships between near-surface hydrology and the The peatland Federseeried in Baden-Württemberg, Germany, spectral behaviour of vegetation. Meingast et al. (2014) find has an area of over 30 km² and has been under nature conser- strong correlation between water table positions and various vation since 1939 (LGRB 2021). The peatland was chosen spectral indices. as the study area because the WTD is monitored at about In general, remote sensing offers great potential for envi- 85 observation wells (Fig. 1). The WTD’s measurements in ronmental and climate protection because solutions can be Federseeried began in 2005 with individual monitoring sites, applied worldwide, long time periods can be studied, and which increased steadily to monitor the rewetted areas. The the methods are very cost-effective and less labour-intensive peat body, which has organic silt underneath, varies greatly compared to in situ measurements. Equally important is the in his thickness. The northern Federseeried is of particu- contactless measuring method of remote sensing, which is lar interest. These meadows were drained in the nineteenth ideal for observing protected ecosystems. century by an elaborate canal system for agricultural use. We use the concept of optical and SAR data fusion as it In summer 2014, construction projects were completed that is used for LULC detection with good results to overcome should lead to the rewetting of the peatland. The potential problems like radar and optical shadowing effects. When of the northern Federseeried for our project lies in the high monitoring peatlands, radar data have the additional advan- density of observation wells (approx. 43) and the fact that tage of penetrating the vegetation cover, but the penetration there is low vegetation. Which means that there is a prospect depends on the plant height and the density of the leaf area. of good measurement results from the SAR backscatter. The Thus, the lake size could be determined independently of management of the peatland is described as extensive. The the vegetation growing on the water surface. The modified vegetation in the northern Federseeried is mown twice a year normalized difference water index (MNDWI) is used for lake and the mown grass is used for agricultural purposes, e.g. 1 3 PFG Fig. 1 Study area peatland Federseeried with lake Federsee, located in Baden-Württemberg, Germany. Location of monitoring wells for WTD measurement (blue) and location of station for lake level measurement (red) pressed as pellets or as bedding for livestock. The aim is to We use band 3 (GREEN = 560 nm), band 8 (NIR = 842 nm) preserve an open landscape and wet meadow. The peatland and band 11 (MIR = 1610 nm). lake Federsee is on average only 1 m deep and surrounded From these bands we calculate the MNDWI according to by a reed belt up to 100 m wide. The reed belt also covers Xu (2006) as follows: large parts of the water surface. GREEN − MIR MNDWI = (1) GREEN + MIR 3 Methodology Furthermore we calculate the NDWI developed by McFeeters (1996). 3.1 Satellite Data GREEN − NIR NDWI = (2) GREEN + NIR Data from ESA’s satellites Sentinel-1 and Sentinel-2 are used. We use Sentinel-1 Ground Range Detected (GRD) Both indices range between −1 and +1 . Water surfaces usu- C-band data acquired using interferometric wide swath ally have a high MNDWI. A high NDWI indicates a high mode (IW). For the study area described above, a Senti- water content of the surface. nel-1 product is downloaded for each month of 2019 and both VH and VV polarisation are used. The backscatter data are preprocessed using SNAP, which involves thermal noise 3.2 In Situ Data: Lake Level and Water Table Depth removal, radiometric calibration, decibel (dB) conversion and terrain correction. These processing steps result in the The water level of the lake Federsee is measured daily (dB) backscatter data we use in the further course. Further- (Fig. 2). In 2019, the lowest lake level was at 53 cm, the more, we use one Sentinel-2 Level 2A product per month of highest at 85 cm and the average at 64 cm above the lake 2019. No cloud-free image are available for the months of bottom. The WTD of the peatland Federseeried is monitored January and November, which is why these months had to be at about 85 observation wells. Figure 3 shows the average omitted from the evaluation. In each case, the product with WTD at all monitoring wells over the year 2019. The aver- the lowest cloud cover and temporally closest to the date of age level was −24 cm with a maximum of −12 cm and a min- recording of the Sentinel-1 data set is selected in order to imum of −54 cm below the ground surface. A groundwater keep the comparability of the data sets as high as possible. level of more than −10 cm below the soil surface is the target 1 3 PFG Lake Level Mean individually and combine them with two methods whose decision basis we know and determine ourselves. To do this, 70 we first consider all three features and use them individu- ally for classification. The three classification results of the individual features are combined and a statement about the 30 probability of the pixels to belong to the water area is made using the mathematical Dempster–Shafer approach. Further- more, the features are extracted, weighted and the lake area is delineated by analysing the histogram of the fusion image. An overview of the proceeding is given in Fig. 4. Date 3.3.1 Supervised Classification with Random Forest Fig. 2 Lake level Federsee 2019 [cm], measured daily. Min: 53 cm, max: 85 cm, mean: 64 cm above the lake bottom Breiman (2001) describes random forest as ensemble of tree predictors that vote for the most popular class. Each tree Date depends on a random selection of features, which proves to be more robust with respect to noise (Breiman 2001). Variable importance can be indicated by internal estimates -10 (Breiman 2001). Compared to the fusion image, random for- est offers the advantage that weights for each data set are -20 learned individually. A disadvantage of the method is that -30 training data is required, which was not available in this -40 project and had to be created manually. As features we use 0 0 the three data sets -VH, -VV and MNDWI. -50 -60 3.3.2 Unsupervised Classification with Feature Analysis WTD Mean 0 0 Fig. 3 Average water table depth (WTD) in