A cross-domain semantic segmentation framework fusing relative depth for high-resolution optical satellite remote sensing imagery

doi:10.11947/j.AGCS.2026.20260149

Abstract

Abstract:

Semantic segmentation of high-resolution optical satellite remote sensing imagery serves as a key technique. It supports the intelligent recognition and dynamic monitoring of large-scale multi-scenario land surface elements. It is of great significance for applications such as land survey, disaster emergency response and urban planning. However, cross-domain discrepancies in imagery can be induced by variations in sensors, satellite imaging angles and regional geographical environments. Such discrepancies lead to significant performance degradation of semantic segmentation models in unseen domains. Existing cross-domain semantic segmentation methods suffer from insufficient exploitation of spatial structure information and heavy reliance on target-domain images for training. To address these issues, this paper proposes a cross-domain semantic segmentation framework with relative depth as spatial prior knowledge. Relative topographic relief extracted by vision foundation models is utilized as cross-domain prior information. Cross-domain pixel-wise semantic category inference is achieved through adaptive multi-scale feature fusion and edge supervision signals. Experiments are conducted on the public satellite remote sensing cross-domain dataset LoveDA, a self-built circumpolar dataset and a cross-sensor building extraction data configuration. Results demonstrate that the introduction of relative depth can effectively improve cross-domain semantic segmentation performance. Without using target-domain images for training, the proposed method can achieve inference performance close to even surpassing that of state-of-the-art unsupervised domain adaptation methods. The necessity of relative depth as prior information, the effectiveness of the proposed framework and its potential in intelligent interpretation of global high-resolution satellite imagery are validated.

Key words: high-resolution satellite remote sensing imagery, cross-domain, semantic segmentation, relative depth, feature fusion

CLC Number:

P237

Yongqi Sun, Chenguang Dai, Zhenchao Zhang, Jinchun Qin, Yu Su. A cross-domain semantic segmentation framework fusing relative depth for high-resolution optical satellite remote sensing imagery[J]. Acta Geodaetica et Cartographica Sinica, 2026, 55(8): 1425-1438.

