DamageScope: An Integrated Pipeline for Building Damage Segmentation, Geospatial Mapping, and Interactive Web-Based Visualization

Abstract

Effective post-disaster damage assessment is crucial for guiding emergency response and resource allocation. This study introduces DamageScope, an integrated deep learning framework designed to detect and classify building damage levels from post-disaster satellite imagery. The proposed system leverages a convolutional neural network trained exclusively on post-event data to segment building footprints and assign them to one of four standardized damage categories: no damage, minor damage, major damage, and destroyed. The model achieves an average F1 score of 0.598 across all damage classes on the test dataset. To support geospatial analysis, the framework extracts the coordinates of damaged structures using embedded metadata, enabling rapid and precise mapping. These results are subsequently visualized through an interactive, web-based platform that facilitates spatial exploration of damage severity. By integrating classification, geolocation, and visualization, DamageScope provides a scalable and operationally relevant tool for disaster management agencies seeking to enhance situational awareness and expedite post-disaster decision making.

1. Introduction

Natural disasters pose profound threats to human safety, infrastructure integrity, and socioeconomic stability. Each year, communities worldwide experience numerous catastrophic events that disrupt societal functions, induce significant economic losses, and result in long-term psychological impacts. Recent trends underscore a concerning rise in both the frequency and intensity of such events, influenced, in part, by population growth, urbanization, and climate variability. In particular, severe weather-related disasters such as tornadoes and hurricanes have become increasingly prevalent. For instance, in 2024, the United States documented approximately 1735 confirmed tornadoes, marking an alarming 45% increase from the 1197 tornadoes recorded the previous year [1,2]. Economic losses associated with natural disasters are correspondingly escalating. According to the National Oceanic and Atmospheric Administration (NOAA), in 2024 alone, the U.S. faced 27 weather and climate disasters, each exceeding USD 1 billion in economic damage. These events encompassed severe storms, flooding, hurricanes, droughts, wildfires, and winter storms, resulting in a cumulative loss of hundreds of lives and billions of dollars [3]. Such statistics highlight the critical need for effective disaster preparedness strategies, rapid response frameworks, and robust recovery planning to mitigate these severe impacts on communities.

The effective disaster management cycle comprises multiple phases, including preparation, response, recovery, and mitigation [4]. Although proactive strategies such as early warning systems, evacuation planning, and resilient infrastructure development play essential roles, the efficacy of immediate post-disaster responses is equally critical. Rapid identification and assessment of structural damage are vital components in this phase, directly influencing the efficiency of search and rescue operations and the allocation of critical resources. Previous research demonstrates that survival probabilities in disaster scenarios decline significantly with delays in rescue operations, underscoring the necessity of swift and accurate damage assessments immediately following disasters [5]. Global Navigation Satellite System (GNSS) technology can support the disaster management cycle from the preparedness stage through mitigation. Multi-constellation GNSS receivers, leveraging signals from GPS, Galileo, BeiDou, and GLONASS, now provide centimeter-grade, real-time positioning, which allows responders to geotag imagery, sensor readings, and eyewitness reports with far greater precision [6,7]. Integrated with low-altitude UAV platforms, this capability links on-site reconnaissance to broader remote sensing workflows, enabling the rapid production of detailed, site-specific damage maps within hours of impact [8,9]. However, altitude limits, battery constraints, and adverse weather still restrict UAV coverage in wide-area events, driving a growing reliance on high-resolution satellite imagery for post-disaster assessment [10].

Satellite imagery offers distinct advantages over traditional ground-based or aerial methods, such as rapid, large-scale coverage, accessibility of remote or hazardous locations, and consistent temporal and spatial resolutions. Such imagery enables comprehensive evaluations of extensive geographic areas affected by disasters, providing crucial data for decision-making processes in emergency management. However, the analysis of satellite imagery poses unique challenges, including large data volumes, computationally intensive processing, varying spatial resolutions, and image complexity. Furthermore, traditional satellite-based damage assessment methods frequently require comparative analyses of images acquired before and after disaster events, necessitating pre-disaster datasets that are often unavailable or incomplete due to the inherently unpredictable nature of disasters.

Recent advances in artificial intelligence, especially deep learning methodologies, offer promising solutions for addressing these analytical challenges. Deep learning models, particularly convolutional neural networks (CNNs) and transformer-based architectures, have demonstrated remarkable capabilities in interpreting complex image data, performing tasks such as semantic segmentation [11] and object detection [12] with high accuracy and speed. While many studies have utilized deep learning approaches requiring paired pre- and post-disaster images, this dependency significantly limits practical applicability, especially in unpredictable disaster scenarios, where pre-event data may not exist or be accessible. Therefore, the development of methodologies capable of effectively analyzing post-disaster imagery independently remains an essential and ongoing research challenge.