Federseeried 2019 [cm] Feature analysis The three data sets -VH, -VV and of all wells, measured weekly. Min: −54 cm, max: −12 cm, mean: MNDWI served also as inputs for an adaptive thresholding −26 cm below the ground surface method (Fig. 5). This method is applied to each data set individually. For , the minimum between the two peaks of in natural peatlands. The measurements are taken weekly by the histogram is selected as the threshold for the separation employee of the NABU Nature Conservation Centre Fed- between the two surfaces, water and land (Fig. 5c). ersee (Naturschutzbund Deutschland-Naturschutzzentrum Because the determination of the threshold value is Federsee). The collection of data is carried out manually. In ambiguous for the MNDWI data examined using the adap- the WTD data set, only the corresponding calendar weeks tive threshold value procedure explained above, a fixed are noted, not the exact day of the measurement. threshold value of −0.25 is chosen. The threshold value The year 2019 was one of the four hottest years in the results from the visual interpretation of the MNDWI images. state of Baden-Württemberg since the beginning of consist- The images were divided into two segments, land and ent weather records in 1881. The water reserves in the soil, lake, using the threshold (Fig. 5c, d). To determine the lake which were greatly reduced by the drought year 2018, were size, only the pixels that fulfils the following conditions were not able to fully regenerate in 2019 (LUBW 2020). considered as lake area: ‘pixel-value < threshold and located in the lake section’ for and ‘pixel-value > threshold and 3.3 Lake Size Measurement located in the lake section’ for MNDWI. The size of the lake mask is therefore directly depend- A quick and easy way to detect the open water area via ent on the threshold. Since there is no reference data for 0 0 the combination of the different features ( -VH, -VV, the lake size, we evaluate the measurements by forming MNDWI) is by using the machine learning algorithm ran- correlations between lake size and lake level. We assume dom forest. However, the random forest algorithm requires that the area of the lake also increases at a higher lake training data, which are not always available and the genera- level. The correlation between lake level and lake size is tion is labour-intensive. Also, analysing the whole model determined using Pearson’s correlation coefficient. and the way the algorithm makes its classification is very challenging. Therefore, we additionally consider the features 1 3 Water Table Depth [cm] Lake Level [cm] 2018-12-31 2018-12-31 2019-01-07 2019-01-07 2019-01-14 2019-01-14 2019-01-21 2019-01-21 2019-01-28 2019-01-28 2019-02-04 2019-02-04 2019-02-11 2019-02-11 2019-02-18 2019-02-18 2019-02-25 2019-02-25 2019-03-04 2019-03-04 2019-03-11 2019-03-11 2019-03-18 2019-03-18 2019-03-25 2019-03-25 2019-04-01 2019-04-01 2019-04-08 2019-04-08 2019-04-15 2019-04-15 2019-04-22 2019-04-22 2019-04-29 2019-04-29 2019-05-06 2019-05-06 2019-05-13 2019-05-13 2019-05-20 2019-05-20 2019-05-27 2019-05-27 2019-06-03 2019-06-03 2019-06-10 2019-06-10 2019-06-17 2019-06-17 2019-06-24 2019-06-24 2019-07-01 2019-07-01 2019-07-08 2019-07-08 2019-07-15 2019-07-15 2019-07-22 2019-07-22 2019-07-29 2019-07-29 2019-08-05 2019-08-05 2019-08-12 2019-08-12 2019-08-19 2019-08-19 2019-08-26 2019-08-26 2019-09-02 2019-09-02 2019-09-09 2019-09-09 2019-09-16 2019-09-16 2019-09-23 2019-09-23 2019-09-30 2019-09-30 2019-10-07 2019-10-07 2019-10-14 2019-10-14 2019-10-21 2019-10-21 2019-10-28 2019-10-28 2019-11-04 2019-11-04 2019-11-11 2019-11-11 2019-11-18 2019-11-18 2019-11-25 2019-11-25 2019-12-02 2019-12-02 2019-12-09 2019-12-09 2019-12-16 2019-12-16 2019-12-23 2019-12-23 2019-12-30 2019-12-30 PFG Sentinel-1 Data Sentinel-2 Data In-Situ Measurement: Lake Level Dataset Processing Processing Extraction Lake Level Adaptive Thresholding Thresholding Scale & invert Dempster-Shafer Adaptive Random Forest Theory Adaptation Thresholding Correlation Analysis r r r VH VV MNDWI Fig. 4 Overview of the methods and data for lake size measurement 0 0 Fig. 5 Method adaptive thresholding a Sentinel-1 data: -VH (dB), pixel (right), d created lake mask of -VH on 2019-07-13, blue: lake b red polygon: lake section, c histogram of -VH of the lake sec- (pixel-value < threshold and located in the lake section), red: misclas- tion with a threshold at −22 dB separating lake-pixel (left) from land- sified area (pixel-value < threshold but not located in the lake section) 0 0 Fusion using weighted features Furthermore, we gen- Fusion image =G ⋅ + G ⋅ VH VV VH VV (3) erate a fusion image by weighting and combining the + G ⋅ MNDWI MNDWI three features. In the first step, the data must be scaled so that features with a large range of values do not domi- The weighting G is set as 1 + r and results from the cor- nate the fusion result. The images are scaled to a value relation coefficients (r ) between lake area and lake level range between zero and one. Due to the feature properties, (Table 2). This ensures that a data set which is considered MNDWI high values are characteristic for lake area, while more accurate has a greater impact on the overall result than for water areas have a small value. Hence, to combine a data set which is considered less accurate. The adaptive the data sets in a meaningful way, part of the data sets thresholding method is then applied on the fusion image. must be inverted. We calculate the fusion image via the The threshold is set at the minimum between the two peaks following expression: of the histogram, followed by the calculation of the lake size as described above. 