🧠 Submission to TGRS 2026：From Rejection to Restoration: Hierarchical Representation Learning for Open-Set Domain Adaptive Remote Sensing Segmentation

🧠 Full source code will be released after the paper is accepted.

👓Abstract

Recent studies have shown that unsupervised domain adaptive semantic segmentation (UDASS) has achieved favorable results in remote sensing (RS). However, existing UDASS methods are mainly applicable to closed-set condition and are not well suited to open-set condition, due to the presence of unseen unknown classes in the target domain. Under open-set condition, unknown classes distort the representation space, thereby blurring discriminative boundaries across classes. Moreover, the presence of multiple unknown classes hinders structured unknown-class representation learning. To address the problem mentioned above, a Hierarchical Open-set Segmentation Network (HOSNet) is proposed. First, to purify the representation space of known classes, a Dual-Consensus Known-Unknown Miner (DCKU-Miner) is proposed. DCKU-Miner jointly exploits prediction consistency, confidence, and distribution discrepancy from two complementary decoders to identify reliable known and unknown target features, thereby reducing the unknown-class interference on known-class representation learning. Moreover, to structure the representation space of unknown classes, a Quota-Preserved Unknown Allocator (QPUA) is proposed. QPUA allocates representation capacity to each unknown subclass, alleviating the optimization bias caused by overrepresented unknown classes. Furthermore, to unify the known and unknown representation space, an Anchor-Space Unified Contrastive Learning (AUCL) module is proposed. AUCL uses momentum-updated anchors and a unified contrastive objective to pull features toward assigned anchors and separate them from others, thereby promoting intra-class compactness and inter-class separability. Extensive experiments on the ISPRS and LoveDA benchmarks show that HOSNet outperforms previous methods, achieving average mIoU improvements of 2.15% and 2.98%, respectively.

✨Highlight

• A novel HOSNet is proposed for RS open-set UDASS, which restores the representation space and improves adaptability through hierarchical purification, structuring, and unification.
• A DCKU-Miner is proposed to purify the known-class representation space. It distinguishes reliable known and unknown target features through dual-consensus mining over complementary decoder branches by jointly considering prediction consistency, confidence, and distribution discrepancy.
• A QPUA is proposed to structure the unknown representation space. It assigns reserved representation capacity to each unknown subclass through quota-preserved top-response selection, thereby reducing the optimization bias caused by overrepresented unknown categories.
• A AUCL module is proposed to unify known and unknown representations. It improves intra-class compactness and inter-class separability through anchor-guided contrastive learning with momentum-updated class anchors.

💡Method Overview

👀Visualization

👀Visualization on the ISPRS and LoveDA datasets.

👀t-SNE visualization of feature representations.

📦Usage

📦Datasets

All datasets including ISPRS dataset and LoveDA dataset.

