A hybrid framework that decouples detection from damage assessment is proposed, combining the precision of CV models with the reasoning power of LVLMs, and the best combination under this framework accurately counts intact, partially damaged and completely destroyed buildings.
Abstract
Rapid and accurate post-disaster building damage assessment is essential, yet remains a challenging task. Unmanned Aerial Vehicle (UAV) imagery offers a timely and high-resolution view of affected areas, but existing Computer Vision (CV) models often demand large annotated datasets, generalize poorly across geographic regions and their assessment policies, and are confined to the specific tasks they were trained for. Large Vision-Language Models (LVLMs) offer a promising alternative through their strong reasoning and generalization capabilities, but fall short on precise, low-level perception tasks such as object detection and accurate bounding box generation. Furthermore, they often require a substantial amount of data for effective fine-tuning on domain-specific tasks. In this paper, we propose a hybrid framework that decouples detection from damage assessment, combining the precision of CV models with the reasoning power of LVLMs. A CV model first detects buildings and generates bounding boxes on the image that are then passed to an LVLM for damage classification and contextual interpretation. We evaluated our framework on two real-world benchmarks: RescueNet and FloodNet. In particular, the best combination under this framework accurately counts intact, partially damaged and completely destroyed buildings, surpassing isolated baselines by up to 2.1 R^2 points, while requiring only limited annotated data for the detection stage. Beyond reporting aggregate gains, we provide a detailed analysis of failure scenarios and edge cases, offering practical insights for practitioners and concrete directions for future work. Our source code and data are publicly available to the research community via the following repository: https://github.com/ungquanghuy-kddi/VLM_GDINO.git
Abstract. In the aftermath of a disaster, whether natural, industrial, or war-related, a rapid and accurate assessment of building damage is crucial for rescue forces to conduct an effective emergency response. Very high-resolution satellite imagery enables such assessments and serves as an important indicator for understanding the scale of destruction, supporting time-critical rescue operations, and guiding resource allocation. While deep learning models have shown promising results in automating building damage assessment (BDA) from pre- and post-disaster optical satellite imagery, they often fail to generalize to new disasters due to domain shifts. This paper studies the challenge of rapid domain adaptation for BDA in the context of the war in Ukraine. We create a new, challenging dataset annotated with damage grades across six cities in Ukraine, using pre- and post-disaster optical imagery. To facilitate rapid adaptation, we propose an efficient fine-tuning workflow using Low-Rank Adaptation. Our experiments show that this approach substantially improves performance in both out-of-domain and in-domain settings, presenting a practical and data-efficient study for deploying BDA models in time-critical emergency scenarios.
Sebastian Gapp, C. Henry, Pablo d'Angelo et al.· ISPRS Annals of the Photogra...· 0 citations
Abstract. In the immediate aftermath of a hurricane, the rapid, accurate assessment of building damage is paramount for effective emergency response and the allocation of resources. Traditional methods of damage assessment, which rely on ground-based surveys, are often slow, hazardous, and subjective. While the advent of remote sensing (RS), through Unmanned Aerial Vehicles (UAVs) and the application of Convolutional Neural Networks (CNNs), has significantly advanced the automation of this process, these models operate on a pixel-level or object-level basis, failing to capture the inherent spatial relationships and contextual information within a disaster zone. Damage patterns are not spatially random; they exhibit strong spatial autocorrelation, a principle encapsulated by Tobler’s First Law of Geography. This paper introduces a novel approach that leverages Graph Attention Networks (GATs) to explicitly model spatial dependencies when evaluating building damage. By representing damaged buildings and their surroundings as nodes and edges in a graph, our model can learn and weigh the influence of neighboring structures and the local environment when assessing their damage level. This spatially-aware methodology moves beyond simple image classification to a more holistic scene understanding. We evaluate the method on DoriaNET, a geo-referenced UAV dataset collected after Hurricane Dorian (2019) that provides masked building patches, GPS centroids, structural metadata, and ordinal FEMA/HAZUS-style damage labels. By incorporating spatial context via a graph-based framework, our GAT model achieves superior performance in building damage classification compared to state-of-the-art CNN-based approaches, producing more coherent and accurate damage maps better suited to real-world disaster management scenarios.
