Spatial-Frequency Domain Learning for Road Extraction from Remote Sensing Images

Extraction of road from remote sensing (RS) images confronts a dual challenge: the inhomogeneous intensity and inconsistent contrast. Historically, conventional approaches predominantly relied on spatial-domain convolutional neural networks (CNNs), which were constrained by their local receptive fields, limiting their ability to effectively capture extensive and intricate road features. In addition, self-attention mechanisms with global awareness have a high computational load, and recent image frequency-domain representation learning makes up for this shortcoming. Based on this, the paper presents a road segmentation network, specifically designed using a U-Net architecture, called lightweight spatial-frequency domain learning U-Net (LSFDLU-Net). In this network, spatial-frequency domain learning module is used as the basic component. Firstly, a spatial-channel Fourier neural operator (SC-FNO) based on frequency-domain learning is proposed. SC-FNO employs gate-controlled filtering to coarsely process the frequency-domain representation of road features in RS images, globally suppressing interference information related to the road. Subsequently, a multi-scale feature fusion module based on spatial-domain learning is devised, consisting of multi-scale depth-wise separable convolutions (MSDSC), feature separation (FS), and attention-guided fusion (AGF). It finely processes RS images by separating the original, high-frequency, and low-frequency spatial domain representations, and employs MSDSC and AGF to effectively cap-ture slender road features. The network significantly improved Fl scores, accuracy, recall, and mean Intersection over Union (mIoU) for the Massachusetts and DeepGlobe public road datasets while maintaining a smaller number of parameters and computations compared to related methods.