MFAGNet: multi-scale frequency attention gating network for land cover classification

Abstract

ABSTRACT As a classic problem of high-resolution remote sensing image tasks, land cover classification has many challenges, for instance, variable scales and low-edge discrimination. Traditional spatial semantic segmentation methods based on single-scale feature extraction are challenging to model multi-scale features. They cannot effectively fit high-resolution remote sensing images with variable scales and are prone to misclassification issues. On the other hand, spatial multi-scale semantic segmentation methods lack edge constraints, which can easily lead to edge segmentation ambiguity issues. We propose a segmentation network based on a multi-scale frequency-domain attention gating mechanism (MFAGNet) to solve the above problems. Specifically, we obtain multi-scale features through multi-scale input and design the discrete cosine transform channel attention module (DCTCAM). In DCTCAM, we extract global low-frequency features and effectively alleviate misclassification caused by scale changes. The edge features of the image are obtained through the high-frequency feature extraction module (HFEM) by using edge detection operators to enhance the model’s ability to recognize edges. Additionally, no further pooling operations are employed except for a single max pooling layer in the backbone network. In this paper, we design a new gating mechanism to dynamically control the extraction ratio of various frequency features so that it can adapt to input images of different scales. This paper conducts comparative experiments on three high-resolution land cover datasets: GID, Vaihingen, and Potsdam. The new algorithm proposed compares with classic semantic segmentation models such as DANet, PSPNet, SegNet, UNet, DeepLabv3+, FcaNet, BiSeNet, AFN, and FastFCN. The experimental results show that MFAGNet is superior to the comparison models. With no significant increase in computing overhead, MFAGNet obtains MIoU of 69.05%, 73.40%, and 79.42% on the standard data sets GID, Vaihingen, and Potsdam, respectively.