Organizations: Department of Geography, National University of Singapore, Singapore, 117570, Singapore · Urban Governance and Design Thrust, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, 511453, Guangdong, China · School of Environment and Spatial Informatics, China University of Mining and Technology, Xuzhou, 221116, Jiangsu, China · Disaster Management Institute, Istanbul Technical University, Istanbul, 34469, Sariyer, Türkiye · Department of Complexity Science and Engineering, University of Tokyo, Chiba, 277-8561, Japan · School of Automation, Southeast University, Nanjing, 211189, Jiangsu, China
Living in a changing climate, human society now faces more frequent and severe natural disasters than ever before. As a consequence, rapid disaster response during the "Golden 72 Hours" of search and rescue becomes a vital humanitarian necessity and community concern. However, traditional disaster damage surveys routinely fail to generalize across distinct urban morphologies and new disaster events. Effective damage mapping typically requires exhaustive and time-consuming manual data annotation. To address this issue, we introduce Smart Transfer, a novel Geospatial Artificial Intelligence (GeoAI) framework, leveraging state-of-the-art vision Foundation Models (FMs) for rapid building damage mapping with post-earthquake Very High Resolution (VHR) imagery. Specifically, we design two novel model transfer strategies: first, Pixel-wise Clustering (PC), ensuring robust prototype-level global feature alignment; second, a Distance-Penalized Triplet (DPT), integrating patch-level spatial autocorrelation patterns by assigning stronger penalties to semantically inconsistent yet spatially adjacent patches. Extensive experiments and ablations from the recent 2023 Turkiye-Syria earthquake show promising performance in multiple cross-region transfer settings, namely Leave One Domain Out (LODO) and Specific Source Domain Combination (SSDC). Moreover, Smart Transfer provides a scalable, automated GeoAI solution to accelerate building damage mapping and support rapid disaster response, offering new opportunities to enhance disaster resilience in climate-vulnerable regions and communities. The data and code are publicly available at https://github.com/ai4city-hkust/SmartTransfer.
Figures & tables
Figure 1: A conceptual diagram of post-earthquake damage mapping, illustrating the Smart Transfer framework’s objective to minimize the rapid response time window for critical disaster response operations.
Figure 2: Study area overview displaying the USGS Modified Mercalli Intensity (MMI) shakemap. Post-disaster VHR imagery is provided for the 9 selected urban regions.
Region
Area (km 2 )
Total Tiles
Total Buildings
Labeled Tiles
Labeled Buildings
Gaziantep
123.03
30,155
173,774
42
45
Hatay
105.30
25,781
47,752
210
258
Kahramanmaras
27.41
18,020
40,419
414
419
Kirikhan
12.18
3,016
11,958
53
51
Nurdagi
15.21
10,032
4,076
225
176
Sakcagozu
4.50
3,024
1,196
74
79
Table 1: Dataset statistics across the 9 selected study regions, including the distribution of total VHR image tiles, building footprints, and ground truth annotations.
Figure 3: Selected examples of multiple building damage levels from 9 study regions, from top to bottom row refer to slight damage, heavy damage, buildings requiring demolition, and collapse building, respectively.
Figure 4: Post-disaster remote sensing imagery of part of Kahramanmaras (a) post-disaster VHR imagery and (b) with annotated damaged buildings (red) and undamaged buildings (green). Color intensity indicates the damage score of the ground truth annotation.
Figure 5: Three Stages of Smart Transfer. Stage 1 is the cold start of foundation models, Stage 2 is the warm-up of foundation models, and Stage 3 is the smart transfer strategy designed for diverse damage mapping scenarios.
Figure 6: Overall network architecture of Smart Transfer, consisting mainly of three parts. (1) the Freeze Vision FM Encoder, (2) Smart Transfer Loss Functions, (3) the Trainable Damage Mapping Decoder.