1 3 2019-01-01 2019-01-13 2019-01-25 2019-02-06 2019-02-18 2019-03-02 2019-03-14 2019-03-26 2019-04-07 2019-04-19 2019-05-01 2019-05-13 2019-05-25 2019-06-06 2019-06-18 2019-06-30 2019-07-12 2019-07-24 2019-08-05 2019-08-17 2019-08-29 2019-09-10 2019-09-22 2019-10-04 2019-10-16 2019-10-28 2019-11-09 2019-11-21 2019-12-03 2019-12-15 Lake Masks 2019-12-27 0 Lake Level [cm] σ -VH (dB) σ -VH (dB) Lake Masks σ -VV (dB) σ -VV (dB) Lake Masks MNDWI MNDWI Class Maps Lake Masks Fusion Image Fusion Image Lake Masks Random Forest Confidence Image PFG Dempster–Shafer theory adaptation Now, the Dempster– the well. The correlation between WTD and or NDWI is Shafer theory is applied to analyse the credibility of the determined using Pearson’s correlation coefficient. created lake masks. The Dempster–Shafer theory of belief The analysis of WTD in our project is limited to indi- functions is based on Dempster’s work on upper and lower vidual points in time over the course of a year. The spatial probabilities in the 1960s and Shafer’s monograph in the correlation between the in situ WTD values and the remote 1970s (Yager 2008). A special feature of the Dempster– sensing data is analysed. In order to analyse the temporal Shafer theory is the evidence that allows uncertainty to be course, which would be interesting for the monitoring of expressed explicitly. An additional advantage of the theory renaturation, it would make sense to generate time series is that it has a combination rule to combine evidence from over several years and with all available SAR data sets. Since multiple independent items. The theory is usually used when the quality of the in situ data is limited (no exact date of the credibility of the individual items is known and the sig- measurement given, only the calendar week) and in order nificance of a common hypothesis is to be determined. In to keep the processing effort low, we only analyse selected this project, the three lake masks (based on the adaptive individual points in time for their spatial correlation in this 0 0 thresholding method of -VH, -VV and MNDWI) serve work. as independent items. However, the credibility of the masks, represented by a probability number, is not known and can- not be determined due to the lack of reference data. There- 4 Results fore, we create two different scenarios of credibility for each mask and analyse the resulting mass function of the com- 4.1 Results on Lake Size mon hypothesis for realism (Table 2). If we consider the two states (lake pixel or non lake pixel) and the three lake masks The results of the lake size measurements with the different ( Mask , Mask , Mask ), each of which giving 0 0 methods are presented in Table 1. The area of the masks is -VH -VV MNDWI an independent statement, eight different classes of combi- calculated by multiplying the number of pixels inside the nations result (Table 3). Assuming the considered pixel is a mask by their size. lake pixel, the mass functions m(A) listed in Table 3 are cal- The largest lake masks are those from the radar data. The culated using the basic probability numbers from Table 2 for smallest are those of class 1. This is logical, since the condi- both scenarios and the formulas to mass functions according tion of class 1 is that a pixel must be classified as a lake by to Yager (2008). all three lake masks that use a single data set. The standard The scenarios are presented in Sect. 4.1. Based on the deviations of the areas range between 0.016 km ( Mask ) FI results, a statement is to be made which scenario achieved a and 0.037 km ( Mask ). For the analysis of the lake MNDWI more trustworthy image. Scenario 1 is based on the results area measurements, we evaluate which method captures of the fusion image, scenario 2 is based on the results of the most pixels that lie within the lake section and the least random forest. A scenario with high probabilities for lake pixels that lie outside this area and are therefore defined as pixels in the lake section and low probabilities for lake pixels misclassified. Furthermore, it is considered which method in the surrounding area is considered probable. has the highest correlation with the lake level. The correla- tion is determined only for the lake masks using one data set. Since the satellite images from Sentinel-1 and Sentinel-2 are 3.4 Water Table Depth Estimation often not taken on the same day, the fusion methods cannot be assigned an exact lake level for a given day. This would Next, we present the methodology used to explore the rela- reduce the significance of the correlation. tionship between the peatland WTD and VH and VV polar- Some of the histograms prove to be suitable for estab- ized as well as the relationship between WTD and NDWI. lishing a separation between lake and land. The histograms Unlike previous approaches the use of a speckle filter was show two peaks, which can be easily distinguished from dispensed with in the preprocessing of the SAR data. The each other and can be assigned to the land or water area reason for this was to counteract a falsification of of the (Fig. 6). The threshold varies nearly every month. The meas- cell covering the water table depth measurement point. In ured values for the water areas are in the range of −25 to −15 contrast to the measurement of lake area, which rely on a dB for the VV polarisation and −22 to −32 dB for the VH certain homogeneity of adjacent pixels, this method is inter- polarisation. The average thresholds are −22.4± 1.1 (VH) ested in the exact values at the given well coordinates. and −16.1± 1.9 (VV). For every observation date we extract data from the WTD The measurement of the lake area via the adaptive thresh- data set measured in the same week. Also we extract pixel- old method using VH polarised shows a higher correla- values from the satellite images. For each observation well, tion with the lake level (r = 0.72) than VV polarised (r = we use only the single pixel-values that cover the location of 1 3 PFG Table 1 Results of lake size [km ] using the different lake masks with mean (Ø) and standard deviation (SD) Acquisition dates, Sentinel-1/Sentinel-2 Mask 0 Mask 0 Mask Mask Mask Mask -VH -VV MNDWI FI RF Class1 January 19th/– 1.463 – – – – – February 19th/February 21th 1.412 1.415 1.422 1.402 1.383 1.269 March 15th/March 21th 1.411 – 1.367 – – – April 20th/April 20th 1.400 1.404 1.377 1.356 1.322 1.338 May 20th/May 25th 1.400 1.369 1.447 1.388 1.375 1.335 June 18th/June 19th 1.404 1.433 1.370 1.397 1.368 1.320 July 13th/July 4th 1.406 1.421 1.380 1.395 1.358 1.331 August 18th/August 12th 1.410 1.406 1.362 1.382 1.367 1.350 September 16th/September 19th 1.403 1.400 1.361 1.401 1.370 1.310 October 11th/October 14th 1.421 1.390 1.326 1.385 1.305 1.307 November 15th/– 1.412 1.376 – – – – December 16th/December 3rd 1.427 1.394 1.328 1.367 1.406 1.310 Ø 1.414 1.401 1.374 1.386 1.362 1.319 SD 0.017 0.020 0.037 0.016 0.031 0.024 0 0 The lake masks Mask , Mask , Mask and Mask are created