UniMGS: Unsupervised Multimodal Guided Segmentation for High-Resolution Remote Sensing Images

Generating pixel-level annotations for the semantic segmentation of high-resolution remote sensing images is labor-intensive and time-consuming. To address this issue, numerous studies attempt to generate pixel-level annotations through automatic or semi-automatic labeling strategies. However, due to the inherent complexity of remote sensing imagery and the presence of unseen classes, achieving both accurate foreground segmentation and precise class labeling remains highly challenging. In this paper, we propose an Unsupervised Multimodal Guided Segmentation (UniMGS) framework that generates pseudo-labels for remote sensing images without requiring any ground-truth annotations or additional prompts. Specifically, We propose a Mask Purification Module (MPM) that refines class-agnostic segmentation masks by applying an area filter and a ring-merging strategy to remove noisy masks and merge over-segmented regions. To assign precise categories to class-agnostic segmentation masks, we propose a Text-Driven Classifier (TDC) module. This module constructs an intra-class maximum spanning tree from a similarity matrix and incorporates cluster size constraints together with a confidence scoring mechanism, ultimately achieving precise class labeling. Experimental results on multiple remote sensing datasets demonstrate that UniMGS outperforms existing state-of-the-art methods. Our source code is available at https://github.com/nanwang0511/UniMGS.