A Neural Network Classification Framework for Monthly and High Spatial Resolution Surface Water Mapping in the Qinghai–Tibet Plateau from Landsat Observations

Abstract The Qinghai–Tibet Plateau, known as the Asian Water Tower, has a significant area of water bodies that provide a wide range of valuable ecosystem services. In the context of climate change, the formation condition of surface water and water extent is changing fast. Thus, there is a critical need for monthly detection algorithms at high spatial resolution (∼30 m) with good accuracy. Multiple sensors’ observations are available, but producing reliable long time series surface water mapping at a subannual temporal frequency still remains a challenge, mainly due to data limitations. In this study, we proposed a neural-network-based monthly surface water classification framework relying on Landsat 5/7/8 images in 2000–20 and topographic indices, and retrieved monthly water mask for the year 2020. The surface water was mainly distributed in the central and western parts of the plateau and the maximum area of permanent surface water (water frequency > 60%) was 26.66 × 103 km2 in 2020. The overall, producer, and user accuracies of our surface water map were 0.96, 0.94, and 0.98, respectively, and the kappa coefficient reached 0.90, demonstrating a better performance than existing products [i.e., Joint Research Centre (JRC) Monthly Water History with overall accuracy 0.94, producer accuracy 0.89, user accuracy 0.99, and kappa coefficient 0.89]. Our framework efficiently solved the problem of missing data in Landsat images referring to the JRC and a priori information and performed well in dealing with ice/snow cover issues. We showed that higher uncertainties exist on wetlands and recommended exploring relationships between water and wetlands in the future. Significance Statement In this paper, we present a new methodology to estimate surface water and its intra-annual changes using Landsat data. Missing data and retrieval errors in the winter are major issues in the existing products (i.e., JRC dataset). This motivated us to develop a new machine learning algorithm to better improve the retrieval scheme. We show that our approach, based on a neural network classifier, delivers a significant improvement compared to the previous estimates. As shown in the literature, JRC data can hardly be used at the monthly level, whereas our retrieval appears to be exploitable at the monthly scale. This is essential to understand the trend in surface water, one of the key elements of the water cycle.