DAGMN: A Geometric and Semantic Detail Aware Gaussian Mixture Deep Neural Network for Land-Use/Land-Cover Classification of High-resolution Remote Sensing Imagery

The deep learning based semantic segmentation of high-resolution remote sensing imagery plays a crucial role in the field of land-cover monitoring. With the gradual improvement of the spatial resolution of remote sensing images, the geometric characteristics and category distribution of the targets in the images are becoming more complex, and more direct modeling methods are needed to acquire precise segmentation results. In this paper, a geometric and semantic detail aware Gaussian mixture deep neural network (DAGMN) is proposed. The encoder of DAGMN consists of several anisotropic geospatial refinement blocks (AGRBs), which contain a hierarchical strip convolution attention module (HSCAM) and a channel enhancement attention module (CEAM). In the AGRB, the HSCAM can enhance objects' geometrical directional information based on strip convolution kernels with different directions and scales, and the channel information interaction is enhanced by the CEAM in each block. To better model the precise distribution of the different categories, a latent semantic distribution modeling (LSDM) classifier is proposed, which employs the generative Gaussian mixture model (GMM) to model the underlying data distribution instead of the soft-max classifier. A hybrid training strategy to jointly optimize the parameters in the network and the GMM is also introduced. The experiments on the ISPRS Vaihingen, ISPRS Potsdam, and LoveDA datasets and a newly developed dataset called NingXia-Seg show the potential of the proposed method, compared with the related semantic segmentation methods. Furthermore, we conducted city-scale semantic segmentation on the area of Yinchuan, China, as a case study in comprehensive land surface monitoring.