Organizations: the School of Computer Science and Technology, Zhejiang University, Hangzhou 310027, China · School of Instrumentation Science and Engineering, Harbin Institute of Technology, Harbin 150001, China, and also with Harbin Institute of Technology Suzhou Research Institute, Suzhou 215000, China · the School of Software Technology, Zhejiang University, Hangzhou 310027, China · School of Information Engineering, North China University of Water Resources and Electric Power, Zhengzhou 450046, China, and also with School of Computer Science, University of Auckland, Auckland, New Zealand · the School of Software Technology, Zhejiang University, Hangzhou 310027, China, and also with the Innovation Center of Yangtze River Delta, Zhejiang University, Jiaxing, Zhejiang 314103, China
In remote sensing imagery, multi-class change detection (MCD) is crucial for fine-grained monitoring, yet it has long been plagued by feature ambiguity in change regions, which obscures subtle changes and hinders fine-grained discrimination under complex scene variations and scarce annotations. To address this, we propose the Tri-path DINO architecture, which adopts a multi-path decoupled complementary feature learning strategy to separately capture coarse-grained semantic features and fine-grained target-edge features. Specifically, we employ the DINOv3 pre-trained model as the backbone feature extraction network to learn coarse-grained features. An auxiliary path also adopts a siamese structure, progressively aggregating intermediate features from the siamese encoder to enhance the learning of fine-grained features. Finally, a multi-scale attention mechanism is introduced to augment the decoder network, where parallel convolutions adaptively capture and enhance contextual information under different receptive fields. The proposed method achieves state-of-the-art performance on Gaza-change, SECOND, and LandsatSCD, improving mIoU/Fscd by 1.01%/2.93%, 0.67%/2.84%, and 1.22%/1.51%, respectively, over the best competing method. Grad-CAM visualizations further confirm that the main and auxiliary paths naturally focus on coarse-grained semantic changes and fine-grained structural details, respectively. This synergistic complementarity provides a robust and interpretable solution for advanced change detection tasks, offering a basis for rapid and accurate damage assessment.
Figures & tables
Figure 1: Overall architecture of the proposed Tri-path DINO. Our main path consists of a Siamese-structured convolutional network designed to extract local information, thus favoring more fine-grained details. In contrast, the third path adopts a transformer architecture to capture global information. Owing to this complementary structural design, the main path and the third branch are induced to learn the features they each excel at, namely coarse-grained and fine-grained features.
Figure 2: Example results on the Gaza-Change dataset.
Figure 3: Qualitative comparison on the Gaza-Change dataset, with color coding: true positives (white), true negatives (black), false positives (red), and false negatives (green).
Figure 4: Example results on the SECOND dataset.
Figure 5: Qualitative comparison on the SECOND dataset, with color coding: true positives (white), true negatives (black), false positives (red), and false negatives (green).
Figure 6: Example results on the LandsatSCD dataset.
Figure 7: Qualitative comparison on the LandsatSCD dataset, with color coding: true positives (white), true negatives (black), false positives (red), and false negatives (green).
Figure 8: Heatmaps based on the GradCAM visualization of the SECOND dataset. Path represents the features learned by the third pathway, while Backbone represents the features learned by the backbone siamese network.
Figure 9: Heatmaps based on the GradCAM visualization of the Gaza-Change dataset. Path represents the features learned by the third pathway, while Backbone represents the features learned by the backbone siamese network.
Figure 10: Heatmaps based on the GradCAM visualization of the LandsatSCD dataset. Path represents the features learned by the third pathway, while Backbone represents the features learned by the backbone siamese network.
Figure 11: Radar chart of comprehensive performance comparison between the proposed method and existing change detection models.
Model Size
Gaza-Change
Method
Params (M)
Trainable Params (M)
FLOPs (G)
OA
mIoU
SeK
Fscd
P
R
BIT
11.89
11.89
35.38
95.74
68.81
8.91
57.85
51.71
56.48
CDxLSTM
16.18
16.18
15.45
96.06
67.97
7.30
55.36
54.79
47.06
ChangeFormer
41.09
41.09
814.43
95.92
66.48
5.29
52.64
52.54
43.15
DDLNet
13.78
13.78
29.44
96.28
69.39
10.28
58.87
57.41
51.73
LGPNet
70.99
70.99
503.14
95.49
68.24
9.16
56.97
50.29
57.91
Table 1: Comprehensive comparison with state-of-the-art methods on the Gaza-Change dataset. Best results are highlighted in bold .
