RSFF: a reconciliation-sparse fusion framework for semantic segmentation of rocky desertification land in multimodal time series

Accurate identification of rocky desertification land (RDL) is crucial for ecological environment monitoring and management in karst regions. However, RDL extraction is often hindered by land-cover confusion, seasonal variability, and the fragmented, small, and irregular spatial distribution, when existing methods insufficiently exploit multi-temporal multimodal information. To address this issue, we construct the first dual-modal RDL semantic segmentation dataset with time series (RD dataset) and propose a reconciliation-sparse fusion framework (RSFF) for multi-temporal multimodal data. The framework adopts a dual-branch encoding architecture with multilevel spatio-temporal interaction to separately extract spatial and temporal features from optical and Synthetic Aperture Radar (SAR) imageries. Through the designed Convolutional Attention Reconciliation Module (CARM), multidimensional enhanced features are reconciled, and the Sparse Cross Fusion Module (SCFM) is employed to improve feature discriminability while reducing computational complexity compared with conventional dense attention mechanisms, enabling deep complementarity of cross-modal information. Experimental results on the RD dataset demonstrate that RSFF can effectively distinguish RDL from similar land-cover types, such as bare land and cropland, and significantly enhance RDL boundary delineation while maintaining high classification consistency. The mean intersection over union (mIoU) and overall accuracy (OA) reach 82.94% and 94.28%, respectively, outperforming other state-of-the-art (SOTA) methods. In addition, experiments on the public PASTIS-R dataset further verify the robustness of RSFF under different landscape characteristics and small, fragmented parcel distributions. The RD dataset and code are available at https://github.com/laopiao-lp/RSFF.