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

Yi Zhou^1,* Xuechao Zou^2,* Shun Zhang² Kai Li³ Shiying Wang^1,† Jingming Chen¹
Congyan Lang² Tengfei Cao¹ Pin Tao^1,3 Yuanchun Shi^1,3

¹Qinghai University ²Beijing Jiaotong University ³Tsinghua University

Paper Code Data Checkpoints

Overall Results Comparison

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.

Methodology

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.

Attention Visualization

Visualization of attention maps from different heads of the CLIP image encoder (a-c) and DINOv3 backbone (d-f).

Experimental Results

Comparison with state-of-the-art methods across six datasets.
(Bold: Best, Underline: Second Best)

Table I: Segmentation results (mIoU %) on WHDLD dataset.

Method	1/24	1/16	1/8	1/4
OnlySup	53.6	55.2	58.0	60.5
FixMatch	56.0	57.6	59.8	60.8
U2PL	57.1	58.3	59.9	61.1
WSCL	56.8	58.6	59.8	61.1
UniMatch	57.4	58.8	60.4	61.5
DWL	56.5	57.0	57.1	58.9
MUCA	56.5	58.2	60.0	60.5
Co2S (Ours)	61.1	61.5	62.2	62.6

Table II: LoveDA (1/40 Labeled Images)

Method	Backgr.	Build.	Road	Water	Barren	Forest	Agri.	mIoU
OnlySup	48.0	37.5	36.1	55.3	27.3	53.2	57.1	45.9
FixMatch	50.3	52.1	53.3	69.9	28.7	56.4	62.3	53.3
U2PL	47.4	47.5	50.9	69.8	30.0	56.5	63.8	52.3
WSCL	52.9	52.1	62.8	68.1	31.2	57.7	63.8	54.1
UniMatch	52.8	55.9	54.6	70.1	30.2	59.3	65.5	55.5
DWL	50.4	57.8	57.7	75.7	36.4	58.5	65.1	57.4
MUCA	49.2	50.8	51.0	73.5	33.9	58.9	64.7	54.6
Co2S (Ours)	51.6	59.7	58.3	76.5	35.9	61.5	64.2	58.2

Table II: LoveDA (1/16 Labeled Images)

Method	Backgr.	Build.	Road	Water	Barren	Forest	Agri.	mIoU
OnlySup	50.8	51.0	49.5	65.9	33.5	57.1	63.6	53.1
FixMatch	55.5	58.4	55.4	72.2	37.4	59.3	66.9	57.9
U2PL	53.5	58.3	56.6	72.6	38.0	59.9	67.3	58.0
WSCL	53.8	55.8	55.2	71.6	37.5	61.3	66.6	57.4
UniMatch	55.7	59.2	57.0	72.4	37.7	60.9	67.6	58.6
DWL	52.7	60.6	59.1	73.7	36.0	62.9	68.2	59.0
MUCA	52.6	54.1	54.5	75.0	34.1	61.2	65.0	56.7
Co2S (Ours)	54.0	62.2	59.0	76.3	40.0	63.6	67.6	60.4

Table II: LoveDA (1/8 Labeled Images)

Method	Backgr.	Build.	Road	Water	Barren	Forest	Agri.	mIoU
OnlySup	52.8	56.4	53.4	69.8	38.0	59.7	66.2	56.6
FixMatch	55.9	60.9	58.8	73.1	40.6	62.3	69.2	60.1
U2PL	54.8	60.9	59.4	72.0	40.3	61.6	68.4	59.6
WSCL	56.6	60.0	58.5	72.9	38.6	61.1	68.0	59.4
UniMatch	56.8	61.1	59.6	73.3	39.4	61.7	68.3	60.0
DWL	56.2	61.5	60.0	78.9	47.1	66.6	69.4	62.8
MUCA	55.5	57.4	58.2	77.3	38.3	65.4	69.2	60.2
Co2S (Ours)	57.2	63.8	60.4	79.3	42.3	66.7	69.2	62.7

Table II: LoveDA (1/4 Labeled Images)

Method	Backgr.	Build.	Road	Water	Barren	Forest	Agri.	mIoU
OnlySup	55.7	58.8	57.9	72.8	42.3	53.2	68.5	58.9
FixMatch	56.9	61.9	60.6	73.8	43.3	64.2	69.8	61.5
U2PL	57.4	61.5	60.6	74.6	42.9	64.3	69.2	61.5
WSCL	57.6	61.3	59.3	74.3	43.4	63.2	68.6	61.1
UniMatch	58.0	61.1	60.8	74.2	44.6	63.5	69.8	61.7
DWL	56.5	61.8	60.6	79.3	44.5	68.8	72.4	63.4
MUCA	55.9	61.8	61.1	78.0	44.8	68.3	71.6	63.1
Co2S (Ours)	57.0	63.4	61.0	79.1	48.3	67.8	71.2	64.0

Table III: Segmentation results (mIoU %) on Potsdam dataset.

