CS-M 2 Former: unified multi-task framework for remote sensing image segmentation based on cross-scale feature fusion

This is the authors' abstract. We don't add key points for this paper.

High-resolution optical remote sensing image segmentation is crucial for applications, including urban planning, land-use classification, and environmental monitoring. However, delineating regions and objects remains challenging due to scale variations, spatial patterns, and intra-class variability across landscapes. Unified frameworks are needed to jointly address semantic and instance segmentation tasks in remote sensing imagery. To address these issues, this letter presents a cross-scale segmentation framework named Cross-Scale MambaMask2Former (CS-M2Former). Specifically, we build MambaVision++ by upgrading the existing MambaVision backbone with shallow convolutional-attention (ConvAttn) blocks, while preserving its deep Mamba–Transformer mixers for global modelling. We further propose a Cross-Scale Bidirectional Feature Pyramid Network (CSB-FPN) for bidirectional cross-scale feature fusion and DualDySample, a DySample-inspired dual-branch adaptive upsampling module, to reduce spatial information loss during long-range feature transmission. On the Forest-Full instance segmentation benchmark, CS-M 2 Former achieves 77.2% AP, exceeding the strongest Mask2Former baseline by 2.0% points; on the LoveDA semantic segmentation benchmark, it obtains 53.38% mIoU. These results demonstrate highly competitive performance among the representative segmentation models evaluated under the unified instance and semantic segmentation setting.