Delineating fragmented grassland patches in the tropical region using multi-seasonal synthetic aperture radar (SAR) and optical satellite images

Abstract

ABSTRACT Globally, grasslands are declining and are in highly degraded conditions. In South Asia, grasslands are neglected and treated as wastelands. They remain unprotected, highly fragmented, and poorly understood, which has led to a loss of unique biodiversity and livelihoods. Mapping grasslands accurately is a challenge and current maps based on optical remote sensing often over- or underestimate grasslands in South Asia due to the prevalent complex landscape matrix, small patch sizes, and obscuring monsoonal clouds. Synthetic Aperture Radar (SAR), fused with moderate spatial resolution optical data, has been used to delineate grasslands, but the high-resolution, freely available European Space Agency (ESA)’s Sentinel-1 (SAR) and Sentinel-2 (Optical) provide an opportunity to map small and fragmented patches. Further, high-resolution imageries require high computing power, which is often limited to stand-alone machines. Here, we demonstrate that using cloud computing and optimal use of multi-seasonal imagery, one can obtain a highly accurate land-cover land use (LCLU) classification for a complex habitat matrix. We performed LCLU classification using the Support Vector Machine (SVM) algorithm of Sentinel-1, Sentinel-2, and Sentinel-1 and Sentinel-2 (Combined) in Google Earth Engine (GEE). GEE is a freely accessible cloud computing platform. We compared the accuracy of grassland classes between a) Sentinel-1 seasonal (pre-, during, and post-monsoon), b) Sentinel-2 post-monsoon, and c) combined Sentinel-1 and Sentinel-2. We tested this method at two sites in a highly fragmented habitat matrix in arid and semi-arid areas of Western and Southern India. The classification results have shown that the overall accuracy for the combined image classification was higher than only Sentinel-2 or Sentinel-1 alone for both sites. Grassland class accuracy was also consistent with combined image classification across the sites. Our results identified newer grassland areas that coarse land use management maps used by the government did not. Since the computation was conducted in GEE, a regular laptop was sufficient for the user and processing was completed rapidly. We, therefore, suggest that this approach of using cloud computing and optimal use of resource-hungry (computation and storage) high-resolution Sentinel-1 and Sentinel-2 data can be used to identify major LCLU classes and patchy grasslands in the arid and semi-arid regions of India and has the potential to map at a continent level.