SECOND
Method
OA
mIoU
Sek
F scd
P
R
BIT
86.77
70.79
19.28
54.52
43.74
25.51
CDxLSTM
86.29
70.38
17.89
52.58
37.25
21.95
ChangeFormer
86.36
70.26
18.35
53.55
36.89
23.46
DDLNet
86.47
71.12
19.06
53.78
40.01
25.88
LGPNet
86.57
69.67
17.49
53.05
45.86
20.86
Table 2: Comprehensive comparison with state-of-the-art methods on the SECOND dataset. Best results are highlighted in bold .
LandsatSCD
Method
OA
mIoU
Sek
F scd
P
R
BIT
93.62
82.80
41.92
80.82
85.79
76.08
CDxLSTM
92.35
79.85
34.41
76.77
81.73
64.23
ChangeFormer
92.79
80.35
35.28
77.41
82.95
57.77
DDLNet
93.52
82.59
41.25
80.75
86.24
74.01
RSmamba
93.32
81.31
38.02
79.11
85.25
61.76
Table 3: Comprehensive comparison with state-of-the-art methods on the LandsatSCD dataset. Best results are highlighted in bold .
Path
MLHA
OA
mIoU
Sek
Fscd
95.63
69.66
11.79
59.42
✓
96.33
70.64
11.73
60.8
✓
✓
96.45
71.35
13.02
62.58
Table 4: Module-level ablation results on the Gaza-change dataset. Best scores are in bold .
Path
MLHA
OA
mIoU
Sek
Fscd
85.46
70.22
18.6
53.26
✓
85.97
71.14
19.7
53.91
✓
✓
86.33
71.62
20.61
55.44
Table 5: Module-level ablation results on the SECOND dataset. Best scores are in bold .
Path
MLHA
OA
mIoU
SeK
Fscd
94.63
85.39
48.06
82.72
✓
94.81
86.12
49.63
84.38
✓
✓
95.09
86.43
50.95
85.31
Table 6: Module-level ablation results on the LandsatSCD dataset. Best scores are in bold .
Path
LoRA
Adapter
Trainable params (M)
Mean F1 (%)
Δ F1 (pp)
SeK (%)
✓
✓
✓
16.38
84.35
0.00
50.95
✓
✓
15.32
76.98
−7.37
39.11
✓
✓
15.42
82.84
−1.51
48.51
Table 7: Adaptation-module ablation results on LandsatSCD. Mean F1 denotes the average F1 over change categories and is distinct from Fscd . Best scores are in bold .
Focal
Dice
Lovász
Mean F1 (%)
SeK (%)
Fscd (%)
mIoU (%)
0.4
0.3
0.3
84.35
50.95
85.31
86.43
0.6
0.2
0.2
83.09
49.01
84.39
85.77
0.2
0.6
0.2
83.14
48.80
84.33
85.69
0.2
0.2
0.6
82.32
47.47
83.85
85.23
1.0
0.0
0.0
82.63
49.79
84.59
86.08
0.0
1.0
0.0
79.83
47.36
83.48
85.40
Table 8: Comparison of loss configurations on LandsatSCD. Best scores are highlighted in bold.
Figure 12: Representative failure cases on the Gaza-Change dataset. Columns show T1, T2, ground truth (GT), and prediction (Pred). From top to bottom: small-change miss, similar-class confusion, and background pseudo-change. Cyan boxes indicate the regions of interest.
Figure 13: Representative failure cases on the SECOND dataset. Columns show T1, T2, ground truth (GT), and prediction (Pred). From top to bottom: small-change miss, fine-boundary error, and similar-class confusion. Cyan boxes indicate the regions of interest.
Figure 14: Representative failure cases on the LandsatSCD dataset. Columns show T1, T2, ground truth (GT), and prediction (Pred). From top to bottom: small-change miss, similar-class confusion, and background false-change. Cyan boxes indicate the regions of interest.
