Organizations: Department of Ecology and Conservation Biology, Texas A&M University, College Station, 77843, Texas, USA · Department of Computer Science, Stanford University, Stanford, 94305, CA, USA · Department of Mechanical Engineering, National University of Singapore, Singapore, 117575, Singapore
Geospatial foundation models (GFMs) are reshaping remote sensing by shifting the field from task-specific modeling toward general-purpose geospatial intelligence. Mapping the damage to buildings stands to benefit substantially from this shift, as GFMs pretrained on large-scale Earth observation data provide generalizable geospatial representations that can facilitate effective responses to sudden-onset disasters. Although parameter-efficient fine-tuning (PEFT) facilitates GFM adaptation, parameter efficiency does not necessarily translate into label efficiency. Building damage mapping remains challenging due to scarce pre-existing labeled data or limited time for annotation. When weak supervision is introduced to alleviate the labeling bottleneck, we identify the difficulty of learning sufficiently discriminative representations as a key challenge for label-efficient GFM adaptation. To address this challenge, we propose a Positive-Unlabeled GFM Adaptation (PUMA) framework that requires annotations for only a small number of damaged buildings. More specifically, PUMA integrates non-contrastive weakly supervised representation refinement to mitigate the impact of noisy supervision on representation learning from a gradient perspective, and latent-state-based label disambiguation to enable reversible supervision refinement without requiring class prior estimation. Extensive experiments on disaster benchmark datasets from around the world demonstrated the effectiveness and robustness of PUMA. A real-world application to the 2025 Los Angeles wildfires further highlighted its potential to accurately map damage under highly limited supervision. More broadly, these findings suggest that weakly supervised PEFT offers a label-efficient pathway from GFMs to real-world applications.
Figures & tables
Figure 1: t-SNE ( van der Maaten and Hinton, 2008 ) visualization of learned representations, representation distributions estimated using kernel density estimation (KDE), and decision boundaries produced by different methods on the hurricane-florence subset of the xBD dataset. VPT-deep was used as the PEFT strategy for all methods.
Figure 2: Spatial distribution of all disaster events considered in this paper.
Name
Disaster event
Event date
Number of images
Used for building damage classification and building localization evaluation
hurricane-florence
Hurricane Florence
Sep 10-19, 2018
1092
joplin-tornado
Joplin, MO Tornado
May 22, 2011
298
santa-rosa-wildfire
Santa Rosa Wildfires
Oct 8-31, 2017
754
Used only for training the building localization model
guatemala-volcano
Guatemala Fuego Volcano Eruption
Jun 3, 2018
56
Table 1: Summary statistics of the xBD dataset.
Figure 3: Overview of the proposed positive-unlabeled GFM adaptation framework.
Method
Tsunami
Hurricane
Tornado
Wildfire
OA
F1
OA
F1
OA
F1
OA
F1
Training-from-scratch methods
nnPU
78.48 ±0.55
81.84 ±0.38
92.76 ±0.52
82.57 ±1.13
95.60 ±0.50
92.87 ±0.74
93.92 ±0.06
87.27 ±0.18
vPU
65.82 ±5.39
48.77 ±12.71
84.62 ±1.91
27.85 ±14.66
72.42 ±1.75
13.79 ±10.01
80.65 ±2.43
27.22 ±15.78
HolisticPU
80.67 ±1.89
78.50 ±2.85
51.90 ±30.25
43.68 ±16.45
82.16 ±18.10
77.98 ±16.52
56.19 ±33.00
54.72 ±28.19
WSC
86.41 ±0.95
85.83 ±1.21
92.59 ±1.23
81.97 ±2.20
95.69 ±0.72
92.92 ±1.02
95.05 ±0.26
89.78 ±0.37
Table 2: Quantitative comparison of different methods. The best result for each metric is highlighted in red.
Figure 4: Visual comparison of different methods on the Hurricane Florence subset of the xBD dataset. Damaged and undamaged buildings are shown in red and green , respectively.
Figure 5: Visual comparison of different methods on the Joplin Tornado subset of the xBD dataset. Damaged and undamaged buildings are shown in red and green , respectively.
Representation
Learning
Pseudo-label
Updating
Tsunami
Hurricane
OA
F1
P
R
OA
F1
P
R
Lrepself
68.70 ±1.90
54.83 ±3.99
97.66 ±0.83
38.18 ±3.76
91.39 ±0.65
70.24 ±3.00
96.87 ±0.24
55.16 ±3.80
Lrepsup
65.32 ±8.56
45.60 ±18.94
97.78 ±0.25
31.22 ±17.58
92.72 ±1.03
76.10 ±4.53
96.18 ±1.05
63.17 ±6.44
Lrepsqrt(k=1)
67.14 ±4.50
51.02 ±10.06
97.73 ±0.67
34.92 ±8.99
93.53 ±0.13
79.36 ±0.35
96.58 ±0.75
67.35 ±0.15
Table 3: Ablation study of the proposed modules. The parameter k in Lrepsqrt was set to 9 by default.
