Target‐oriented habitat and wildlife management: estimating forage quantity and quality of semi‐natural grasslands with Sentinel‐1 and Sentinel‐2 data

Abstract

Semi‐natural grasslands represent ecosystems with high biodiversity. Their conservation depends on the removal of biomass, for example, through grazing by livestock or wildlife. For this, spatially explicit information about grassland forage quantity and quality is a prerequisite for efficient management. The recent advancements of the Sentinel satellite mission offer new possibilities to support the conservation of semi‐natural grasslands. In this study, the combined use of radar (Sentinel‐1) and multispectral (Sentinel‐2) data to predict forage quantity and quality indicators of semi‐natural grassland in Germany was investigated. Field data for organic acid detergent fibre concentration (oADF), crude protein concentration (CP), compressed sward height (CSH) and standing biomass dry weight (DM) collected between 2015 and 2017 were related to remote sensing data using the random forest regression algorithm. In total, 102 optical‐ and radar‐based predictor variables were used to derive an optimized dataset, maximizing the predictive power of the respective model. High R2 values were obtained for the grassland quality indicators oADF (R2 = 0.79, RMSE = 2.29%) and CP (R2 = 0.72, RMSE = 1.70%) using 15 and 8 predictor variables respectively. Lower R2 values were achieved for the quantity indicators CSH (R2 = 0.60, RMSE = 2.77 cm) and DM (R2 = 0.45, RMSE = 90.84 g/m²). A permutation‐based variable importance measure indicated a strong contribution of simple ratio‐based optical indices to the model performance. In particular, the ratios between the narrow near‐infrared and red‐edge region were among the most important variables. The model performance for oADF, CP and CSH was only marginally increased by adding Sentinel‐1 data. For DM, no positive effect on the model performance was observed by combining Sentinel‐1 and Sentinel‐2 data. Thus, optical Sentinel‐2 data might be sufficient to accurately predict forage quality, and to some extent also quantity indicators of semi‐natural grassland.

Spatiotemporal Patterns of Pasture Quality Based on NDVI Time-Series in Mediterranean Montado Ecosystem

“…The difference between both tools (with a slight advantage of PS over RS) can be explained by the fact that there is a temporal gap between the date of RS capture and the date of field biomass collection. These determination coefficients are similar to those obtained for pasture quality indicators (CP and NDF) by Pullanagari et al [26] from proximal sensors (R 2 of 0.80) and higher than those recorded from satellite images for example by Lugassi et al [19] or Raab et al [23] (R 2 about 0.70), Fernández-Habas et al [4] (R 2 about 0.65), or Zhao et al [27] (R 2 about 0.60). According to Fava et al [28] and Fernández-Habas et al [4] heterogeneous pastures with multiple functional groups and different phenological stages might have contradictory effects on the relationship between pasture quality variables and reflectance, leading to a wide variability of spectral responses.…”

“…This general pattern was represented in simplified form by a polynomial equation in Figures 9-14 with excellent coefficients of determination in the two years of study and in the six experimental fields (0.80-0.95). Since most nitrogen in plant tissue is contained in chlorophyll-protein complexes [22], the well established relationship between leaf chlorophyll pigments and nitrogen (and, therefore, crude protein) helps to explain the behavior of NDVI as robust predictor for spatial distribution of pasture vegetative vigor [23].…”

“…Grazed pastures are recognized as complex and very spatio-temporal dynamic systems [30], composed of a mosaic of different landscape features [23] in close interaction with animals which graze them selectively in time and space [31]. This complexity also demands further research in development of model-data fusion tools [29,32] and present particular challenges for sensor applications [30], preferably integrating the potential complementarity of remote and proximal sensing.…”

Contributions to Satellite-Based Land Cover Classification, Vegetation Quantification and Grassland Monitoring in Central Asian Highlands Using Sentinel-2 and MODIS Data

“…Numerous machine-learning algorithms for modeling and assessing related research questions have been proposed and tested in remote sensing literature. Commonly applied methods for classification or regression include random forest (RF), conventional decision trees, support vector machines (SVM), maximum likelihood classifiers, Mahalanobis distance, artificial neural networks, fuzzy adaptive resonance theory-supervised predictive mapping, K-nearest neighbors, boosting techniques with decision trees, quadratic discriminant analysis, Extreme Gradient Boosting and many more (Grabska et al, 2020;Lapini et al, 2020;Raab et al, 2020;Talukdar et al, 2020). Among respective methods, RF is considered as the most widely used classification algorithm (Phan et al, 2020) and a large number of recent studies state good or better performance of the RF approach compared to other techniques (Verrelst et al, 2019;Diesing 2020;Lapini et al, 2020;Talukdar et al, 2020).…”

Landscape Structure and Seasonality: Effects on Wildlife Species Richness and Occupancy in a Fragmented Dry Forest in Coastal Ecuador

“…Remote sensing has been an important tool for wildlife management, especially for the conservation of large animals over extensive areas [31][32][33]. The recent availability of finer sensors and the development of advanced classification approaches has made the assessment of habitat quality for smaller areas possible [34]. From a conservation planning perspective, grouping spatially explicit information from multiple species is a powerful combination of tools, allowing us to understand species' responses to changes in their landscapes and to accurately design management strategies according to the species realities [31][32][33].…”