Model
FMs
Precision
Recall
Accuracy
F1
IoU 0
IoU 1
mIoU
ResNet-18
✗
0.14 ±
0.06
0.68 ±
0.07
0.44 ±
0.03
0.22 ±
0.08
0.38 ±
0.05
0.12 ±
0.05
0.25 ±
0.02
ResNet-152
✗
0.22 ±
0.09
0.57 ±
0.10
0.70 ±
0.02
0.30 ±
0.10
0.68 ±
0.04
0.18 ±
0.07
0.43 ±
0.02
YOLO-like
✗
0.23 ±
0.10
0.55 ±
0.15
0.73 ±
0.06
0.31 ±
0.11
0.71 ±
0.07
0.19 ±
0.08
0.45 ±
0.03
DINOv3
✓
0.55 ±
0.09
0.75 ±
0.08
0.90 ±
0.02
0.62 ±
0.09
0.89 ±
0.03
0.46 ±
0.09
0.68 ±
0.04
Table 2: Full-supervision overall performance for building damage assessment, evaluating both classification and segmentation metrics (from five stratified folds). FMs indicate whether the model is a Foundation Model ( ✓ ) or not ( ✗ ).
mIoU
F1 Score
DINO
ResNet-18
ResNet-152
YOLO-like
DINO
ResNet-18
ResNet-152
YOLO-like
FMs
✓
✗
✗
✗
✓
✗
✗
✗
Gaziantep
0.61 ±
0.05
0.25 ±
0.06
0.41 ±
0.05
0.45 ±
0.06
0.47 ±
0.08
0.11 ±
0.03
0.16 ±
0.04
0.19 ±
0.08
Hatay
0.63 ±
0.03
0.26 ±
0.07
0.41 ±
0.06
0.44 ±
0.03
0.55 ±
0.05
0.17 ±
0.01
0.24 ±
0.04
0.21 ±
0.02
Kahramanmaras
0.72 ±
0.02
0.27 ±
0.08
0.47 ±
0.08
0.42 ±
0.02
0.74 ±
0.02
0.35 ±
0.03
0.48 ±
0.07
0.44 ±
0.03
Kirikhan
0.68 ±
0.03
0.27 ±
0.08
0.42 ±
0.06
0.44 ±
0.05
0.59 ±
0.04
0.14 ±
0.05
0.20 ±
0.04
0.18 ±
0.08
Table 3: Region-wise results for mIoU and F1 scores under full supervision (from five stratified folds). FMs indicate whether the model is a Foundation Model ( ✓ ) or not ( ✗ ).
mIoU
F1 Score
Target Domain
FS
PC
DPT
PC+DPT
FS
PC
DPT
PC+DPT
Gaziantep
0.61 ±
0.05
0.65 ±
0.04
0.65 ±
0.03
0.66 ±
0.03
0.47 ±
0.08
0.51 ±
0.07
0.52 ±
0.06
0.53 ±
0.05
Hatay
0.63 ±
0.03
0.64 ±
0.03
0.63 ±
0.05
0.63 ±
0.03
0.55 ±
0.05
0.56 ±
0.05
0.55 ±
0.07
0.55 ±
0.05
Kahramanmaras
0.72 ±
0.02
0.69 ±
0.03
0.68 ±
0.05
0.67 ±
0.03
0.74 ±
0.02
0.67 ±
0.05
0.64 ±
0.07
0.63 ±
0.06
Kirikhan
0.68 ±
0.03
0.70 ±
0.07
0.71 ±
0.04
0.70 ±
0.07
0.59 ±
0.04
0.62 ±
0.11
0.64 ±
0.05
0.62 ±
0.11
Nurdagi
0.68 ±
0.02
0.71 ±
0.02
0.70 ±
0.03
0.70 ±
0.03
0.65 ±
0.03
0.69 ±
0.03
0.68 ±
0.03
0.68 ±
0.04
Table 4: Performance w.r.t mIoU and F1 scores for building damage assessment under the LODO setting (from five stratified folds). FS refers to the fully supervised DINOv3 model, which serves as a competitive baseline.