Method	1/32	1/16	1/8	1/4
OnlySup	61.5	64.9	70.2	73.8
FixMatch	69.4	71.5	73.5	75.7
U2PL	67.2	71.3	73.3	75.9
WSCL	69.9	72.3	73.9	75.4
UniMatch	70.7	72.6	74.8	76.4
DWL	74.2	78.3	79.8	80.3
MUCA	66.9	71.5	76.0	79.2
Co2S (Ours)	74.3	76.6	79.8	80.2

Table IV: MER (1/8 Labeled Images)

Method	Soil	Sands	Gravel	Bedrock	Rocks	Tracks	Shadows	Unk.	Bg.	mIoU
OnlySup	4.7	55.1	71.4	61.7	38.0	49.5	34.2	73.0	87.0	52.8
FixMatch	2.0	53.6	73.5	65.8	40.0	54.8	31.5	72.4	90.4	53.8
U2PL	5.2	51.8	73.2	70.5	45.7	54.6	26.2	73.8	93.2	54.9
WSCL	3.5	54.8	69.9	62.8	36.0	56.9	26.1	76.0	80.9	51.9
UniMatch	5.5	54.4	72.7	68.5	44.7	55.2	38.0	73.8	91.1	56.0
DWL	5.9	51.0	74.9	67.2	41.4	61.9	28.3	69.7	90.4	54.5
MUCA	5.3	56.0	72.6	60.5	36.1	52.9	34.5	67.7	87.8	52.6
Co2S (Ours)	3.0	56.3	73.1	67.9	37.0	66.3	36.5	79.7	91.3	56.8

Table IV: MER (1/4 Labeled Images)

Method	Soil	Sands	Gravel	Bedrock	Rocks	Tracks	Shadows	Unk.	Bg.	mIoU
OnlySup	21.6	56.3	74.4	66.3	38.0	52.3	22.1	74.5	92.2	55.3
FixMatch	19.4	55.4	74.0	57.1	43.6	57.3	39.8	75.7	90.9	58.1
U2PL	7.1	56.4	73.7	69.7	41.9	60.8	23.3	81.4	90.3	56.1
WSCL	8.2	52.8	71.1	65.5	35.7	61.0	31.4	79.0	89.0	54.9
UniMatch	19.7	54.6	74.3	69.3	44.5	54.9	37.4	77.4	90.5	58.1
DWL	13.8	55.9	74.8	70.3	36.2	57.9	28.8	66.0	89.7	54.7
MUCA	1.0	55.1	73.8	65.6	38.4	56.3	28.6	73.7	84.9	53.0
Co2S (Ours)	13.1	63.3	76.1	69.4	37.1	63.1	41.1	77.1	91.4	59.1

Table V: MSL (1/8 Labeled Images)

Method	Soil	Sands	Gravel	Bedrock	Rocks	Tracks	Shadows	Unk.	Bg.	mIoU
OnlySup	26.8	61.6	66.8	72.2	37.9	45.5	37.0	49.1	85.1	53.6
FixMatch	31.0	65.4	69.6	75.0	40.4	49.6	39.8	54.2	86.5	56.8
U2PL	26.7	64.7	70.3	73.3	39.8	43.1	40.3	53.1	87.7	55.5
WSCL	25.8	65.0	68.3	73.9	38.8	69.6	39.6	57.0	86.0	58.2
UniMatch	29.3	66.7	69.8	75.4	39.2	50.8	43.0	57.9	89.2	57.9
DWL	33.5	65.9	71.2	74.7	44.7	48.8	44.8	56.3	90.9	59.0
MUCA	15.4	63.9	67.0	71.5	43.1	41.0	41.0	55.9	82.9	52.4
Co2S (Ours)	36.7	68.4	71.7	76.0	42.6	58.9	44.1	60.9	89.3	60.9

Table V: MSL (1/4 Labeled Images)

Method	Soil	Sands	Gravel	Bedrock	Rocks	Tracks	Shadows	Unk.	Bg.	mIoU
OnlySup	32.2	65.0	69.0	75.1	43.6	57.6	44.0	57.9	86.6	59.0
FixMatch	29.7	65.5	69.9	77.1	42.5	63.8	48.1	60.7	87.5	60.5
U2PL	36.1	62.7	72.7	76.3	43.3	63.7	44.2	65.6	90.1	61.6
WSCL	30.9	64.7	70.4	75.5	42.0	58.3	42.8	65.0	89.7	59.9
UniMatch	32.2	67.7	72.6	77.0	42.5	71.8	49.1	61.9	90.8	62.8
DWL	33.8	68.2	72.3	76.1	43.2	57.9	48.2	70.7	91.9	63.6
MUCA	33.7	69.3	72.9	76.9	46.2	40.4	49.9	55.6	89.2	59.3
Co2S (Ours)	36.0	67.7	71.8	77.8	45.5	77.0	49.2	76.7	91.7	65.9

Table VI: Segmentation results (mIoU %) on GID-15 dataset.

Method	1/8	1/4
OnlySup	67.5	74.1
FixMatch	70.0	74.8
U2PL	67.2	75.3
WSCL	72.8	75.4
UniMatch	73.9	75.9
DWL	72.6	76.9
MUCA	67.4	72.6
Co2S (Ours)	75.4	77.7

Visual Comparisons & Qualitative Results

Swipe to view results across different datasets

Visual Comparison WHDLD Dataset

Visual Comparison on LoveDA Dataset

Visual Comparison MER Dataset

Visual Comparison MSL Dataset

Citation

@article{Co2S, title={Toward Stable Semi-Supervised Remote Sensing Segmentation via Co-Guidance and Co-Fusion}, author={Zhou, Yi and Zou, Xuechao and Zhang, Shun and Li, Kai and Wang, Shiying and Chen, Jingming and Lang, Congyan and Cao, Tengfei and Tao, Pin and Shi, Yuanchun}, year={2025}, }

Toward Stable Semi-Supervised Remote SensingSegmentation via Co-Guidance and Co-Fusion