DAKBNet: Multiscale Fusion With Dynamic Assembly Kernels and Bilateral Feature Enhancement for Land Use Classification of Remote Sensing Images

Abstract

Land use classification poses significant challenges when applied to remote sensing images. Due to the complex textures, spatial layouts, and scale variations of images, many methods solve these problems by ignoring the bias between low-level and high-level features caused by multiple semantic information associated with each pixel and the effectiveness of multiscale fusion. To tackle the challenges, we propose a novel bidirectional feature enhancement network based on dynamic assembled kernels, which captures both low-level spatial and high-level semantic information of the corrected image through mutual guidance between deep and shallow features. Specifically, we employ high-level semantic features derived from the bilateral structure to compute the semantic deviation of each pixel in the low-level features. Meanwhile, we use the low-level features to resolve redundant information in the high-level components and enhance their global and local context through mutual guidance. On the other hand, we generate kernels by dynamically assembling elementary weight matrices stored in the weight library. The kernel construction is data driven, providing greater flexibility to multiscale features. We have conducted extensive objective and subjective comparative experiments, as well as ablation studies, on the International Society for Photogrammetry and Remote Sensing (ISPRS) Vaihingen, ISPRS Potsdam, and GaoFen Image Dataset. In conclusion, our method has demonstrated notable superiority over other prevailing methods, as evidenced by numerous experimental results.