Land-cover classification with remote sensing images based on low-rank fusion of multimodal features

Abstract

Multimodal remote-sensing land classification aims to achieve more accurate and comprehensive extraction of land features in remote sensing images by integrating feature information from multiple remote sensing data sources. This article proposes a unified multimodal remote sensing feature classification network, which includes: a weight sharing backbone network responsible for extracting preliminary feature representations from the input data of each modality; The multimodal feature low rank fusion module performs cross modal transmission on high-level semantic features to enhance semantic interaction between modalities; The upsampling operation is responsible for restoring the fused feature map to the same resolution as the input image. This algorithm achieved 91.23% OA and 83.28% mIoU in remote sensing land feature classification tasks, effectively alleviating the problems of insufficient accuracy and insufficient utilization of multimodal information faced by traditional single modal remote sensing classification methods through feature low rank fusion technology, thereby significantly improving the performance of land feature classification.