HRRF: a hierarchical recursive reasoning framework for high-resolution remote sensing semantic segmentation

Semantic segmentation of high-resolution remote sensing imagery is impeded by a trio of persistent challenges: the obscuration of small-scale objects, the fragmentation of linear geofeatures, and high-level topological inconsistencies. Existing methods often struggle to address these intertwined issues simultaneously, lacking a holistic reasoning process. To overcome these limitations, this paper proposes the Hierarchical Recursive Reasoning Framework (HRRF), a multi-stage deep learning system designed for comprehensive scene interpretation. First, a Dual-stream Context-Aware Encoder robustly captures multi-scale global and local features. Second, a Bi-directional Recursive Refinement Decoder (BRRD) employs a novel Recursive Flow-Modulated Inter-level Aggregation (RFMIA) module with recursive feedback to enhance detail and context. Third, a Graph-based Topological Consistency Correction (GTC) stage enforces high-level spatial-structural priors. Evaluations on the challenging ISPRS Potsdam, Vaihingen, and LoveDA datasets demonstrate that HRRF achieves state-of-the-art or highly competitive performance.