An optimized machine learning approach to water pollution variation monitoring with time-series Landsat images

Non-point source (NPS) pollution has greatly threatened socio-economic development and human health due to water environment degradation. It is very important to quantitatively analyze spatio-temporal variation rules of NPS pollution sources surrounding drinking water source area (DWSA) and their impact on the water environment with time-series satellite images. In this paper, we study a systematic remote sensing monitoring method on DWSA of upper Huangpu River, Shanghai. Firstly, an optimized Extreme Learning Machine (ELM) classification algorithm, namely Mixed Kernel ELM with Particle Swarm Optimization (PSO-MK-ELM) was constructed. Based on the PSO-MK-ELM, four NPS pollution sources- farmland, building land, woodland, and water were identified accurately and efficiently. Then their corresponding spatiotemporal analysis was performed with 30 years (1989–2019) Landsat images. On the basis of NPS pollution source area and census data from 1989 to 2017, the principal pollutants discharged into DWSA were also calculated with the common Export Coefficient Model (ECM). Finally, the contributions of the spatial and temporal changes of NPS pollution sources on pollutant emissions were analyzed. The result indicates the PSO-MK-ELM has an advantage of efficiency and accuracy in NPS pollution source extraction and our results are expected to provide a scientific basis and data support for NPS pollution control and DWSA protection for better practices for environmental management in megacities worldwide.