National Water Reservoirs Monitoring through Time Series Remote Sensing

Abstract

Efficient management of water resources plays a critical role in countries characterized by high aridity levels and vulnerability to floods, as seen in the case of Pakistan. However, challenges arise, largely stemming from the scarcity of resources required for effective water management. Fortunately, Remote Sensing field has emerged as a significant player, providing time series data, and allowing the mapping of surface water bodies in an automated and efficient manner. Therefore, this study predominantly focuses on curating and leveraging time series RGB and 12-band Sentinel-L2A imagery to carry out the segmentation of surface water bodies. The study presents a methodology for gathering and preparing multi-temporal RGB and multispectral (12- band) Sentinel-L2A imagery data of surface water bodies in Pakistan. Additionally, accurate ground truth water masks were generated using the manual annotation tool, LabelMe. Furthermore, we have analyzed the performance of variants of a state-of-the-art deep learning segmentation model to segment the multi-temporal surface water bodies from both true-color RGB and 12-band Sentinel-2 surface water bodies images. The variants of the segmentation model are essentially used to yield insights into the strengths and limitations of different spectral information configurations. This study also provides different training and validating strategies, enhancing the deep learning model’s robustness and its ability to generalize effectively on our collected dataset. Experiments compared RGB and 12-band image results, with RGB images outperforming due to the uniform resolution of their bands. The study’s applications include calculating water surface area, aiding flood and drought forecasting, and supporting strategic decision-making for water resource management and environmental monitoring.