Automated estimation of daily surface water fraction from MODIS and Landsat images using Gaussian process regression

Abstract

ABSTRACT Satellite remote sensing has been widely used to monitor surface water, but its application in observing rapid inundation changes remains challenging. Observations relying on only one sensor could hardly achieve both high temporal and high spatial resolutions. High spatial resolution images are not frequent enough to capture the fast-changing inundation, while high temporal resolution images do not provide sufficient detail on spatial variations. To address this resolution compromise for rapid response, here we propose a new automated method to estimate daily sub-pixel water cover, the surface water fraction, from Moderate Resolution Imaging Spectroradiometer (MODIS) and Landsat images. First, water indices are derived from MODIS and Landsat images, and the Landsat derived water indices are used to generate a binary water map. We then aggregate the Landsat water map to the coarser MODIS resolution, making a surface water fraction map. Second, the relationship between the MODIS derived water indices and the Landsat derived surface water fraction is fitted using a Gaussian Process Regression (GPR) model. Lastly, the fitted GPR model is applied to new MODIS imagery to estimate surface water fraction. The experiment results showed that the proposed method could provide water fraction with Root Mean Squared Errors of less than 7%. By mapping fractional water cover automatically and regularly, the proposed method could facilitate daily emergency responses and long-term analyses.