Multiscale and Multidirection Feature Extraction Network for Hyperspectral and LiDAR Classification

Abstract

Deep learning (DL) plays an increasingly important role in Earth observation by multisource remote sensing. However, the current DL-based methods do not make a fully use of the complementary information among multisource remote sensing data, such as hyperspectral image and light detection and ranging data, and lack the consideration of multiscale, directional, and fine-grained features. To address these issues, a multiscale and multidirection feature extraction network is proposed in this article. Specifically, the multiscale spatial feature (MSSpaF) module is designed to extract the MSSpaFs, and then, these features are fused by feature concatenation operation. In addition, the multidirection spatial feature module is designed to further extract multidirection and frequency information, employing cross-layer connection and multiscale feature fusion strategy to improve the fineness of the proposed network. Moreover, the spectral feature module is employed to provide detailed spectral information for enhancing the expression ability of multiscale features. Experimental results on three different datasets demonstrate the superior classification performance of the proposed framework.