Add to citation manager EndNote|Reference Manager|ProCite|BibTeX|RefWorks

URL: http://xb.chinasmp.com/EN/10.11947/j.AGCS.2026.20260149

http://xb.chinasmp.com/EN/Y2026/V55/I8/1425

Figures/Tables 14

Fig. 1

Fig. 2

Fig. 3

Tab. 1

Fig. 4

Tab. 2

Tab. 3

Tab. 4

Fig. 5

Tab. 5

Tab. 6

Tab. 7

Tab. 8

Tab. 9

References 32

[1]	Wu Kang, Zhang Yingying, Ru Lixiang, et al. A semantic-enhanced multi-modal remote sensing foundation model for Earth observation[J]. Nature Machine Intelligence, 2025, 7(8): 1235-1249.
[2]	Li Yansheng, Shi Te, Zhang Yongjun, et al. Learning deep semantic segmentation network under multiple weakly-supervised constraints for cross-domain remote sensing image semantic segmentation[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2021, 175: 20-33.
[3]	沈秭扬, 倪欢, 管海燕. 遥感图像跨域语义分割的无监督域自适应对齐方法[J]. 测绘学报, 2023, 52(12): 2115-2126. DOI: . doi: 10.11947/j.AGCS.2023.20220483
	Shen Ziyang, Ni Huan, Guan Haiyan. Unsupervised domain adaptation alignment method for cross-domain semantic segmentation of remote sensing images[J]. Acta Geodaetica et Cartographica Sinica, 2023, 52(12): 2115-2126. DOI: . doi: 10.11947/j.AGCS.2023.20220483
[4]	张继贤, 顾海燕, 杨懿, 等. 高分辨率遥感影像智能解译研究进展与趋势[J]. 遥感学报, 2021, 25(11): 2198-2210.
	Zhang Jixian, Gu Haiyan, Yang Yi, et al. Research progress and trend of high-resolution remote sensing imagery intelligent interpretation[J]. Journal of Remote Sensing, 2021, 25(11): 2198-2210.
[5]	顾海燕, 杨懿, 李海涛, 等. 高分辨率遥感影像样本库动态构建与智能解译应用[J]. 测绘学报, 2024, 53(6): 1165-1179. DOI: . doi: 10.11947/j.AGCS.2024.20230469
	Gu Haiyan, Yang Yi, Li Haitao, et al. Dynamic construction of high-resolution remote sensing image sample datasets and intelligent interpretation applications[J]. Acta Geodaetica et Cartographica Sinica, 2024, 53(6): 1165-1179. DOI: . doi: 10.11947/j.AGCS.2024.20230469
[6]	Li Da, Yang Yongxin, Song Yizhe, et al. Learning to generalize: Meta-learning for domain generalization[C]//Proceedings of 2018 AAAI Conference on Artificial Intelligence. 2018, 32(1): 11596.
[7]	Kim D, Yoo Y, Park S, et al. SelfReg: self-supervised contrastive regularization for domain generalization[C]//Proceedings of 2021 IEEE/CVF International Conference on Computer Vision (ICCV). Montreal: IEEE, 2021: 9599-9608.
[8]	Pan Xingang, Luo Ping, Shi Jianping, et al. Two at once: enhancing learning and generalization capacities via IBN-net[C]//Proceedings of 2018 Computer Vision. Cham: Springer International Publishing, 2018: 484-500.
[9]	Hoyer L, Dai Dengxin, Van Gool L. DAFormer: improving network architectures and training strategies for domain-adaptive semantic segmentation[C]//Proceedings of 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New Orleans: IEEE, 2022: 9914-9925.
[10]	Hoyer L, Dai Dengxin, Wang Haoran, et al. MIC: masked image consistency for context-enhanced domain adaptation[C]//Proceedings of 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Vancouver: IEEE, 2023: 11721-11732.
[11]	Hoyer L, Dai Dengxin, Van Gool L. Domain adaptive and generalizable network architectures and training strategies for semantic image segmentation[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024, 46(1): 220-235.
[12]	Zhao Qi, Lyu Shuchang, Zhao Hongbo, et al. Self-training guided disentangled adaptation for cross-domain remote sensing image semantic segmentation[J]. International Journal of Applied Earth Observation and Geoinformation, 2024, 127: 103646.
[13]	Chen Jie, Zhu Jingru, He Peien, et al. Unsupervised domain adaptation for building extraction of high-resolution remote sensing imagery based on decoupling style and semantic features[J]. IEEE Transactions on Geoscience and Remote Sensing, 2024, 62: 4406917.
[14]	Hu Wenshuai, Li Wei, Li Hengchao, et al. Unsupervised domain adaptation with hierarchical masked dual-adversarial network for end-to-end classification of multisource remote sensing data[J]. IEEE Transactions on Geoscience and Remote Sensing, 2025, 63: 4409917.
[15]	Guo Xin, Lao Jiangwei, Dang Bo, et al. SkySense: a multi-modal remote sensing foundation model towards universal interpretation for earth observation imagery[C]//Proceedings of 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle: IEEE, 2024: 27662-27673.
[16]	Li Xuyang, Li Chenyu, Ghamisi P, et al. FlexiMo: a flexible remote sensing foundation model[J]. IEEE Transactions on Geoscience and Remote Sensing, 2026, 64: 5606516.
[17]	Li Xuyang, Li Chenyu, Vivone G, et al. SeaMo: a season-aware multimodal foundation model for remote sensing[J]. Information Fusion, 2026, 125: 103334.
[18]	Wang Junjue, Zheng Zhuo, Ma Ailong, et al. LoveDA: a remote sensing land-cover dataset for domain adaptive semantic segmentation[PP/OL]. V6. arXiv (2022-05-31) [2026-04-16]. https://doi.org/10.48550/arXiv.2110.08733.
[19]	Yang Lihe, Kang Bingyi, Huang Zilong, et al. Depth anything: unleashing the power of large-scale unlabeled data[C]//Proceedings of 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Seattle: IEEE, 2024: 10371-10381.
[20]	Liu Zhuang, Mao Hanzi, Wu Chaoyuan, et al. A ConvNet for the 2020s[C]//Proceedings of 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition. New Orleans: IEEE, 2022: 11966-11976.
[21]	Liu Ze, Lin Yutong, Cao Yue, et al. Swin transformer: hierarchical vision transformer using shifted windows[C]//Proceedings of 2021 IEEE/CVF International Conference on Computer Vision. Montreal: IEEE, 2021: 9992-10002.
[22]	Berman M, Triki A R, Blaschko M B. The lovasz-softmax loss: a tractable surrogate for the optimization of the intersection-over-union measure in neural networks[C]//Proceedings of 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City: IEEE, 2018: 4413-4421.
[23]	Ji Shunping, Wei Shiqing, Lu Meng. Fully convolutional networks for multisource building extraction from an open aerial and satellite imagery data set[J]. IEEE Transactions on Geoscience and Remote Sensing, 2019, 57(1): 574-586.
[24]	Luo Yawei, Zheng Liang, Guan Tao, et al. Taking a closer look at domain shift: category-level adversaries for semantics consistent domain adaptation[C]//Proceedings of 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Long Beach: IEEE, 2019: 2502-2511.
[25]	Tsai Y H, Hung W C, Schulter S, et al. Learning to adapt structured output space for semantic segmentation[C]//Proceedings of 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City: IEEE, 2018: 7472-7481.
[26]	Mei Ke, Zhu Chuang, Zou Jiaqi, et al. Instance adaptive self-training for unsupervised domain adaptation[C]//Proceedings of 2020 Computer Vision. Cham: Springer International Publishing, 2020: 415-430.
[27]	Wu Linshan, Lu Ming, Fang Leyuan. Deep covariance alignment for domain adaptive remote sensing image segmentation[J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 5620811.
[28]	Xie Enze, Wang Wenhai, Yu Zhiding, et al. segFormer: Simple and efficient design for semantic segmentation with transformers[J]. Advances in Neural Information Processing Systems, 2021, 34: 12077-12090.
[29]	Xiao Tete, Liu Yingcheng, Zhou Bolei, et al. Unified perceptual parsing for scene understanding[C]//Proceedings of 2018 Computer Vision-ECCV 2018. Cham: Springer International Publishing, 2018: 432-448.
[30]	Yang Lihe, Kang Bingyi, Huang Zilong, et al. Depth anything V2[PP/OL]. V2. arXiv (2024-10-20) [2026-04-16]. https://doi.org/10.48550/arXiv.2406.09414.
[31]	Ebrahimi S, Arik S O, Nama T, et al. CROME: cross-modal adapters for efficient multimodal LLM[PP/OL]. arXiv (2024-08-13) [2026-04-16]. https://doi.org/10.48550/arXiv.2408.06610.
[32]	Zhang Haodi, Yu Anzhu, Gao Kuiliang, et al. M²Caps: learning multi-modal capsules of optical and SAR images for land cover classification[J]. International Journal of Digital Earth, 2025, 18: 2447347.