Method
OA
F1
P
R
uPU
83.48 ±2.31
81.73 ±3.15
90.88 ±1.86
74.40 ±5.32
uPU+ Lrepsqrt
89.34 ±0.31
89.11 ±0.58
90.85 ±1.78
87.52 ±2.78
nnPU
88.18 ±1.82
87.48 ±2.07
92.67 ±1.04
82.87 ±3.12
nnPU+ Lrepsqrt
90.03 ±0.24
90.10 ±0.46
89.25 ±1.54
91.04 ±2.50
Table 4: Effect of representation refinement loss on existing PU learning methods.
Method
Tsunami
Hurricane
OA
F1
OA
F1
Full fine-tuning
50.45 ±0.95
66.44 ±0.23
46.64 ±39.64
48.14 ±28.41
Linear probing
90.80 ±0.26
91.12 ±0.26
94.57 ±0.75
86.53 ±1.56
PEFT
BitFit
92.92 ±0.84
93.07 ±0.73
97.34 ±0.07
92.86 ±0.29
VPT-shallow
90.83 ±0.70
91.11 ±0.57
95.81 ±0.15
89.35 ±0.41
VPT-deep
93.63 ±0.32
93.71 ±0.33
97.32 ±0.60
92.73 ±1.80
Table 5: Performance of PUMA with different PEFT strategies.
Pre-trained
model
Backbone
Tsunami
Hurricane
OA
F1
OA
F1
scratch
ViT-B
80.98 ±2.25
83.34 ±1.44
91.20 ±1.90
80.01 ±3.22
CLIP
ResNet-50
91.18 ±0.09
91.39 ±0.37
96.95 ±0.69
92.07 ±1.57
ViT-B
93.79 ±0.57
93.85 ±0.49
97.53 ±0.30
93.26 ±0.81
LCLIP
ViT-B
93.76 ±0.75
93.83 ±0.69
97.46 ±0.29
93.12 ±0.86
Table 6: Performance of PUMA with different GFMs.
Disaster
Building Localization
Building Damage Classification
Event
AP Box
AP Mask
IoU
P Pixel
R Pixel
F1 Pixel
P
R
F1
Hurricane
56.7 ±0.2
52.2 ±0.1
77.8 ±0.6
79.75 ±0.26
88.97 ±0.26
84.11 ±0.26
75.89 ±0.34
84.63 ±0.38
80.02 ±0.36
Tornado
52.1 ±0.4
48.2 ±0.2
83.0 ±0.4
84.26 ±0.14
91.99 ±0.10
87.96 ±0.12
83.38 ±0.25
76.88 ±0.23
80.00 ±0.24
Wildfire
52.1 ±0.3
47.3 ±0.2
72.9 ±2.5
80.58 ±0.16
89.71 ±0.12
84.90 ±0.15
80.61 ±0.08
84.09 ±0.08
82.32 ±0.08
Table 7: Building localization and damage classification performance using automatically extracted building footprints.
Figure 6: Hyperparameter analyses of PUMA for the tsunami and hurricane events.
Figure 7: Building damage mapping in the Eaton subarea of the Los Angeles wildfires. The left column shows pre-disaster imagery with building footprints, and the right column shows post-disaster imagery with the corresponding damage mapping results. Damaged and undamaged buildings are shown in red and green , respectively.
Method
Eaton Subarea
Palisades Subarea
OA
F1
P
R
OA
F1
P
R
∣Dp∣=50 damaged buildings
Naive Supervision
94.03 ±0.76
76.41 ±3.69
90.11 ±0.92
66.40 ±5.02
50.63 ±2.23
38.46 ±4.26
98.70 ±0.93
23.93 ±3.28
Standard Supervision
91.73 ±0.74
77.54 ±1.47
64.34 ±2.20
97.63 ±0.75
92.87 ±1.08
94.26 ±0.93
98.26 ±0.29
90.58 ±1.80
PUMA
97.81 ±0.28
92.81 ±0.86
89.07 ±1.53
96.90 ±1.03
92.48 ±1.47
93.89 ±1.24
98.79 ±0.46
89.46 ±1.88
Full Supervision
97.60 ±0.62
92.30 ±1.85
87.16 ±3.12
98.11 ±0.34
97.44 ±0.04
98.01 ±0.03
98.26 ±0.42
97.77 ±0.44
Table 8: Building damage mapping results for the Eaton and Palisades subareas during the 2025 Los Angeles wildfires. Naive Supervision treats all unlabeled samples as negative; Standard Supervision additionally incorporates labeled undamaged buildings in equal number to the labeled damaged buildings; and Full Supervision assigns ground-truth labels to all unlabeled samples.