using adaptive thresholding and the data sets -VH, -VV, MNDWI 0 0 -VH -VV MNDWI FI and the fusion image (FI). The lake mask Mask is created using the random forest algorithm (RF). The lake mask Mask is the area, that RF Class1 meets the conditions of class 1 of the Dempster–Shafer theory adaption. For all masks, only the areas that lie within the lake section were con- sidered Table 2 Correlation coefficient r for the relation between lake size which the lake mask was classified as the most important feature by (from adaptive thresholding) and lake level, basic probability num- the random forest algorithm) and testing feature importance score bers for scenario 1 (S1) and scenario 2 (S2) for analysing lake size (average feature accuracy output from the random forest method in according to Dempster–Shafer theory, results of the random forest SNAP) (RF) algorithm in SNAP: main feature (number of evaluations in Data set r Basic probability numbers RF: main feature RF: importance Basic prob- S1 ability num- bers S2 Mask 0 0.72 0.9 2/9 0.04 0.7 -VH Mask 0.46 0.7 2/9 0.02 0.6 -VV Mask 0.61 0.8 5/9 0.12 0.9 MNDWI 0.46). The determined lake size over the fixed threshold of with the lake level. The lake mask that achieves the high- −0.25 also shows a relatively strong correlation (r = 0.61) est correlation with lake level is assigned a basic prob- (Fig. 7). ability number of 0.9, the lake mask with the next highest In terms of the misclassified areas in the whole image correlation is downgraded by 0.1 to 0.8, just as the third section, the method of the weighted fusion image and the lake mask is downgraded to 0.7. The distance between the random forest algorithm showed the best results (both basic probability numbers also approximately mirrors the Ø = 0.2%). The misclassified areas for the adaptive thresh- distance between the determined correlation coefficients. 0 0 olding method for -VH, -VV and MNDWI are signifi- In scenario 2 (S2) the basic probability numbers are set cantly larger (Ø = 7.0%, Ø = 6.6% and Ø = 2.2%) (see according to the results of the random forest algorithm in Fig. 5d). SNAP. The feature (here lake mask) which the algorithm In addition to the label image, the random forest outputs selects to be more relevant for the classification is evalu- a confidence image. This shows an increased uncertainty in ated. In this case, the algorithm classifies one lake mask as the lakeshore area (Fig. 8). the most important feature in five out of nine months with Next we consider the analysis of the areas following an importance score of 0.12. This lake mask is assigned the Dempster–Shafer theory of belief functions. First, we a basic probability number of 0.9 and the mask with the distinguish two scenarios: In scenario 1 (S1), as with the lowest result is set to 0.6. A linear equation can be calcu- weighted fusion image, the basic probability numbers are lated from these two pairs of values that can be used to set based on the correlation of the calculated lake areas 1 3 PFG Table 3 Classes, conditions and mass functions m(A) after Demp- can be clearly assigned to the lakeshore region. Refer- ster–Shafer theory for scenario 1 (S1) and scenario 2 (S2), m(A) is ring to the condition of class 2 (MNDWI = no lake, calculated using the basic probability numbers from Table 2 and the 0 0 -VH = lake, -VV = lake) the lake masks include the formulas to mass functions according to Yager (2008) lakeshore region, while the MNDWI mask excludes this Class Mask 0 Mask 0 Mask m(A) m(A) -VH -VV MNDWI S1 S2 region. 0 0 Class 3 (MNDWI = lake, -VH = lake, -VV = 1 Lake pixel Lake pixel Lake pixel 99% 99% no lake) has the second highest mass function after sce- 2 Lake pixel Lake pixel Non lake 87% 41% pixel nario 1 with 94%. The pixels that meet this condition are 3 Lake pixel Non lake Lake pixel 94% 93% in the lake area, but they are small areas compared to class pixel 1 (Ø 0.028 km ). 0 0 4 Lake pixel Non lake Non lake 35% 8% Class 8 (MNDWI = lake, -VH = no lake, -VV = pixel pixel lake) has the second highest mass function in scenario 2 5 Non lake Non lake Non lake 1% 1% at 88%. This class has less area than class 3 (Ø 0.007 km ) pixel pixel pixel and mostly pixels in the lakeshore region. 6 Non lake Lake pixel Non lake 4% 4% For a more accurate assessment, the following considers pixel pixel how much class 1 and 3 (S1), and class 1 and 8 (S2) do 7 Non lake Non lake Lake pixel 11% 53% pixel pixel not lie in the lake section and thus counts as misclassified 8 Non lake Lake pixel Lake pixel 61% 88% area. Scenario 1, which is based on the weighted fusion pixel image, has less misclassification on average (Ø = 0.11%). Scenario 2, which is based on the classification according to the random forest algorithm, misclassifies on average determine the value for the third mask. This results in a a slightly higher area surrounding the lake as lake (Ø = value of 0.7 (Table 2). 0.13%). For the evaluation, the areas with the calculated mass functions are considered and it is analyzed which scenario is estimated to be more likely. Pixels that are recognized as 4.2 Results on Water Table Depth Estimation lake in all three masks have the highest body of evidence 2 0 of being a lake pixel (99% in both scenarios, Ø 1.319 km ). The correlation coefficients between data and WTD are Here, as in the previous fusion image, it is evident that the shown in Fig. 11. The observation of the correlation coef- combination of the various data sets can highly reduce the ficients mainly shows no or weak correlations between misclassified areas (Fig. 9). the WTD and . Moreover, no stable correlation can be Especially relevant for the evaluation are class 2, observed over all data sets. If only the observation wells because the difference between the determined mass func- of the northern Federseeried are used for the evaluation, tion is the highest (S1: 87%, S2: 41%), and classes 3 and the correlations stabilise (Fig. 11b). 