Although deep learning techniques for building damage assessment from satellite imagery have achieved notable progress, their practical integration into real-time disaster response operations remains limited. To address these limitations, this research introduces an integrated framework, DamageScope, designed specifically for rapid, accurate, and scalable damage assessment using only post-disaster satellite imagery. The main contributions of this work are summarized as follows:

First, we propose an end-to-end framework for post-disaster building damage assessment, combining deep learning-based segmentation, geolocation extraction, and web-based visualization. The system integrates damage classification using a transformer-based encoder, geographic coordinate derivation for detected structures, and an interactive mapping platform for visualizing results. This unified workflow enables efficient, accurate, and accessible structural damage assessment suitable for rapid deployment during disaster response efforts.

Second, we specifically validate the use of a transformer-enhanced encoder, trained exclusively on post-disaster satellite imagery for semantic segmentation tasks. The study demonstrates that transformer-based architectures, when properly adapted, can deliver robust performance without reliance on pre-disaster reference images. This evaluation provides practical insights for developing post-event-only models that are more adaptable to operational emergency management contexts.

Third, we develop and deploy a lightweight, cloud-based web visualization platform that transforms segmentation outputs into an intuitive, interactive damage map using geolocation data derived from satellite imagery. Built using open-source technologies (Leaflet and Nginx) and hosted on AWS, the platform supports real-time accessibility across devices, facilitating rapid situational awareness for emergency responders.

2. Background and Literature Review

Over the past decade, research on satellite-based building damage assessment has accelerated, driven by both the increasing number of natural disasters and the wide adoption of deep learning techniques. This rapid growth reflects three converging trends. First, extreme earthquakes, floods, hurricanes, wildfires, and other disasters continue to grow in frequency and economic impact, placing greater emphasis on rapid, objective evaluation of building damage. Second, the availability of high-resolution satellite imagery and open archives has increased, enabling global observation of affected areas. Third, advances in artificial techniques like machine learning and deep learning, particularly convolutional neural networks (CNNs), transformers, semantic segmentation, and related models, have demonstrated greater capability to interpret raw pixels into actionable post-event intelligence.

Guided by recent research trajectories, we explored studies that employ modern deep-learning architectures, such as convolutional neural networks, vision transformers, and hybrid encoder–decoder models, to automate building damage assessment in post-disaster satellite imagery. The literature review opens with a concise technical overview of satellite sensor platforms and remote sensing modalities. Afterwards, the discussion moves from manual interpretation of satellite imagery, through feature-based machine learning classifiers, and culminates in modern end-to-end deep learning pipelines for semantic segmentation and damage assessment.

2.1. Overview of Remote Sensing and Satellite Technology

Remote sensing is defined as the practice of gathering information about objects or phenomena from a distance, without direct physical interaction, typically by detecting reflected or emitted electromagnetic radiation [13]. Historically, remote sensing originated with aerial photography in the 19th century, initially utilized for military reconnaissance [14]. Satellite remote sensing has advanced rapidly with the launch of space-borne platforms that deliver uniform, global coverage. Onboard sensors measure radiation reflected, emitted, or absorbed by the Earth across visible, infrared, and microwave bands, translating these electromagnetic interactions into geospatial data. In large-scale hazards such as earthquakes, hurricanes, and wildfires, satellites supply rapid revisit times and imagery of regions too dangerous or inaccessible for ground teams, sharpening situational awareness and operational planning. Their synoptic vantage point supports swift damage delineation, guiding resource allocation and rescue priorities, while scheduled overpasses build multi-temporal stacks for tracking evolving impacts. Normally, sensors used in satellite remote sensing fall into two categories: passive sensors, which record sunlight reflected or thermal energy emitted by the Earth, and active sensors, such as SAR or LiDAR, which emit their own signals and analyze the returned energy to characterize surface properties. Some commonly used passive sensors are infrared, multispectral, and hyperspectral sensors. These distinct sensor types serve diverse disciplines, including disaster response, environmental management, urban planning, and agriculture. Although multispectral images have long been used by researchers to collect various information, hyperspectral imaging extends observations across numerous visible and infrared bands, surpassing multispectral imaging in quality and detail [15]. Hyperspectral images capture detailed spectral signatures that allow for the precise identification of surface materials and conditions.