方法	类型	模态	IoU							mIoU
方法	类型	模态	背景	建筑物	道路	水系	裸地	森林	农田	mIoU
CLAN^[24]	对抗训练	光学	22.93	44.78	25.99	46.81	10.54	37.21	24.45	30.39
AdaptSegNet^[25]	对抗训练		26.89	40.53	30.65	50.09	16.97	32.51	28.25	32.27
IAST^[26]	自训练		29.97	49.48	28.29	64.49	2.13	33.36	61.37	38.44
SegFormer^[28]	域泛化		26.60	55.80	35.62	65.44	17.13	40.13	39.65	40.05
DCA^[27]	自训练		36.38	55.89	40.46	62.03	22.01	38.92	60.52	45.17
DAFormer^[9]	自训练		37.39	52.84	41.99	72.05	11.46	46.79	61.27	46.25
HRDA^[11]	自训练		36.80	64.14	40.92	71.80	14.36	47.17	68.37	49.08
MIC^[10]	自训练		39.23	61.92	43.85	72.38	13.92	47.67	68.87	49.69
ST-DASegNet（SegFormer）^[12]	对抗训练+自训练		36.78	59.83	43.77	73.83	19.38	49.96	67.01	50.08
ConvNeXt-B（基线模型）	域泛化	光学+相对深度	35.51	56.14	40.88	70.05	13.99	48.42	64.91	47.13
ConvNeXt-B（光学）+ConvNeXt-S（相对深度）	域泛化		36.79	59.27	46.97	71.31	19.87	41.62	67.05	48.98
ConvNeXt-XL（光学）+ConvNeXt-B（相对深度）	域泛化		37.23	63.05	42.69	71.71	19.91	45.03	68.36	49.71
ConvNeXt-B（光学）+ConvNeXt-S（相对深度）+DF+BAAL	域泛化		38.77	65.72	39.32	72.38	16.70	48.36	68.06	49.90

