A Unified Frequency and Spatial Domain Network for Intelligent Remote Sensing Image Analysis

The semantic segmentation of remote sensing images seeks to provide semantic labels for every pixel in remote sensing images to ensure accurate and fine-grained interpretations of objects on the ground. However, it still faces some challenges, such as interference from noise, limited inter- class difference, uneven scales of objects, and complicated correlations between spatial and frequency domains. The existing CNN networks have limitations in extracting long- range contextual information, and most of the existing transformer networks are mainly focused on feature modeling in the spatial domain, ignoring frequency domain features. Therefore, in this paper, a unified frequency and spatial domain network for intelligent remote sensing image analysis is presented to solve these problems. The key idea of this paper is to use a preprocessing-assisted frequency and spatial domain fusion strategy, in which frequency domain features are learned by using Fourier decomposition to emphasize high-frequency details, and spatial domain features are used to preserve global semantic structures. Moreover, a multi-scale context modeling mechanism is introduced to adapt to scale changes, and channel-spatial collaborative attention is used to optimize feature representations in different dimensions to improve the recognition accuracy of small targets. Significant experiments performed on benchmark remote sensing datasets show that the suggested method surpasses existing state-of-the-art approaches in terms of performance metrics. Ablation experiments also confirm the effectiveness of frequency-spatial fusion and multi-scale modeling in disambiguating inter-class confusion and improving semantic boundary accuracy.