8, because they each have the second highest values. Fig- It can be seen that the correlation coefficients from Jan- ure 10 shows the areas that fulfil the condition according uary to May for both polarisation are in a similar range of to class 2 (Ø 0.080 km ). It is noticeable that especially on the days in September, October and December the areas 2019-09-16, peaks, minimum between the peaks and threshold are Fig. 6 Example histograms of the lake section for a -VH 2019-10- more difficult to identify 11, the histogram shows two peaks, which can be easily distinguished from each other and are assigned to the land or water area. b -VV 1 3 PFG 1.470 1.440 1.460 01-19 1.460 1.430 1.440 02-21 1.450 1.420 1.420 1.440 1.410 1.400 12-16 1.430 1.400 1.380 1.420 1.390 1.360 1.410 1.380 06-18 1.340 1.400 1.370 12-02 10-14 1.390 1.360 1.320 55 60 65 70 75 80 55 60 65 70 75 55 60 65 70 75 80 85 Lake Level [cm] Lake Level [cm] Lake Level [cm] (a) (b) (c) VH Linear (VH) VV Linear (VV) MNDWI Linear (MNDWI) Fig. 7 Relation between lake level [cm] and measured lake size r = 0.72, r = 0.46, r = 0.61, the outliers are labelled with VH VV MNDWI [km ] from a Mask 0 , b Mask 0 and c Mask by adap- the data, all data are related to the study period 2019 -VH -VV MNDWI tive thresholding method, resulting Pearson’s correlation coefficients: values (about 0.2–0.4), with the exception of the values of 5 Discussion the VV polarisation on March 15th, which, however, gen- erally show conspicuousness. The VH polarisation shows Looking at the correlation between the lake masks of the stronger correlations with the WTD than the VV polarisa- individual data sets and the lake level, we see that the -VH tion. On June 18th and July 13th, a negative correlation is data has the highest correlation with 0.7. The fact that the calculated for both polarisation (− 0.10 and −0.29 ). F rom measurement of water area via the adaptive thresholding of August to December both polarisation showed a large vari- VH polarisation has the highest correlation with lake level ability in correlation coefficients ( −0.22 to 0.46). and scenario 1 which weights -VH stronger has a slightly Next, the correlations between WTD and NDWI are lower misclassification leads us to the statement that this presented (Fig. 12). A consideration of only the northern polarisation is more suitable for monitoring the peatland in Federseeried with lower vegetation is also carried out here this project. This die ff rs from earlier assumptions from other (Fig. 12b). The correlation coefficients are higher in Sep- projects such as Huang and Jin (2019). The better suitability tember 19th, October 14th and December 3rd (0.41, 0.44, of VH polarisation could be due to the fact that it is not as 0.41) and lower on February 21th and March 21th (0.25). susceptible to the double bounce effect as VV polarisation Again, the opposite behaviour of June and August (− 0.44 , caused by the reeds on the lake. −0.33 ) is conspicuous. The correlation could have increased if more time points had been included in the evaluation. Schwatke et al. (2019) use for the measurement of lake size five different water indices for the time period between 1984 and 2018. They improve correlations between lake level and the lake size they measured on average from 0.610 of up to 0.834 with their method of filling data gaps by using a long-term water probability mask. Our work shows that lake size can also achieve high correlations via radar measurements. Schwatke et al. (2019) describes, as we do, that it was not possible to read a suitable threshold for MNDWI from the histograms. They solve this problem by developing a new approach for automated threshold computation by combining the informa- tion of all five indexes. The challenge of finding a suitable threshold for masking the water area is also described by Pulvirenti et al. (2011). They suggest that the problem with Fig. 8 Random forest results of 2019-12-16 (Sentinel-1) and 2019- setting the threshold for SAR data is based on the roughen- 12-03 (Sentinel-2). a Label image, purple: lake. b Confidence image shows an increased uncertainty in the shore area of the lake ing effect, which is caused by wind speed. In our case, the 1 3 Lake Size [km²] Lake Size [km²] Lake Size [km²] PFG Fig. 9 Lake masks for the entire study area in black (without the Theory adaption, Condition: Mask = lake, Mask = lake, MNDWI -VH lake section-condition), Sentinel-1 data from July 13th and Senti- Mask = lake. m(A) : 99%, m(A) : 99%. The fusion of all three -VV S1 S2 nel-2 data from July 4th. a Mask , b Mask , c Mask , data sets results in almost no areas outside the lake being misclassi- 0 0 -VH -VV MNDWI d Mask from the fusion image (FI), e Mask from the random for- fied as lake FI RF est (RF) classification, f Mask . Class 1 after Dempster–Shafer- Class1 water surface is mainly rough because of the dense plant data sets. In situ data are used to create the fusion image, cover of reeds. For January and March it was not possible but not for the random forest classification. The fact that to find a suitable threshold in our SAR data. In January this the classification using the random forest algorithm yields was probably due to the ice cover on the lake. In March we comparable results to the thresholding of the fusion image can only explain the noisy SAR image by a sensor problem. suggests that a good estimate of the lake area is possible In general, to improve the thresholding an automated and without in situ data. However, it must be noted that even the mathematically more accurate method for adaptive thresh- fusion image does not necessarily provide the best estimate olding can be applied. of the lake area, because it is questioned whether a rigid Now, we will consider the fusion of the data sets. First of weighting over the entire year is useful, since it is shown all, it has to be said that problems can occur when combin- that the significance of the different data sets varies strongly ing data from different satellites, because e.g. the overflight over the year (Table 1). Therefore, the random forest with its times do not match and a time gap occurs. Second, some adjusted weighting may even be more appropriate. Sentinel-2 data cannot be used due to cloud cover. In addi- Considering both the random forest algorithm and the tion, there are uncertainties resulting from the different spa- Dempster–Shafer theory, it is noticeable that the lakeshore tial resolution of the data (Sentinel-1 5 m × 20 m; Sentinel-2 region has a large uncertainty in the classification. Schwatke VIS: 10 m, SWIR: 20 m; WTD point data). Nevertheless, et al. (2019) also see uncertainties in the lakeshore area in the results show