GeoSeC: Geometric-Semantic Collaborative Learning for Weakly Supervised Remote Sensing Image Segmentation.

Abstract

Despite the economic advantages of weakly supervised semantic segmentation (WSSS) in remote sensing imagery (RSI), existing VLM- and VFM-based approaches still struggle with domain-specific semantic ambiguities and geometric discontinuities. To address these challenges, we propose GeoSeC, the Geometric-Semantic Collaborative Learning framework that integrates semantic cues from VLMs with the geometric coherence of VFMs. GeoSeC facilitates cross-model interactions to enhance both the semantic awareness and spatial consistency of prediction maps. Specifically, a semantic evidence perception module strengthens cross-modal interactions to mitigate the remote sensing domain gap. A geometric region transport module then employs an optimal transport plan to assign semantic identities to VFM features, generating spatially continuous geometric references. Finally, a co-enhancement module refines feature representations via consistency constraints to produce complete and coherent activation regions. In terms of M-mIoU, GeoSeC outperforms CTFA by 6.52 percentage points on the iSAID dataset and OME by 18.35 and 7.51 percentage points on the ISPRS Potsdam and Vaihingen datasets, respectively. The code will be released at https://github.com/byyurjaa/GeoSeC.