Remote sensing monitoring method for flooded farmland based on domestic GF-3 radar images

Against the backdrop of global warming, increasingly frequent floods become a primary agricultural disaster that causes reduced crop production in China. Radar remote sensing technology, possessing all-weather earth observation capabilities, serves as a critical means for rapid monitoring of regional flood information. With the advancement in artificial intelligence, machine learning methods have been extensively applied in the remote sensing-based monitoring of floods. Despite the high accuracy of their algorithms, their training processes often entail extensive field investigations or numerous samples for remote sensing image interpretation. This study aims to overcome sample labeling limitations and improve regional flood monitoring accuracy. Based on the catastrophic flood that occurred in northern Henan on July 20, 2021, this study constructed a flooded crop monitoring method based on the weakly supervised Gaussian mixture model (GMM) using domestic GF-3 HH-HV radar images. Then this method was applied to extract the flooding range of farmland in some areas of northern Henan. Compared to four typical machine learning methods, i.e., random forest (RF), support vector machine (SVM), K-nearest neighbor classification, and parallelepiped classification, the weakly supervised GMM in this study enjoyed the highest accuracy, with overall precision of 0.95 and a Kappa coefficient of 0.90. This study holds great significance for enhancing the accuracy and universality of regional crop flooding monitoring based on remote sensing technology and synthetic aperture radars (SARs).