that combining the data is superior to using the analysis of the five water indices. They consider the rea- one single data set. The Dempster–Shafer approach has also son therefore is mainly the different colour in shallow water. shown that combining both satellite data gives a very reli- However, this phenomenon also occurred in our study when able estimate of the lake area. The weighted fusion image comparing radar data with a water index. Class 2 clearly and the random forest algorithm, both using Sentinel-1 and shows that the MNDWI mask is smaller than the radar Sentinel-2 data, show in terms of misclassification that they masks. Martinis et al. (2022) create water masks based on are better suited for water area detection than the individual analysis of Sentinel-2 data and use Sentinel-1 data to fill data 1 3 PFG Fig. 10 Class 2 after Dempster–Shafer–theory adaption (Condition: located in the lakeshore region. The lake masks resulting from the Mask = no-lake, Mask = lake, Mask = lake; m(A) : SAR data measure a larger lake area than the lake masks resulting 0 0 MNDWI -VH -VV S1 87%, m(A) : 41%) Pixels that meet the conditions of class 2 are often from the optical data S2 0.50 0.50 0.40 0.40 0.30 0.30 0.20 0.20 0.10 0.10 0.00 0.00 -0.10 -0.10 -0.20 -0.20 -0.30 -0.30 -0.40 -0.40 -0.50 -0.50 (a) (b) r (VH) r (VV) r (VH) r (VV) 0 0 Fig. 11 Correlation coefficients (r) for water table depth (WTD) and -VH or -VV. a Whole peatland, b only northern Federseeried gaps due to cloud cover. They note that in the shore area, underestimation of the water area. Peña-Luque et al. (2021) when the water is clear, optical sensors can receive signals also present a methodology to generate large-scale water from the bottom of the water body and this can lead to an maps by merging Sentinel-1 and Sentinel-2 data. In contrast 1 3 2019-01-19 2019-02-19 2019-03-15 2019-04-20 2019-05-20 2019-06-18 2019-07-13 2019-08-12 2019-09-16 2019-10-11 2019-11-15 2019-12-16 2019-01-19 2019-02-19 2019-03-15 2019-04-20 2019-05-20 2019-06-18 2019-07-13 2019-08-12 2019-09-16 2019-10-11 2019-11-15 2019-12-16 PFG 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0 0 -0.1 -0.1 -0.2 -0.2 -0.3 -0.3 -0.4 -0.4 -0.5 -0.5 (a)(b) Fig. 12 Correlation coefficients (r) for water table depth (WTD) and NDWI. a Whole peatland, b only northern Federseeried to Martinis et al. (2022) they observe, just like Kaplan et al. characteristics result, which can strongly influence both the (2019), that radar would underestimate the areas more than optical and SAR data. We can therefore confirm the effect optical methods. We cannot confirm the poor penetration of cuts in our data. of the radar signal, which Kaplan et al. (2019) mention as Asmuß et al. (2019) consider soil and vegetation informa- a reason for the different lake sizes from our research. The tion in their method. The difference between the correlation fact that our MNDWI lake masks are smaller than the radar coefficients of the whole peatland and those of the northern masks is probably due to the lower threshold value of −0.25 Federseeried shows that vegetation has a strong influence on instead of 0, which is normally chosen as threshold for the correlation between the backscatter of the SAR measure- MNDWI. In addition to the choice of threshold, we also see ment and the WTD. The vegetation height is much lower in that a clear ecological boundary between land and water is the northern Federseeried and can therefore be penetrated difficult to define in the peatland ecosystem. Higher resolu- more easily by the radar signals. tion and the inclusion of ecological data could reduce these Large parts of the Federseeried can be described as natu- uncertainties. ral peatlands. Although the northern Federseeried started Since the amount of data we analysed regarding WTD is to be rewetted four years ago, our results with the higher small compared to other publications, we discuss our results correlations according to Bechtold et al. (2018) resemble an primarily with the findings from other research rather than agriculturally used drained peatland. deriving conclusions from our data. Asmuß et al. (2019) In further work we could consider more additional infor- obtained a temporal spearman correlation coefficient of mation like Asmuß et al. (2019) and we could extend the 0.45 (± 0.17) between and WTD on three study sites time period and the study area to generate more stable and in a backscatter time series of two years. Bechtold et al. robust information. Another problem we see is the compa- (2018) studied 17 peatlands in Germany over three years and rability of the different spatial resolutions of our data. Given achieved correlations of 0.38 for natural peatlands and 0.54 that the water level measurements are point data and the for drained peatlands used for agriculture. Their spatial reso- remote sensing data are raster data, uncertainty arises. Also, lution is about 100 m (Asmuß et al. 2019) and about 500 m as in the lake size methods a fusion of the remote sensing (Bechtold et al. 2018) is coarser than ours with about 20 m. data can be considered. Similar to us, Asmuß et al. (2019) observe a loss of correlation during summer. In both of our Sentinel data, anomalies can be observed in the summer months of June 18th/19th and August 12th/18th (in Sentinel-1 also in July 6 Conclusion and Outlook 13th). So as mentioned by Asmuß et al. (2019) and also by Bechtold et al. (2018) one possible explanation is mowing. In the field of environmental protection, there is a need for According to the NABU nature conservation centre Federsee evaluating methods, which the actors can use over long peri- the vegetation growing in the northern Federseeried is mown ods of time to identify trends in the protected areas. The use twice a year, in June and August, in order to maintain the of freely available Sentinel data and working with the open- meadows as open landscape and wet meadows. If the mown source software SNAP and QGIS as well as the easy-to-use grass is left on the meadow to dry, entirely different surface random forest algorithm make this possible. 