High-spatiotemporal-resolution dynamic water monitoring using LightGBM model and Sentinel-2 MSI data

Abstract

Lake shrinkage and water scarcity are frequent problems in arid and semiarid regions; monitoring the variations of surface water using remote sensing images is useful for disaster prevention and water resource management. Here, we proposed models using Sentinel-2 images based on the light gradient boosting machine (LightGBM) to quantify the monthly surface water dynamics in the middle farming-pastoral ecotone of the Northern China (M-FPENC) region, which was facing severe water security challenges from 2016 to 2021, at a 10-m resolution. The results show that the proposed models perform very well, with average values over 99.9% derived for four classification metrics; LightGBM is time-efficient and robust in different scenarios for surface water dynamic monitoring. The maps produced by these models could capture the details of surface water accurately in the M-FPENC region; the surface water area of the M-FPENC region showed a clear seasonal pattern, showing the largest average extent in June (856.9 km2) and the smallest average extent in November (486.6 km2); the annual average maximum water area was 1030.9 km2, of which 630.9 km2 was seasonal and 400 km2 was permanent. The proposed product can provide decision support for local water resources planning.

Keywords

• Sentinel-2
• Surface water dynamics
• Farming-pastoral ecotone of Northern China (FPENC)
• Machine learning
• LightGBM