An improved surface water extraction method by integrating multi-type priori information from remote sensing

Surface water mapping based on historical, neighbourhood, and other priori information has shown improved accuracy. However, the accuracy can be compromised due to the lack of consideration for water dynamics in the proximity period and the limited utilization of quantitative methods for integrating multiple types of priori information. In this study, an unsupervised surface water extraction method that integrates historical, proximity, and neighbourhood priori information from remote sensing is proposed to enhance water extraction accuracy. The experiments were conducted in Poyang Lake, a region characterized by active hydrological phenomena. Coarsely extracted water extents using the OTSU method from all available Sentinel-1/2 images within a one-month period priori to the current moment were utilized as proximity priori information to estimate water probability (WP). The optimized WP was then obtained by combining the estimated probabilities from historical, proximity, and neighbourhood priori information using the Bayesian Model Averaging (BMA) method. The experimental results demonstrate that the proposed method outperforms traditional unsupervised water body extraction methods such as K-Means, IsoData, and OTSU in terms of accuracy. Moreover, the integration of historical, proximity, and neighbourhood priori information based on the BMA method results in higher accuracy compared to using each priori information separately. Specifically, the incorporation of proximity priori information significantly enhances the estimation accuracy of WP. In summary, the proposed water extraction method offers high accuracy, automation, and strong robustness, making it applicable to areas characterized by active and complex hydrological changes.