Journal of Geo-information Science

MDSNet : A Multi-Scale Depth Supervision Method for High-Resolution Remote Sensing Image Semantic Segmentation

SHAN Huilin, WANG Xingtao, LIU Wenxing, WU Xinyue, GAO Runze, LI Hongxu

Journal of Geo-information Science, 2025, 27(6): 1381-1400. DOI: 10.12082/dqxxkx.2025.250009

Backbone				mIoU	mF1	mRecall	OA
SDCM	Double Conv	ResT-Mamba	CEB	mIoU	mF1	mRecall	OA
	√		√	85.87	92.32	92.73	92.50
	√	√		86.07	92.43	92.89	92.62
√			√	86.02	92.39	92.88	92.58
√		√		86.49	92.68	93.19	92.87

Tab. 2 Quantitative results of different branch stems (Vaihingen) (%)

Other figure/table from this article

Fig. 1 The model structure of MDSNet
Fig. 2 Space De-redundant convolution module
Fig. 3 Feature expressiveness enhances visualization
Fig. 4 The structure of channel reweight concat
Fig. 5 The structure of ResT-Mamba
Fig. 6 The structure of multi-scale convolutional feature fusion module
Fig. 7 Multi-scale convolutional attention module
Tab. 1 Experimental configuration information
Fig. 8 Examples of ISPRS dataset
Fig. 9 Comparison of wavelet transform ablation experiments
Tab. 3 The effect comparison test of each module (Vaihingen)
Fig. 10 Comparison of CRC ablation experiments
Tab. 4 Segmentation results under different network models on the Vaihingen dataset (%)
Fig. 11 Vaihingen data on the distribution of recall rates
Tab. 5 Segmentation results under different network models on the Potsdam dataset (%)
Fig. 12 Segmentation of low vegetation and trees in Potsdam
Fig. 13 Segmentation of Impervious surfaces and buildings in Potsdam
Fig. 14 Segmentation of Impervious surfaces and buildings in Potsdam
Fig. 15 Potsdam overall segmentation diagram