Historical information fusion of dense multi-source satellite image time series for flood extent mapping

Abstract

Existing lightweight flood extent mapping (FEM) methods, leveraging satellite imagery, offer respite from the burden of extensive data processing. However, these methods often grapple with suboptimal accuracy of the FEM results. To tackle this challenge, this paper introduces an innovative lightweight historical information fusion framework aimed at enhancing FEM results through the utilization of dense multi-source satellite image time series (MSSITS). Firstly, we propose an ultra-lightweight convolutional neural network comprising fundamental network modules, such as convolution layers, batch normalization layers, and parametric rectified linear unit layers. This network swiftly extracts initial FEM results from dense MSSITS. Secondly, we devise a historical information fusion strategy to refine the initial FEM results of the current image by fusing historical FEM results exhibiting high temporal correlation. During the fusion stage, we harness the complementary attributes of dense MSSITS. Specifically, the fusion of diverse satellite sensor data aids in mitigating false alarms in the initial FEM results, thereby facilitating the acquisition of refined FEM results. This approach effectively diminishes the impact of noise interferences, including terrain shadows and surging flood waves in synthetic aperture radar images, as well as cloud occlusions and cloud shadows in optical images. Experimental findings, based on the dense MSSITS of four authentic flood events, demonstrate that our proposed method yields relatively superior FEM performance while maintaining a low FEM complexity in comparison with commonly used semantic segmentation methods.