Figure 7: Visualization of the regional-wise Class Activation Maps (CAMs) of the fully supervised baselines compared to different smart transfer variants.
mIoU
F1 Score
Source Domain
FS
PC
DPT
PC+DPT
FS
PC
DPT
PC+DPT
Kah
0.68 ± 0.04
0.68 ± 0.06
0.68 ± 0.06
0.67 ± 0.06
0.62 ± 0.09
0.62 ± 0.11
0.63 ± 0.11
0.63 ± 0.11
Nur
0.68 ± 0.04
0.68 ± 0.05
0.68 ± 0.04
0.68 ± 0.04
0.62 ± 0.09
0.59 ± 0.09
0.61 ± 0.07
0.60 ± 0.08
Kah+Nur
0.68 ± 0.04
0.71 ± 0.04
0.70 ± 0.05
0.71 ± 0.04
0.62 ± 0.09
0.66 ± 0.09
0.64 ± 0.10
0.65 ± 0.09
Kah+Hat
0.68 ± 0.04
0.70 ± 0.04
0.71 ± 0.04
0.70 ± 0.04
0.62 ± 0.09
0.65 ± 0.09
0.66 ± 0.08
0.64 ± 0.09
Nur+Hat
0.68 ± 0.04
0.70 ± 0.03
0.69 ± 0.03
0.69 ± 0.03
0.62 ± 0.09
0.63 ± 0.06
0.62 ± 0.06
0.62 ± 0.07
Table 5: Performance w.r.t mIoU and F1 scores for building damage assessment under the SSDC setting (from five stratified folds). FS refers to the fully supervised DINOv3 model, which serves as a competitive baseline.
Figure 8: Influence of different train ratios on model performances (w.r.t mIoU and F1 scores) across the 6 selected source domains under the SSDC setting.
Metric
LODO
SSDC
Raw
PC
DPT
PC+DPT
Raw
PC
DPT
PC+DPT
Smart Transfer
✗
✓
✓
✓
✗
✓
✓
✓
Precision
0.96
0.91
0.95
0.94
0.94
0.95
0.93
0.92
Recall
0.22
0.30
0.24
0.19
0.20
0.17
0.18
0.16
Accuracy
0.74
0.76
0.74
0.72
0.73
0.72
0.72
0.71
F1
0.36
0.46
0.38
0.32
0.33
0.29
0.31
0.28
Table 6: Performance comparison of post-disaster building damage classification in Kahramanmaras under both LODO and SSDC settings. Raw refer to the non-adapted transfer.
Figure 9: Post-disaster Building Damage Classification for a part of Kahramanmaras. (a) Raw, (b) PC, (c) DPT, (d) PC+DPT.
Setting
Shot
Finetune
Raw
PC
DPT
PC+DPT
LODO
0-shot
✗
0.6851 ± 0.0370
0.6932 ± 0.0325
0.6876 ± 0.0345
0.6885 ± 0.0332
1-shot
✓
0.6865 ± 0.0358
0.6944 ± 0.0320
0.6891 ± 0.0335
0.6897 ± 0.0323
3-shot
✓
0.6867 ± 0.0364
0.6946 ± 0.0327
0.6892 ± 0.0344
0.6900 ± 0.0331
5-shot
✓
0.6868 ± 0.0361
0.6949 ± 0.0328
0.6895 ± 0.0341
0.6902 ± 0.0331
10-shot
✓
0.6878 ± 0.0358
0.6916 ± 0.0399
0.6902 ± 0.0338
0.6910 ± 0.0328
SSDC
0-shot
✗
0.6924 ± 0.0131
0.6933 ± 0.0115
0.6942 ± 0.0125
0.6933 ± 0.0094
Table 7: Few-shot learning results (mIoU) under both LODO and SSDC, (from five stratified folds).
Figure 10: Tradeoff between PC and DPT considering different weighting strategies. Herein, λpc and λdpt refer to the loss function weights of PC and DPT, respectively.
College of Geodesy and Geomatics, Shandong University of Science and Technology, Qingdao 266590, China · Professorship of Big Geospatial Data Management, School of Engineering and Design, Technical University of Munich, Munich, 85521, Bavaria, Germany · State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan 430079, China +6