Fuad Hasan, C. Yeum, Ali Lesani et al.· ISPRS Annals of the Photogra...· 1 citation
Recently, deep learning has enabled unmanned aerial vehicles (UAVs) to detect human bodies in aerial imagery, which is of particular importance in post-disaster situations such as floods and storms. Yet progress in this domain remains constrained by a familiar obstacle: the shortage of annotated training data. Neural networks, while powerful, are highly sensitive to data volume and diversity. Existing augmentation strategies help reduce this gap but typically introduce only incremental novelty, especially with respect to viewpoint variation, thereby limiting dataset richness. In this work, we propose a complementary strategy that leverages three-dimensional human models reconstructed via photogrammetric techniques. By situating these models within a controlled rendering environment, we generate synthetic imagery across a broad range of elevations and camera angles—perspectives that are rarely captured in conventional UAV datasets. These additions are designed to increase both the variability and the resilience of the training corpus. To evaluate the contribution of this approach, a custom CNN deep convolutional neural classifier was trained and benchmarked on a UAV human vs. non-human patch dataset of 4000 baseline images (128 × 128 px; 2800 train, 600 validation, 600 test), expanded with 3000 photogrammetry-derived synthetic patches (balanced by class) to 7000 total images for the 3DG setting. The primary metric was classification accuracy on the held-out test set, consistent with patch-level evaluation practice; detection-style metrics such as AP/IoU were not applicable to this binary classification protocol. Averaged over five independent training runs, the proposed augmentation improved classification accuracy by 3.02 percentage points over the baseline (88.06 ± 0.97% → 91.08 ± 1.03%), with consistent gains in precision, recall, and F1-score. When combined with standard augmentations (rotation, translation, scaling, flipping), accuracy reached 95.21 ± 0.61%, a gain of 7.15 percentage points over the baseline. These results suggest that photogrammetry-based augmentation offers a practical and effective enhancement for UAV-based human detection pipelines where timely, reliable identification is critical.
M. Varshosaz, Kamyar Hassanpoor, Vahid Mousavi et al.· Remote Sensing· 0 citations
A novel lightweight Local-Global Interaction Network (LGINet) for efficient BDD, which achieves the best balance between accuracy and efficiency, outperforming existing methods.
Wei Li, Guorui Ma, Lunjun Fan et al.· The International Archives o...· 0 citations
One of the worst natural disasters, floods seriously harm infrastructure, economy, and society. The manual interpretation of satellite or aerial imagery is a major component of traditional post-disaster assessment methods, but is time-consuming, labor-intensive, and prone to human error. Automated and precise flood mapping systems that facilitate quick emergency response is becoming more necessary as deep-learning advances and high-resolution UAV datasets become available. This paper focuses on semantic segmentation of flood-affected regions from aerial images by developing a hybrid light weighted and semi-supervised deep learning framework. In this research, from floodNet dataset both unlabeled and labeled UAV imagery, with 10 semantic classes is used and a multi-class segmentation using UNet++ model with ResNet34 and EfficientNet-B4 as backbone encoders is performed. The model is initially trained using annotated images in a supervised learning manner and later is enhanced using a semi-supervised learning approach of pseudo labeling. This training approach is applied to leverage more than 1000 unlabeled images, improving the generalization gradually. With a mean Dice score of 0.84 and mean IoU score of 0.77 for the validation set, EfficientNetB4 performs better while ResNet34 is faster by 3ms of inference time. In effect, the results show that preprocessing and segmentation using proposed models improve further processing of the results in disaster management systems. In addition, pseudo-labeling not only improves performance of minority classes but also offers a scalable method, to manage unlabelled data in substantial amounts, in real-world deployments. Overall, this work testifies the potential of combining the robust encoder-decoder architectures, with semi-supervised learning approach in order to deliver a reliable and efficient preprocessing for automated flood assessment technique. It also provides insights by highlighting the potential for extending the model for preprocessing in real-world applications, such as damage quantification, risk prioritization, etc.
Voore Manaswini, R. K· 2026 International Conferenc...· 0 citations
This study explores AI-driven image classification to expedite damage evaluation by identifying damaged buildings from post-disaster photos much faster than conventional methods, providing a more detailed understanding of structural integrity across affected areas.
M. Kovačević, F. Đorđević, Đorđe Nedeljković et al.· Bulletin of Earthquake Engin...· 0 citations