🚀Training

CUDA_VISIBLE_DEVICES=1 nohup python -u tools/train.py > train.log 2>&1 &

📊 Results

📊 Results on the ISPRS dataset

Method	Domain	Surf	Bldg	Vegt	Tree	Car	Bkgd	mIoU (%)	Domain	Surf	Bldg	Vegt	Tree	Car	Bkgd	mIoU (%)
DAFormer	P2V	67.98	77.92	43.72	64.09	43.72	0.01	49.57	PRGB2V	70.01	78.93	15.83	18.02	51.62	0.05	39.08
HRDA	P2V	70.07	74.23	40.34	63.99	50.38	0.61	49.94	PRGB2V	71.33	71.43	14.04	26.71	51.54	0.70	39.29
MIC	P2V	65.28	76.92	45.05	63.85	54.37	0.03	50.92	PRGB2V	58.60	70.85	17.89	23.64	62.15	0.21	38.89
SimT	P2V	70.39	81.28	46.75	64.18	47.13	0.35	51.68	PRGB2V	63.67	75.79	22.96	47.75	46.62	0.08	42.81
MAOSDAN	P2V	70.57	74.82	55.24	36.16	74.16	0.17	51.85	PRGB2V	61.51	62.64	48.36	12.12	70.78	0.09	42.58
GLC++	P2V	69.42	80.55	52.81	61.14	48.69	0.08	52.12	PRGB2V	56.74	68.18	37.84	52.86	41.11	0.68	42.90
BUS	P2V	67.31	79.50	49.93	63.83	53.04	0.03	52.27	PRGB2V	57.88	67.36	31.81	56.20	47.26	0.29	43.47
HOSNet	P2V	68.16	85.94	46.48	73.15	48.64	0.62	53.83	PRGB2V	58.08	76.58	24.17	62.69	48.56	0.71	45.13
DAFormer	V2P	65.51	67.21	53.63	29.19	68.70	2.71	47.82	V2PRGB	58.94	70.40	28.68	22.89	70.58	2.38	42.31
HRDA	V2P	67.53	69.74	54.46	35.50	71.54	2.70	50.24	V2PRGB	66.97	68.77	26.73	24.45	76.33	2.32	44.26
MIC	V2P	70.53	76.78	48.13	8.25	82.62	3.80	48.35	V2PRGB	57.91	75.65	44.80	9.55	78.27	3.30	44.91
SimT	V2P	67.14	78.22	52.79	27.61	71.15	3.97	50.15	V2PRGB	63.57	71.10	51.84	10.34	67.59	3.96	44.73
MAOSDAN	V2P	68.39	72.64	54.52	33.40	70.21	4.98	50.69	V2PRGB	63.46	65.47	51.27	15.61	69.25	4.73	44.97
GLC++	V2P	67.24	72.00	54.22	34.02	70.26	4.77	50.42	V2PRGB	62.45	64.48	50.90	14.01	67.00	4.99	43.97
BUS	V2P	65.56	75.91	55.07	32.63	71.59	3.58	50.72	V2PRGB	57.30	71.66	44.86	24.17	71.06	3.82	45.48
HOSNet	V2P	73.08	77.17	55.81	24.76	83.25	6.15	53.37	V2PRGB	66.36	77.36	48.09	9.75	80.81	6.74	48.19

📊 Results on the LoveDA dataset

Method	Domain	Bkgd	Bldg	Rd	Wtr	Barr	Frst	Agri	mIoU (%)	Domain	Bkgd	Bldg	Rd	Wtr	Barr	Frst	Agri	mIoU (%)
DAFormer	U2R	28.98	31.92	27.12	38.09	13.72	16.88	5.07	23.11	R2U	42.43	41.04	33.71	63.54	27.95	47.60	5.85	37.45
HRDA	U2R	29.23	32.34	27.99	49.38	13.72	5.28	5.92	23.41	R2U	45.89	40.65	33.15	65.20	28.67	44.79	4.96	37.62
MIC	U2R	33.05	29.85	26.37	45.03	13.39	13.28	4.75	23.67	R2U	43.75	40.62	33.08	65.51	26.64	44.36	5.12	37.01
SimT	U2R	28.18	33.13	29.35	47.57	14.82	10.24	5.16	24.06	R2U	43.78	42.09	32.74	59.18	33.84	45.86	5.11	37.51
MAOSDAN	U2R	30.16	33.17	30.42	43.55	14.81	16.14	5.69	24.85	R2U	43.68	41.88	33.36	59.81	32.20	44.26	5.29	37.21
GLC++	U2R	37.08	30.31	24.50	47.93	14.83	18.04	7.03	25.67	R2U	46.51	40.21	39.63	61.19	28.70	47.71	5.53	38.50
BUS	U2R	35.01	32.93	29.83	47.32	14.67	17.76	6.33	26.26	R2U	45.74	46.46	39.50	65.54	30.70	47.53	6.78	40.32
HOSNet	U2R	47.89	31.16	32.43	57.26	9.00	13.97	8.31	28.57	R2U	28.78	54.96	52.33	72.25	41.45	50.20	7.84	43.97

📧Contact

If you encounter any problems or bugs, please don't hesitate to contact me at yiweifang@hhu.edu.cn.