RSReason: Remote Sensing Reasoning Segmentation via Rotation-aware Large Vision-Language Model

Remote Sensing Referring Expression Segmentation (RSRES) has laid the groundwork for mapping natural language instructions to pixel-level segmentation in remote sensing imagery, where objects frequently occur at diverse scales, dense spatial distributions, and arbitrary orientations. However, existing approaches predominantly rely on explicit, category-driven queries, thereby limiting their ability to reason over spatial relationships, geometric properties, and both commonsense and domain-specific knowledge. To address these challenges, we introduce Remote Sensing Reasoning Segmentation (RSReasonSeg), a new task requiring models to perform segmentation and interpretation based on implicit, reasoning-centric, commonsense-rich language instructions. To support this, we present the RSReason benchmark, a large-scale dataset comprising 5,138 samples with reasoning-centric queries, expert-annotated masks, and explanatory answers. We further propose the RSReason-Net, a large vision-language model equipped with a Rotation Awareness Optimization (RAO) module. RAO models irregular masks as 2D Gaussians to align predictions with geometric and orientational cues derived from language. Extensive experiments demonstrate that RSReasonNet achieves superior performance over the classical LISA baseline on RSReasonSeg. Moreover, RAO consistently improves the recent state-of-the-art RSRefSeg 2 model, establishing a new state-of-the-art result. On RRSIS-D, RAO enhances RSRefSeg 2 across all evaluation metrics on both the validation and test sets, without introducing additional inference-time overhead.