Recursive Bayesian Classification for Water Mapping

Terrestrial surface water is an important resource that is significantly affected by climate change and urbanization. This results in changes in the magnitude and frequency of flooding, drought, and sometimes waterborne disease outbreaks. Hence, monitoring and prediction of surface water extent is essential for flood hazard/damage assessment and decision-making processes. Satellite-based sensors of varying spatial, spectral and temporal resolution have been used to extract and analyze surface water characteristics. The Landsat and Sentinel satellites are commonly used optical sensors in surface water and land cover classification research. Typical methods for performing water extraction from image data consist of thresholding using either single or multiple image bands. Although band thresholding is simple and computationally inexpensive, recent works considered the use of machine learning (ML) techniques to improve surface water mapping accuracy. However, existing ML strategies used in water mapping face difficulties integrating contextual information, such as spatial, spectral and temporal domains, while taking into account the scarcity of labeled training data that exists in most remote sensing applications. In this work, we propose a method to classify water pixels from Sentinel images recursively capable of real-time operation which integrates a Bayesian statistical formulation with modern ML approaches. The model is able to integrate temporal information to accurately distinguish water from pixels with similar spectral signatures. Experimental results based on images acquired over the Oroville dam show that the proposed method is able to improve the classification accuracy of water pixels of tributaries when compared to one common Modified Normalized Difference Water Index (MNDWI), while using a limited amount of training samples.