One of the limitations of passive sensors is that they cannot penetrate cloud cover, and visible-spectrum instruments operate only under daylight conditions [16,17]. By contrast, active sensors generate and transmit their own signals and measure the reflected returns. For example, radar satellites emit microwave pulses and record the backscatter, enabling observations independent of sunlight and often even through clouds. Such active sensors are indispensable for mapping flood inundation during hurricanes or monsoon rains, when optical imagery is often obstructed and rapid assessment of water extent is critical. To map topography and monitor changes in sea level or ice sheet thickness, satellite altimetric sensors are used [18]. These sensors measure surface elevation by timing the return of transmitted signals. Gravimetric remote sensing missions detect spatial variations in Earth’s gravitational field caused by changes in mass distribution; for instance, the GRACE satellite pair measures tiny fluctuations in inter-satellite separation, enabling the estimation of groundwater depletion and glacial mass loss [19]. The data gathered from these sensors are transmitted to ground stations, where sophisticated processing transforms them into actionable imagery and analytical products, critical for informed decision making in various sectors, particularly disaster management.

However, satellite-based remote sensing presents several inherent challenges. Spatial resolution limitations—ranging from tens of centimeters to several meters per pixel—may impede the detailed assessment of subtle structural damages. Atmospheric conditions such as clouds, smoke, fog, and haze frequently obstruct optical satellite sensors, complicating image collection and analysis [20]. Although radar-based satellite technologies can mitigate atmospheric interference through all-weather observation capabilities, not all satellite platforms possess radar sensors, thus limiting universal applicability. Furthermore, accurate geolocation and alignment of multi-temporal imagery require computationally intensive processing methods and specialized expertise, posing significant logistical challenges, particularly in resource-constrained environments [21]. Despite these challenges, ongoing advancements in sensor technologies, computational infrastructure, and artificial intelligence-driven analytical methods continue to improve satellite remote sensing capabilities. Innovations in these areas promise to overcome existing limitations, enhancing the reliability, accuracy, and speed of satellite-based methods in disaster response and recovery planning.

2.2. Identification of Damaged Structures from Satellite Images

The identification and classification of structural damage from satellite imagery have become critical components of disaster response and recovery planning. Historically, disaster response agencies have relied heavily on manual interpretation of satellite images to detect and assess building damages, a practice evident in major disaster events such as the 2010 Haiti earthquake [21]. However, manual interpretation of satellite images is highly time consuming, labor intensive, and often impractical for rapid, large-scale assessment immediately following disasters. To overcome these limitations, researchers have increasingly turned toward automated computational methods, notably machine learning and deep learning algorithms. The subsequent paragraphs discuss different machine learning and deep learning approaches developed over the years to identify the damaged structures using satellite imagery.

2.2.1. Early Feature-Based Approaches

Early efforts typically employed traditional machine learning methods, such as support vector machines (SVMs), decision trees, logistic regression, and simple artificial neural networks, utilizing hand-crafted image features including spectral indices, texture measures, and edge densities [22]. Object-based image analysis (OBIA) was also common, particularly in combining multisource features such as optical, LiDAR, and SAR data [23,24]. These early approaches provided foundational insight into automated damage assessment but required extensive manual feature engineering and expert tuning, limiting scalability and generalization [25].

2.2.2. Advancement of Convolutional Neural Networks (CNNs)

With significant improvements in computational resources and the rise of convolutional neural networks (CNNs), more advanced methods have emerged, significantly enhancing accuracy and efficiency in identifying damaged structures. CNNs have evolved from simple binary image classification (damaged vs. undamaged) [26] to more sophisticated object detection frameworks like YOLO (You Only Look Once), capable of rapidly localizing damaged structures within large-scale imagery [27]. For instance, Ji et al. [28] successfully employed the SqueezeNet CNN to classify collapsed buildings following the devastating Haiti earthquake. Similarly, Miura et al. [29] analyzed post-event multispectral aerial images from the 2016 Kumamoto earthquake and showed that a customized CNN could distinguish collapsed, intact, and blue tarp-covered roofs with about 95% accuracy, confirming the promise of deep learning for rapid building damage mapping. Such CNN-based methodologies exhibited superior performance compared to traditional feature-based methods, highlighting CNNs’ capability for discriminative feature extraction [30]. In addition, the release of benchmark datasets like xBD further triggered advancements, facilitating research focused explicitly on building damage assessment using CNN architecture. However, researchers recognized that single-image CNN classifiers, while useful, struggled to generalize to new disaster scenes [31]. This led to the development of more complex architectures combining localization and classification, as well as techniques to utilize pre- and post-event image pairs for change detection.

