FS2CNet: A Dynamic Frequency Selection and Similarity Constraints Change Detection Network Within Cropland for Remote Sensing Images

Cropland non-agriculturalization refers to the conversion of cropland into non-agricultural land that is being used for purposes apart from agricultural production. Using deep learning-based change detection methods for large-area cropland non-agriculturalization detection has great potential. However, most existing deep learning approaches operate primarily in the spatial domain, making them susceptible to fragmented change regions, high-frequency noise (e.g., seasonal crop variations), low-frequency structural ambiguities (gradual changes in topography) in the cropland remote sensing images, and ignores the differences of the change regions in the frequency domain. To address these issues, this paper proposes a dynamic frequency selection and similarity restriction network FS2CNet. Unlike conventional frequency-domain methods that apply fixed filters or uniform frequency processing, our approach introduces a dynamic frequency selection mechanism that adaptively recalibrates features in the frequency domain. Specifically, the model employs learnable filters to decompose features into high-frequency and low-frequency components. It then applies channel-wise attention to dynamically adjust the contributions of different frequency bands—suppressing high-frequency noise (e.g., crop texture disturbances) while enhancing structurally salient low-frequency information (e.g., buildings and roads). Furthermore, we optimize the feature reconstruction process by using similarity constraints on the features across the hierarchical levels that enhances the detection consistency of fragmented regions. Experiments on three non-agriculturalization change detection datasets, CLCD, PX-CLCD and LuojiaSET-CLCD, show that FS2CNet outperforms the current state-of-the-art methods.