1 3 2019-02-21 2019-03-21 2019-04-20 2019-05-25 2019-06-19 2019-07-04 2019-08-18 2019-09-19 2019-10-14 2019-12-03 2019-02-21 2019-03-21 2019-04-20 2019-05-25 2019-06-19 2019-07-04 2019-08-18 2019-09-19 2019-10-14 2019-12-03 PFG Our project shows despite the good suitability of VH References polarised that the meaningful combination of different Asmuß T, Bechtold M, Tiemeyer B (2019) On the potential of Senti- remote sensing data is superior to the consideration of single 0 0 nel-1 for high resolution monitoring of water table dynamics in features. The two methods that combined -VH, -VV and grasslands on organic soils. Remote Sens 11(14):1–19. https://doi. MNDWI, the weighted fusion image and the random forest org/ 10. 3390/ rs111 41659 algorithm, produced comparable results. We consider the Baghdadi N, Zribi M (eds) (2016) Land surface remote sensing in continental hydrology, Remote sensing observations of conti- replacement of lake level measurement by lake size meas- nental surfaces set, vol 4. ISTE Press and Elsevier, London and urement through remote sensing to be a suitable alternative. Kidlington, Oxford Potentials lie in the inclusion of further environmental con- Bagwan WA, Sopan Gavali R (2021) Dam-triggered land use land ditions such as temperature and precipitation, in the com- cover change detection and comparison (transition matrix method) of Urmodi River Watershed of Maharashtra, India: a remote sens- bination with biological and ecological knowledge, and in ing and GIS approach. Geol Ecol Landsc. https://doi. or g/10. 1080/ the application of other machine learning or deep learning 24749 508. 2021. 19527 62 methods. This way, the complex interrelationships within the Bechtold M, Schlaffer S, Tiemeyer B, de Lannoy G (2018) Inferring ecosystem could be learned and not only could the current water table depth dynamics from ENVISAT-ASAR C-band back- scatter over a range of peatlands from deeply-drained to natural state be represented, but also could the future performance conditions. Remote Sens 10(4):2–21. https:// doi. or g/ 10. 3390/ of the peatland be estimated. We see measuring water levels rs100 40536 in organic soils as a big task and have identified the compa- Bhunia GS (2021) Assessment of automatic extraction of surface rability of different spatial resolutions, the effects of peatland water dynamism using multi-temporal satellite data. Earth Sci Inf 14(3):1433–1446. https:// doi. org/ 10. 1007/ s12145- 021- 00612-7 management (e.g. grass cuts) and the acquisition of suitable Boni G, Ferraris L, Pulvirenti L, Squicciarino G, Pierdicca N, Candela in situ data for further research as future challenges. L, Pisani AR, Zoffoli S, Onori R, Proietti C, Pagliara P (2016) A The advantage of remote sensing is that data are collected prototype system for flood monitoring based on flood forecast worldwide and therefore methods that use this data can eas- combined with COSMO-SkyMed and Sentinel-1 data. IEEE J Sel Top Appl Earth Obs Remote Sens 9(6):2794–2805. https:// ily be applied all over the world. However, environmen- doi. org/ 10. 1109/ JSTARS. 2016. 25144 02 tal conditions are diverse and vary from region to region. Breiman L (2001) Random Forests. Mach Learn 45(1):5–32. https:// Whether methods that have been tested successfully in one doi. org/ 10. 1023/A: 10109 33404 324 region can be transferred to other regions is not certain. The Bryant R, Baird A (2003) The spectral behaviour of Sphagnum cano- pies under varying hydrological conditions. Geophys Res Lett challenge will be to develop unified procedures and provide 30:1134–1138. https:// doi. org/ 10. 1029/ 2002G L0160 53 accurate and comparable results despite the complexity of Cazals C, Rapinel S, Frison PL, Bonis A, Mercier G, Mallet C, Corgne ecosystems. So remote sensing can be a tool to create data- S, Rudant JP (2016) Mapping and characterization of hydrologi- bases, which can be used for example for recording the cli- cal dynamics in a coastal marsh using high temporal resolution Sentinel-1A images. Remote Sens 8(7):1–17. https:// doi. org/ 10. mate protection services of ecosystems. The comparability 3390/ rs807 0570 of the results is particularly relevant because climate protec- Dong Z, Wang Z, Liu D, Song K, Li L, Jia M, Ding Z (2014) Map- tion services are to be traded on a CO market in the future ping wetland areas using Landsat-derived NDVI and LSWI: a (Michel 2021). case study of West Songnen Plain, Northeast China. J Indian Soc Remote Sens 42(3):569–576. h t t p s : / / d o i . o r g / 1 0 . 1 0 0 7 / s12524- 013- 0357-1 Acknowledgements We thank the NABU-Naturschutzzentrum Feder- Guo M, Li J, Sheng C, Xu J, Wu L (2017) A review of wetland remote see for the support to make this research project possible. Many thanks sensing. Sensors. https:// doi. org/ 10. 3390/ s1704 0777 to Dr. Katrin Fritzsch, Head of NABU-Naturschutzzentrum Federsee, Haque MI, Basak R (2017) Land cover change detection using GIS and and to Judith Engelke from Regierungspräsidium Tübingen for giving remote sensing techniques: a spatio-temporal study on Tanguar us access to in situ data. Haor, Sunamganj, Bangladesh. Egypt J Remote Sens Space Sci 20(2):251–263. https:// doi. org/ 10. 1016/j. ejrs. 2016. 12. 003 Funding Open Access funding enabled and organized by Projekt Harris A, Bryant RG (2009) A multi-scale remote sensing approach for DEAL. monitoring northern peatland hydrology: present possibilities and future challenges. J Environ Manag 90(7):2178–2188. https://doi. Open Access This article is licensed under a Creative Commons Attri- org/ 10. 1016/j. jenvm an. 2007. 06. 025 bution 4.0 International License, which permits use, sharing, adapta- Huang M, Jin S (2019) Water level and morphological changes of wet- tion, distribution and reproduction in any medium or format, as long lands in the Poyang Lake using Sentinel-1 data. In: 2019 photonics as you give appropriate credit to the original author(s) and the source, and electromagnetics research symposium—Fall, pp 3159–3163. provide a link to the Creative Commons licence, and indicate if changes https:// doi. org/ 10. 1109/ PIERS- Fall4 8861. 2019. 90213 03 were made. The images or other third party material in this article are Huang C, Chen Y, Zhang S, Wu J (2018) Detecting, extracting, and included in the article's Creative Commons licence, unless indicated monitoring surface water from space using optical sensors: a otherwise in a credit line to the material. If material is not included in review. Rev Geophys 56(2):333–360. h t t p s : / / d o i . o rg / 1 0 . 