Dynamic mapping of subarctic surface water by fusion of microwave and optical satellite data using conditional adversarial networks

Abstract

Surface water monitoring with fine spatiotemporal resolution in the subarctic is important for understanding the impact of climate change upon hydrological cycles in the region. This study provides dynamic water mapping with daily frequency and a moderate (500 m) resolution over a heterogeneous thermokarst landscape in eastern Siberia. A combination of random forest and conditional generative adversarial networks (pix2pix) machine learning (ML) methods were applied to data fusion between the Moderate Resolution Imaging Spectroradiometer (MODIS) and the Advanced Microwave Scanning Radiometer 2, with the addition of ancillary hydrometeorological information. The results show that our algorithm successfully filled in observational gaps in the MODIS data caused by cloud interference, thereby improving MODIS data availability from 30.3% to almost 100%. The water fraction estimated by our algorithm was consistent with that derived from the reference MODIS data (relative mean bias: −2.43%; relative root mean squared error: 14.7%), and effectively rendered the seasonality and heterogeneous distribution of the Lena River and the thermokarst lakes. Practical knowledge of the application of ML to surface water monitoring also resulted from the preliminary experiments involving the random forest method, including timing of the water-index thresholding and selection of the input features for ML training.