Semantic change detection (SCD) aims to simultaneously locate land-cover changes and identify semantic categories before and after transition. However, existing methods suffer from insufficient cross-temporal alignment, weak multi-scale representation, and poor robustness to pseudo-changes caused by illumination, season, and registration noise. To address these issues, we propose a novel end-to-end semantic change detection network named SemDINO, which integrates a dual-branch encoder, multi-scale temporal interaction, semantic purification, change enhancement, and decoupled multi-task prediction into a unified framework. Specifically, we construct a dual-branch encoder that combines a CNN backbone and frozen DINOv3 features via gated pyramid fusion, enabling rich multi-scale semantic representation. Then, a multi-scale temporal bidirectional transformer interaction (M-TBTT) module is proposed to achieve global cross-temporal feature alignment and information interaction. To further enhance genuine changes and suppress pseudo-variations, we introduce semantic purification (SCP), bidirectional change enhancement (BiChangeEnhance), and multi-scale change enhancement (MCE) modules collaboratively. Finally, a multi-branch CD prediction head is designed to jointly output binary change mask, bi-temporal semantic maps, and edge constraint. Extensive experiments on public remote sensing CD datasets demonstrate that SemDINO achieves superior performance and generalization ability against state-of-the-art methods, especially in complex scenarios with interference factors.
Xinyu Tong, Meihua Zhou, Jinxiao Sun +2
Xinjiang Institute of Ecology and Geography, Chinese Academy of Sciences, Urumqi 830011, China · University of Chinese Academy of Sciences, Beijing 100049, China · School of Computer Science, Xiangtan University, Xiangtan 411105, China +1
Vision foundation models (VFMs) such as DINO are pretrained for single-image representation, whereas remote sensing change detection requires reasoning over a bi-temporal pair. Existing VFM-based methods usually encode the two images independently and compare them only afterward, leaving the VFM backbone unaware of cross-temporal relations. To bridge this mismatch, we present AdaDINO, a pair-aware in-backbone adaptation framework that equips a frozen DINO encoder with bi-temporal interaction for efficient change detection. Its core component, Change-aware Gated Local Adaptation (CGLA), couples the two streams after selected frozen blocks and injects a shared temporal residual into them with opposite signs, enhancing genuine change responses while preserving the pair midpoint. Batch-Shared Chunk Selection (BSCS) further reduces feed-forward network (FFN) computation by retaining a batch-shared subset of channel chunks that can be executed as a compact dense FFN. A CGLA-Prior-Guided Refinement (CPGR) decoder reuses encoder-side change responses for coarse-to-fine prediction. Experiments on four remote sensing change detection benchmarks show that AdaDINO achieves competitive or superior performance against VFM-based baselines, with the largest gain on the category-agnostic SYSU-CD dataset. With 62.5% of the FFN hidden width removed, AdaDINO still achieves an F1 score of 85.29% on SYSU-CD while delivering a 1.41× throughput speedup. The code will be released.
Xu Zhang, Xinqing Li, Jianpeng Xie +3
VCIP, College of Computer Science, Nankai University · School of Computer Science and Engineering, Tianjin University of Technology · Academy for Advanced Interdisciplinary Studies, Nankai University +1
Change detection (CD) in remote sensing is vital for applications such as urban monitoring and disaster assessment, yet traditional methods struggle with generalization across diverse scenarios. We present OmniCD, a foundational framework that unifies and enhances remote sensing CD through multimodal semantic guidance. OmniCD incorporates image and text prompts -- such as textual descriptions, semantic maps, and geospatial metadata -- into a unified architecture, supporting tasks from binary CD to zero-shot semantic change understanding. The framework integrates a hierarchical scene retrieval module and a change detection module, reinforced by a style disentanglement mechanism for improved cross-domain robustness. We further introduce RSITCD, a large-scale multimodal dataset with 300K+ annotated image-text pairs. Extensive experiments show that OmniCD achieves state-of-the-art performance across benchmarks, demonstrating strong adaptability and setting a solid foundation for general-purpose CD systems in remote sensing.