1 0 2 9 / the article's Creative Commons licence and your intended use is not 2018R G0005 98 permitted by statutory regulation or exceeds the permitted use, you will Joosten H, Clarke D (2002) Wise use of mires and peatlands: back- need to obtain permission directly from the copyright holder. To view a ground and principles including a framework for decision-making. copy of this licence, visit http://cr eativ ecommons. or g/licen ses/ b y/4.0/ . Internat, Mire Conservation Group, Totnes 1 3 PFG Kaplan G, Avdan U (2018) Sentinel-1 and Sentinel-2 data fusion for the Biebrza Floodplain (Poland). Remote Sens 10(2):1–19. https:// wetlands mapping: Balikdami, Turkey. Int Arch Photogramm doi. org/ 10. 3390/ rs100 10078 Remote Sens Spat Inf Sci XLII-3:729–734. https:// doi. org/ 10. Moser L, Schmitt A, Wendleder A, Roth A (2016) Monitoring of the 5194/ isprs- archi ves- XLII-3- 729- 2018 Lac Bam Wetland extent using dual-polarized X-band SAR data. Kaplan G, Yigit Avdan Z, Avdan U (2019) Mapping and monitoring Remote Sens 8(4):302. https:// doi. org/ 10. 3390/ rs804 0302 wetland dynamics using thermal, optical, and SAR remote sensing Muro J, Canty M, Conradsen K, Hüttich C, Nielsen A, Skriver H, Remy data. In: Gökçe D (ed) Wetlands management—assessing risk and F, Strauch A, Thonfeld F, Menz G (2016) Short-term change sustainable solutions. IntechOpen, pp S.87–107 detection in wetlands using Sentinel-1 time series. Remote Sens Kim JW, Lu Z, Gutenberg L, Zhu Z (2017) Characterizing hydrologic 8(10):795. https:// doi. org/ 10. 3390/ rs810 0795 changes of the Great Dismal Swamp using SAR/InSAR. Remote Peña-Luque S, Ferrant S, Cordeiro MCR, Ledauphin T, Maxant J, Sens Environ 198:187–202. https:// doi. org/ 10. 1016/j. rse. 2017. Martinez JM (2021) Sentinel-1 &2 multitemporal water surface 06. 009 detection accuracies, evaluated at regional and reservoirs level. LGRB (2021) Federseeried. https://lg rbwissen. lg rb-b w.de/ g eotour ism Remote Sens 13(16):3279. https:// doi. org/ 10. 3390/ rs131 63279 us/ moore/ feder seeri ed. Accessed 28 Jan 2022 Pulvirenti L, Pierdicca N, Chini M, Guerriero L (2011) An algorithm Liao A, Chen L, Chen J, He C, Cao X, Chen J, Peng S, Sun F, Gong for operational flood mapping from Synthetic Aperture Radar P (2014) High-resolution remote sensing mapping of global land (SAR) data using fuzzy logic. Nat Hazard 11(2):529–540. https:// water. Sci China Earth Sci 57(10):2305–2316. https://d oi.o rg/1 0. doi. org/ 10. 5194/ nhess- 11- 529- 2011 1007/ s11430- 014- 4918-0 Schwatke C, Scherer D, Dettmering D (2019) Automated extraction Liu S, Qi Z, Li X, Yeh A (2019) Integration of convolutional neural of consistent time-variable water surfaces of lakes and reser- networks and object-based post-classification refinement for land voirs based on Landsat and Sentinel-2. Remote Sens 11(9):1010. use and land cover mapping with optical and SAR data. Remote https:// doi. org/ 10. 3390/ rs110 91010 Sens 11(6):1–25. https:// doi. org/ 10. 3390/ rs110 60690 Succow M, Joosten H (eds) (2012) Landschaftsökologische LUBW (2020) Wieder außergewöhnlich warm und heiß, mit Nach- Moorkunde: Mit 10 Farbbildern, 223 Abbildungen, 136 Tabellen wirkungen des Trockenjahrs 2018: Eine klimatische Einordnung im Text sowie auf 2 Beilagen, 2nd edn. Schweizerbart Science des Jahres 2019 für Baden-Württemberg. https:// pd. lub w . de/ Publishers, Stuttgart 10102. Accessed 29 June 2022 Tsyganskaya V, Martinis S, Marzahn P, Ludwig R (2018) Detection of Martinis S, Kersten J, Twele A (2015) A fully automated TerraSAR-X temporary flooded vegetation using Sentinel-1 time series data. based flood service. ISPRS J Photogramm Remote Sens 104:203– Remote Sens. https:// doi. org/ 10. 3390/ rs100 81286 212. https:// doi. org/ 10. 1016/j. isprs jprs. 2014. 07. 014 Tubiello F, Biancalani R, Salvatore M, Rossi S, Conchedda G (2016) A world- Martinis S, Groth S, Wieland M, Knopp L, Rättich M (2022) Towards a wide assessment of greenhouse gas emissions from drained organic soils. global seasonal and permanent reference water product from Sen- Sustainability 8(4):371. https:// doi. org/ 10. 3390/ su804 0371 tinel-1/2 data for improved flood mapping. Remote Sens Environ UBA (2021) Submission under the United Nations framework conven- 278:113077. https:// doi. org/ 10. 1016/j. rse. 2022. 113077 tion on climate change and the Kyoto Protocol 2021. National McFeeters SK (1996) The use of the Normalized Difference Water Inventory Report for the German Greenhouse Gas Inventory Index (NDWI) in the delineation of open water features. Int J 1990–2019 Remote Sens 17(7):1425–1432. https:// doi. org/ 10. 1080/ 01431 White L, Brisco B, Dabboor M, Schmitt A, Pratt A (2015) A collec- 16960 89487 14 tion of SAR methodologies for monitoring wetlands. Remote Sens Meingast KM, Falkowski MJ, Kane ES, Potvin LR, Benscoter BW, 7(6):7615–7645. https:// doi. org/ 10. 3390/ rs706 07615 Smith AM, Bourgeau-Chavez LL, Miller ME (2014) Spectral Wilson D, Blain D, Couwenberg J (2016) Greenhouse gas emission detection of near-surface moisture content and water-table posi- factors associated with rewetting of organic soils. Mires Peat tion in northern peatland ecosystems. Remote Sens Environ 17:1–28. https:// doi. org/ 10. 19189/ MaP. 2016. OMB. 222 152:536–546. https:// doi. org/ 10. 1016/j. rse. 2014. 07. 014 Xu H (2006) Modification of Normalised Difference Water Index Michel J (2021) Das bedeutet das EU-Klimagesetz für Landwirte: Kli- (NDWI) to enhance open water features in remotely sensed maschutz und nachhaltige Investitionen. https://www .ag rarheute. imagery. Int J Remote Sens 27(14):3025–3033. https:// doi. org/ c om/ p ol it i k/ b ed e u t e t - e u- k li ma g es e t z- f ue r- l an dw i r t e - 58 04 90. 10. 1080/ 01431 16060 05891 79 Accessed 28 Jan 2022 Yager RR (ed) (2008) Classic works of the Dempster–Shafer theory Mleczko M, Mróz M (2018) Wetland mapping using SAR data from of belief functions, Studies in fuzziness and soft computing, vol the Sentinel-1A and TanDEM-X missions: a comparative study in 219. Springer, Berlin 1 3
"PFG – Journal of Photogrammetry, Remote Sensing and Geoinformation Science" – Springer Journals
Published: Sep 6, 2022
Keywords: Fusion; SAR; NDWI; Peatland; Water area; Water table depth
Access the full text.
Sign up today, get DeepDyve free for 14 days.