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

Abstract

The peripheral setting of cold drylands in Asian mountains makes remote sensing tools essential for respective monitoring. However, low vegetation cover and a lack of meteorological stations lead to uncertainties in vegetation modeling, and obstruct uncovering of driving degradation factors. We therefore analyzed the importance of promising variables, including soil-adjusted indices and high-resolution snow metrics, for vegetation quantification and classification in Afghanistan’s Wakhan region using Sentinel-2 and field data with a random forest algorithm. To increase insights on remotely derived climate proxies, we incorporated a temporal correlation analysis of MODIS snow data (NDSI) compared to field measured vegetation and MODIS-NDVI anomalies. Repeated spatial cross-validation showed good performance of the classification (80–81% overall accuracy) and foliar vegetation model (R2 0.77–0.8, RMSE 11.23–12.85). Omitting the spatial cross-validation approach led to a positive evaluation bias of 0.1 in the overall accuracy of the classification and 25% in RMSE of the cover models, demonstrating that studies not considering the spatial structure of environmental data must be treated with caution. The 500-repeated Boruta-algorithm highlighted MSACRI, MSAVI, NDVI and the short-wave infrared Band-12 as the most important variables. This indicates that, complementary to traditional indices, soil-adjusted variables and the short-wave infrared region are essential for vegetation modeling in cold grasslands. Snow variables also showed high importance but they did not improve the overall performance of the models. Single-variable models, which were restricted to areas with very low vegetation cover (<20%), resulted in poor performance of NDVI for cover prediction and better performance of snow variables. Our temporal analysis provides evidence that snow variables are important climate proxies by showing highly significant correlations of spring snow data with MODIS-NDVI during 2001–2020 (Pearson’s r 0.68) and field measured vegetation during 2006, 2007, 2016 and 2018 (R 0.3). Strong spatial differences were visible with higher correlations in alpine grasslands (MODIS NDVI: 0.72, field data: 0.74) compared to other regions and lowest correlations in riparian grasslands. We thereby show new monitoring approaches to grassland dynamics that enable the development of sustainable management strategies, and the mitigation of threats affecting cold grasslands of Central Asia.

Assessing the accuracy of spectral indices obtained from Sentinel images using field research to estimate land degradation

“…The examination of the Root Mean Square Error (RMSE) between the models obtained from the Albedo index and each of the spectral indices derived from remote sensing and field data revealed that the intensity of degradation derived from the MSAVI-Albedo index has the highest proximity to the field data with an RMSE of 0.65. This indicates a closer agreement than the other indices, NDVI and SAVI, which is consistent with Zandler et al’s study in 2022, which demonstrated that the NDVI index is not suitable for dryland with sparse vegetation covers [ 40 ]. Furthermore, the intensity of degradation obtained from the NDVI-TGSI model with an RMSE of 0.66 is higher than other models, indicating that the degradation intensity map derived from these two indices has higher accuracy compared to vegetation cover indices with Albedo [ 41 ].…”

Multisite and Multitemporal Grassland Yield Estimation Using UAV-Borne Hyperspectral Data

“…However, that method delivered unsatisfactory results in a study on estimation of the quality parameter starch content of the grassland species red clover from hyperspectral data [44]. Another promising method for variable reduction in ML approaches previously employed on grassland data is the Boruta algorithm [45]. Due to its high computational intensity [46], however, Boruta was ruled out for the present study.…”

Cooling Potential of Urban Tree Species during Extreme Heat and Drought: A Thermal Remote Sensing Assessment

“…To also include a non-parametric machine learning method in our analysis, we assessed the importance of the different variables on surface temperatures using a 500-repeated Boruta algorithm, as this is considered a powerful method for quantifying influential predictors [68,69]. In this approach, the dependent variable is modeled using random forest regression, whereby shadow variables are created by randomly shuffling original predictors, which are then compared to the performance of original variables for a quantitative importance assessment [70].…”

“…We used T diff as the dependent variable, and tree species, NDVI, and imperviousness within a 91 m radius, as outlined in Ziter et al [56], as independent variables to assess if all variables are considered important in a collective model. The performance of the model was assessed using 100-repeated, 10-fold cross-validation, which is considered to yield an unbiased error estimate [68]. Formulas for performance metrics are outlined in Zandler et al [72].…”