Detection of Grassland Degradation In Azerbaijan By Combining Multi-Decadal NDVI Time Series And Fractional Cover Estimates Based On DESIS Data

Abstract

Degradation is one of the most pressing global environmental problems and is projected to worsen due to climate change and land use pressure. Grassland ecosystems used for pasturing are especially prone to degradation. In the Caucasus region, livestock farming is an important part of the agricultural sector and subsistence farming is commonplace, hence threats to pastures can significantly impact livelihoods. The grassland areas of Azerbaijan are under heavy anthropogenic pressure, leading to vegetation cover loss and erosion, especially on community pasture land. Degradation is generally assessed in remote sensing by quantifying changes in vegetation indices (VIs). This is a challenging task, as information from long time series is needed to detect trends, and frequent observations are needed to distinguish degradation from phenological variability. For such time series, only multispectral data is available. However, especially in regions where vegetation is sparse, information on the fractions of ground cover such as photoactive vegetation (PV), non-photoactive vegetation (NPV) or soil is important, as soil reflectance affects VIs. Hyperspectral data are particularly valuable in this regard as they have the spectral resolution required to distinguish soil, vital and degraded vegetation. In this study we therefore investigate the potential of combining multispectral time series with hyperspectral data. In a first step, a nationwide land cover map is generated. During two field campaigns in August and October 2018, 296 plots in grassland, cropland and shrubland were visited, for which land cover, coverage and erosion intensity were recorded. In addition, samples of urban areas, soil, water and forests were collected from Google Earth Engine (GEE) imagery. 70 spectral-temporal metrics of the Sentinel-2 imagery of 2018 were used as input features together with the field data in a random forest classifier. Land cover is modeled with an overall accuracy of 83 % (Asam et al. 2019, ESA LPS). A Normalized Difference Vegetation Index (NDVI) time series is used to identify grassland degradation on a national scale. Acquisitions from the Landsat Missions (TM, ETM+, OLI; 1984-2020) are harmonized and each image is masked using fmask on the GEE platform. For each year, median NDVI of the grassland areas are generated and trends are calculated using the Sen's slope and the Mann-Kendall test. For 2019 – 2021, 9 DESIS acquisitions are available with cloud coverage < 25% and recorded with sun angle < 40°, covering parts of the western lowlands and Lesser Caucasus of Azerbaijan. For each scene, fractional cover was calculated using the “fCover” processor. It derives pure material signatures using the Spatial-Spectral Endmember Extraction (Rogge et al. 2012, JSTARS, 5(1)), which are then classified into the classes PV, NPV and soil, using a pre-trained random forest. fCovers are then calculated using a Multiple Endmember Spectral Mixture Analysis (Bachmann et al. 2009, 6th EARSeL-SIG-IS) with each pixel treated as a linear combination of each spectral class. Originally developed for hyperspectral sensors covering the full VNIR-SWIR range, this method was successfully applied also to VNIR-only DESIS data (Marshall et al., 2021, 1st DESIS User Workshop). First results indicate that 5.4 % of Azerbaijan’s grasslands show a significant (p < 0.05) negative NDVI trend, pointing to potential degradation hotspots. PV could be derived from DESIS with a mean absolute error of 8,94 %. Next, areas showing a degradation trend are intersected with PV and NPV fractions and analyzed regarding their statistical relationship. First results show that pixels with a high PV coverage are less degraded. In addition, effects of topography and degradation time scales will be analyzed. Using this approach, a country-wide multi-decadal assessment of vegetation changes can be enhanced by adding canopy structure information from hyperspectral DESIS data.