Figure 8: Building damage mapping in the Palisades subarea of the Los Angeles wildfires. The left column shows pre-disaster imagery with building footprints, and the right column shows post-disaster imagery with the corresponding damage mapping results. Damaged and undamaged buildings are shown in red and green , respectively.
Figure 9: Label efficiency of PUMA compared with supervised PEFT methods in the Eaton subarea of the Los Angeles wildfires. A fixed set of 4,000 unlabeled buildings is used in all experiments, and 1× corresponds to 100 annotated samples. (a) F1 scores under varying numbers of labeled damaged buildings. PUMA maintains performance close to Full Supervision and consistently outperforms Standard Supervision across annotation budgets. (b) PUMA requires only 1/40 of the annotated samples used by Standard Supervision.
Wildfire burned-area mapping is essential for damage assessment, emissions modeling, and understanding fire-climate interactions across diverse ecological regions. Recent geospatial foundation models provide strong general-purpose representations for satellite imagery, yet there is still no clear understanding of how to efficiently adapt these models for downstream Earth observation tasks, particularly under geographic and temporal domain shift. This study evaluates three state-of-the-art Geospatial Foundation Models (GFMs) - Terramind, DINOv3, and Prithvi-v2 - for burned-area mapping across the United States and Canada using Sentinel-2 data. Leveraging 3,820 wildfire events from 2017-2023, we conduct spatial and temporal generalization tests across diverse biomes. We systematically compare full fine-tuning, decoder-only fine-tuning, and Low-Rank Adaptation (LoRA) for adapting each model. Across all experiments, LoRA provides the strongest cross-domain generalization while updating less than 1% of parameters, demonstrating a favorable trade-off between accuracy and efficiency. Prithvi-v2 with LoRA achieves the highest overall accuracy and the largest improvement compared to full fine-tuning. These findings indicate that geospatial foundation models, when adapted using lightweight parameter-efficient methods such as LoRA, offer a robust and scalable solution for large-scale burned-area mapping. Code is available at https://github.com/alishibli97/wildfire-lora-gfm.
Ali Shibli, Andrea Nascetti, Yifang Ban
KTH Royal Institute of Technology Stockholm, Sweden
Geospatial foundation models can support disaster damage assessment, but high accuracy on familiar disasters may hide shortcut learning: prediction from event or scene characteristics rather than evidence that transfers to a new disaster. We study embeddings from TerraMind, Prithvi-EO-2.0, DOFA, and DINOv3 on the multi-disaster datasets xView2 and BRIGHT. Holding out an entire disaster, instead of randomly splitting image tiles, reduces mean balanced accuracy from 0.807 to 0.638 on xView2 and from 0.818 to 0.568 on BRIGHT. Four sets of tests examine what information supports prediction: event and acquisition context, before/after image inputs, building count, and correct image pairing. Pre-event images remain predictive, and a classifier given only the annotated number of buildings reaches balanced accuracy of 0.629 and 0.653. Combining before/after features performs best on average on xView2, but does not improve over pre-only features on BRIGHT. Shuffling image pairs within the same disaster and damage class causes modest losses. Motivated by these findings, we test whether feature transformations learned only from training disasters improve generalization to unseen disasters. These transformations improve performance, with supervised contrastive learning performing best on xView2 and PCA on BRIGHT. Our results highlight the importance of evaluation on unseen domains for assessing model robustness and show that simple feature-level transformations can improve generalization. Code and experiments can be found at https://github.com/alishibli97/gfm-shortcut-learning.
Ali Shibli, Andrea Nascetti, Yifang Ban
Division of Geoinformatics KTH Royal Institute of Technology Stockholm, Sweden
Geospatial foundation models (GFMs) have been proposed as generalizable backbones for disaster response, land-cover mapping, food-security monitoring, and other high-stakes Earth-observation tasks. Yet the published work about these models does not give reviewers or users enough information to tell which model fits a given task. We argue that nobody knows what the current state of the art is in geospatial foundation models. The methods may be useful, but the GFM literature does not standardize evaluations, training and testing protocols, released weights, or pretraining controls well enough for anyone to compare or rank them. In a 152-paper audit, we find 47 cross-paper disagreements of at least 5 percentage points, including 18 of at least 10 points, within model, benchmark, and protocol groups; 94/126 papers with pretraining data use a configuration no other paper uses; and 39% of GFM papers release no model weights. This lack of community standards can be solved. We propose six concrete expectations: named-license weight release, shared core evaluations, copied-versus-rerun baseline annotations, variance reporting, one shared evaluation harness, and data-vs-architecture-vs-algorithm controls. These gaps are a coordination failure, not a fault of any individual lab; the authors of this paper, like many others in the GFM community, have contributed to them. Rather than just critiquing the community, we aim to provide concrete steps toward a shared understanding of how to evaluate and improve GFMs.
Isaac Corley, Nils Lehmann, Caleb Robinson +6
Taylor Geospatial · Technical University of Munich · Microsoft AI for Good Research Lab +6