CPFR: category-phenology decoupling and frequency-aware resampling model for satellite image time series semantic segmentation

Precise delineation of crop spatial distributions using satellite image time series (SITS) is crucial for optimizing agricultural monitoring and management. However, constrained by the inherent spatial sampling limitations of imaging sensors, spatial aggregation in deep neural networks frequently induces heterogeneous spectral mixing and high-frequency detail distortion. This poses significant challenges for accurately discriminating crop categories and identifying parcel boundaries in complex agricultural scenarios. To address these issues, we propose a novel Category-Phenology Decoupling and Frequency-Aware Resampling Network (CPFR). We design a set of learnable category queries to explicitly decouple category-specific crop phenological patterns from the mixed temporal signals at the bottleneck layer. Furthermore, we extract temporal consensus high-frequency information to guide the adaptive aggregation of sampling points toward high-frequency regions, thereby mitigating spatial aliasing during the downsampling process. Correspondingly, we restore high-frequency information via feature demodulation and incorporate masked attention to facilitate effective information interaction within foreground regions. Extensive experiments conducted on the PASTIS, JM, and CF datasets demonstrate that our method achieves notable performance. The proposed model further exhibits robust adaptability under more challenging spatial resolution degradation scenarios.