Toward Stable Semi-Supervised Remote Sensing Segmentation via Co-Guidance and Co-Fusion

TL;DR

This work tackles pseudo-label drift in semi-supervised remote sensing segmentation by introducing Co2S, a drift-resistant dual-student framework that jointly leverages CLIP-based explicit semantic priors and DINOv3-based implicit priors. It couples these priors through an explicit-implicit semantic collaborative guidance mechanism and reinforces them with a global-local feature fusion strategy to balance semantic consistency with boundary precision. The approach demonstrates leading performance across six diverse RS benchmarks, notably under extreme label scarcity, and is supported by extensive ablations highlighting the benefits of heterogeneous priors and stability-driven losses. The proposed framework has practical implications for reducing annotation requirements while delivering accurate segmentation in complex RS scenes.

Abstract

Semi-supervised remote sensing (RS) image semantic segmentation offers a promising solution to alleviate the burden of exhaustive annotation, yet it fundamentally struggles with pseudo-label drift, a phenomenon where confirmation bias leads to the accumulation of errors during training. In this work, we propose Co2S, a stable semi-supervised RS segmentation framework that synergistically fuses priors from vision-language models and self-supervised models. Specifically, we construct a heterogeneous dual-student architecture comprising two distinct ViT-based vision foundation models initialized with pretrained CLIP and DINOv3 to mitigate error accumulation and pseudo-label drift. To effectively incorporate these distinct priors, an explicit-implicit semantic co-guidance mechanism is introduced that utilizes text embeddings and learnable queries to provide explicit and implicit class-level guidance, respectively, thereby jointly enhancing semantic consistency. Furthermore, a global-local feature collaborative fusion strategy is developed to effectively fuse the global contextual information captured by CLIP with the local details produced by DINOv3, enabling the model to generate highly precise segmentation results. Extensive experiments on six popular datasets demonstrate the superiority of the proposed method, which consistently achieves leading performance across various partition protocols and diverse scenarios. Project page is available at https://xavierjiezou.github.io/Co2S/.

Figures (5)

• Figure 1: The radar chart compares the mIoU(%) of different semi-supervised semantic segmentation methods across six remote sensing datasets under the 1/8 labeled ratio. Co2S consistently maintains leading performance across all benchmarks.
• Figure 2: Overview of the proposed Co2S framework. It integrates a CLIP-based student (top) using text embeddings for explicit semantic guidance and a DINOv3-based student (bottom) using learnable queries for implicit guidance. For unlabeled data, the global-local collaborative fusion strategy enforces training stability by arbitrating supervision based on pixel-wise confidence.
• Figure 3: Visualization of attention maps from different heads of the CLIP image encoder (a-c) and DINOv3 backbone (d-f).
• Figure 4: Visual comparison of semantic segmentation results on the six datasets under the 1/8 labeled ratio.
• Figure 5: Evolution of pseudo-label accuracy during the early training phase (first 10 epochs) on the WHDLD dataset (1/24).