Scholarly record

WATER SURFACE MAPPING USING SENTINEL-1 AND SENTINEL-2 IMAGERY: A COMPARATIVE STUDY OF INDEX AND MACHINE LEARNING APPROACHES

Adrian Yordanov

Abstract

This study compares multiple approaches for surface water mapping using multi-sensor satellite data. Open-access imagery from Sentinel-2 (optical) and Sentinel-1 (SAR) was used for three study sites in Bulgaria. The optical dataset consisted of three Level-2A images acquired under near cloud-free conditions, while the radar dataset included three Ground Range Detected images acquired on dates closest to the optical imagery. Surface water extraction was performed using single- and multi-index approaches for the optical data, threshold-based segmentation for the SAR data, and a machine learning approach based on the Random Forest (RF) algorithm. The RF classifier was trained and evaluated using three input configurations: optical-only, SAR-only, and combined optical-SAR datasets, in order to assess differences in classification accuracy and processing efficiency. The models were trained on independent reservoirs and tested on separate validation sites to assess cross-site transferability. Pre-processing was conducted in the Sentinel Application Platform (SNAP), including resampling of optical bands, calculation of spectral indices, radiometric calibration of SAR data, terrain correction, and speckle filtering. The datasets were subsequently collocated and prepared for classification and validation. Validation was performed using high-resolution multispectral imagery from the Beijing-3A (50 cm spatial resolution) and Beijing-3B (30 cm spatial resolution) satellite missions. The study evaluates the performance of single-sensor index-based methods versus multi-sensor machine learning approaches and examines the benefits of optical-SAR data fusion for robust surface water mapping across different water bodies.

Publication details

Back to publication list