方法	类型	模态	IoU							mIoU
方法	类型	模态	背景	建筑物	道路	水系	裸地	森林	农田	mIoU
CLAN	对抗训练	光学	43.41	25.42	13.75	79.25	13.71	30.44	25.80	33.11
AdaptSegNet	对抗训练		42.35	23.73	15.61	81.95	13.62	28.70	22.05	32.68
IAST	自训练		48.57	31.51	28.73	86.01	20.29	31.77	36.50	40.48
SegFormer	域泛化		44.98	43.93	27.46	85.82	16.24	37.02	30.48	40.85
DCA	自训练		45.82	49.60	51.65	80.88	16.70	42.93	36.92	46.36
DAFormer	自训练		50.94	56.66	62.83	89.41	11.99	45.81	25.26	48.99
HRDA	自训练		48.25	45.24	59.16	87.17	18.82	44.94	27.61	47.31
MIC	自训练		50.66	49.56	49.46	87.87	19.51	43.48	34.66	49.31
ST-DASegNet（SegFormer）	对抗训练+自训练		51.01	54.23	60.52	87.31	15.18	47.23	36.26	50.28
ConvNeXt-B（基线模型）	域泛化		46.77	46.52	37.27	82.30	17.33	42.34	36.27	44.11
ConvNeXt-B（光学）+ConvNeXt-S（相对深度）	域泛化		48.23	52.68	42.70	77.18	16.62	40.85	48.86	46.73
ConvNeXt-XL（光学）+ConvNeXt-B（相对深度）	域泛化	光学+相对深度	50.56	51.81	42.89	85.11	16.69	42.67	44.00	47.68
ConvNeXt-B（光学）+ConvNeXt-S（相对深度）+DF+BAAL	域泛化	光学+相对深度	48.95	52.80	52.26	84.29	18.18	44.27	46.49	49.61

方法	类型	模态	IoU									mIoU
			细粒度类别（道路）			背景	粗粒度类别
			跑道	滑行道和保险道	其他道路	背景	植被	水系	裸地	建筑物	冰雪
Swin-Base	域泛化	光学	62.08	11.50	34.65	37.19	89.64	33.11	51.69	44.68	42.38	45.21
SegFormer（MiT-B5）	域泛化		47.19	10.81	33.27	37.26	90.26	43.58	50.57	43.17	46.84	44.77
Mask2Former（Swin-Base）	域泛化		49.08	18.38	37.06	31.60	89.65	50.82	51.05	46.15	45.06	46.54
DAFormer	自训练	光学	46.27	14.63	36.27	36.62	86.58	46.90	53.45	42.19	33.74	44.07
HRDA	自训练		55.36	15.90	33.47	32.91	90.22	44.58	52.82	39.15	47.57	45.59
MIC	自训练		53.86	21.25	31.85	37.44	88.45	48.96	50.96	40.82	37.73	45.70
ConvNeXt-B（基线模型）	域泛化	光学+相对深度	49.71	14.72	36.10	32.57	90.36	27.38	50.19	38.61	46.91	42.95
ConvNeXt-XL	域泛化		59.76	9.91	27.87	35.10	90.80	48.82	52.64	36.26	48.76	45.55
ConvNeXt-B（光学）+ConvNeXt-S（相对深度）	域泛化		67.56	17.47	35.19	34.02	89.09	29.72	51.01	48.24	42.07	46.04
ConvNeXt-XL（光学）+ConvNeXt-B（相对深度）	域泛化		66.35	14.45	35.83	37.21	89.65	39.06	52.28	46.44	42.91	47.13
ConvNeXt-B（光学）+ConvNeXt-S（相对深度）+DF+BAAL	域泛化		61.30	20.56	37.45	39.64	89.20	52.62	46.53	47.27	35.66	47.80

方法	类型	模态	建筑物IoU/（%）
DAFormer	自训练	光学	67.81
HRDA	自训练	光学	73.05
MIC	自训练	光学	74.05
ConvNeXt-B（光学）+ConvNeXt-S（相对深度）+DF+BAAL	域泛化	光学+相对深度	74.58

模型	模态	总参数量（M）	FLOPs（GMacs）	FPS/（张/s）	SUSAN数据集mIoU/（%）
ConvNeXt-B	光学	120.81	291.54	67.78	42.95
ConvNeXt-XL	光学	388.67	532.91	45.61	45.55
ConvNeXt-B（光学）+ConvNeXt-S（相对深度）+DF+BAAL	光学+相对深度	169.97	349.48	42.48	47.80