MMSRC: A Multidirection Multiscale Spectral–Spatial Residual Network for Hyperspectral Multiclass Change Detection
Hongmei Ge, Yongsheng Tang, Zuolin Bi, Tianming Zhan, Yang Xu, Aibo Song
Abstract
Recently, deep convolutional neural network (CNN) hyperspectral change detection methods have achieved significant improvement. However, most CNN hyperspectral change detection methods do not make full use of spectral-spatial feature information. In this paper, we propose a novel multi-direction and multi-scale spectral-spatial residual network for hyperspectral multi-class change detection. Specifically, multi-scale structure and a multi-direction mechanism are introduced to investigate feature variation of hyperspectral images and improve the accuracy of hyperspectral change detection. Experiments on multiple hyperspectral datasets show that the proposed method achieves improved performance in comparison with other advanced hyperspectral multi-class change detection methods.