Among various methodologies, semantic segmentation techniques have become increasingly prominent due to their ability to produce pixel-wise maps categorizing each pixel into specific damage classes. The U-Net architecture, characterized by its encoder-decoder design and skip connections, is particularly popular in this context. U-Net effectively preserves spatial and contextual information, thus enabling precise localization and accurate classification of damaged structures [32]. Recent advancements have introduced enhancements such as attention mechanisms within the U-Net architecture. Attention U-Nets, for example, utilize attention gates to selectively focus on relevant features, significantly improving segmentation accuracy, especially in complex disaster scenes. Wu et al. [33] demonstrated improved performance by incorporating attention mechanisms within Siamese U-Nets, achieving enhanced accuracy in classifying building damages across diverse disaster types. With time, U-Net-based architectures such as UNet++, ResU-Net, and BD-SKUNet have further enhanced segmentation accuracy for diverse damage patterns [34,35,36]. These advances enable more precise delineation of buildings and the identification of varying degrees of damage, resulting in detailed maps that indicate which structures are affected and to what extent.

2.2.3. Transformer and Hybrid Architectures

Recent developments in deep learning techniques, especially transformer-based architectures, have advanced automated damage assessment by enabling accurate pixel-level analyses. Transformer models, such as Vision Transformers and hybrid CNN–transformer architectures, have gained prominence due to their capability of capturing global contextual information essential in complex disaster imagery. Models such as SDAFormer by Da et al. [37] integrate Swin-Transformers within Siamese Frameworks, considerably outperforming CNN-based models while identifying subtle damage distinctions and improving global context modeling. Kaur et al. [38] similarly, introduced DAHiTrA, a Damage Assessment model using Hierarchical Transformers, which achieved excellent performance on large-scale disaster datasets by capturing multi-scale features with transformer encoders. Another innovative approach is the High-Resolution Transformer for Building Damage Assessment (HRTBDA) by Chen et al. [39], which uses a high-resolution vision transformer backbone to enhance multi-scale feature extraction, along with a Cross-Attention-based Spatial Fusion module that merges information from dual inputs. The attention mechanism in the HRTBDA enables the model to identify subtle damaged details that might be missed by purely convolutional models.

Another prominent advancement includes Siamese network architectures, which utilize pairs of pre- and post-disaster images to directly infer damage through change detection approaches [40]. However, many existing semantic segmentation methodologies rely on paired pre- and post-disaster images. While such paired-image approaches can achieve high accuracy [41], they are constrained by the availability and quality of pre-event imagery. The unpredictable nature of natural disasters frequently limits the accessibility of pre-disaster datasets, complicating assessments reliant on comparative analysis. Thus, there is a critical need for developing robust frameworks that exclusively utilize post-disaster imagery.

3. Methodology

This section details the methodology employed to develop DamageScope, an integrated deep learning framework designed for detecting and classifying building damage levels from post-disaster satellite imagery and subsequently visualizing these results via an interactive web platform to facilitate operational decision making. The workflow begins by utilizing post-disaster satellite imagery from the xBD dataset [42], a comprehensive dataset encompassing a wide range of natural disasters with annotated structural damage levels. The selected images are systematically partitioned into training, validation, and testing subsets to ensure thorough model training and rigorous performance assessment. The proposed deep learning architecture integrates the FastViT [43] transformer-based encoder with a U-Net-based decoder. FastViT is a Fast Hybrid Vision Transformer that can effectively capture rich spatial and semantic context, while the U-Net decoder leverages these extracted features for precise pixel-level segmentation, classifying damage into four categories: no damage, minor damage, major damage, and destroyed. After model training and validation on the xBD dataset, the trained network is specifically applied to post-disaster images collected following Hurricane Ian to demonstrate practical utility and validate the approach in a real-world disaster scenario.

Following model inference, the resulting segmentation masks enable the classification and spatial delineation of damaged buildings. Subsequently, geospatial coordinates corresponding to these identified damaged structures are accurately derived using georeferencing metadata embedded within the satellite imagery. Finally, the classified damage information, along with extracted geolocation data, is integrated into an interactive, web-based visualization platform. This platform supports rapid situational assessment, enabling emergency responders, humanitarian organizations, and disaster management authorities to visually explore damage patterns and prioritize resource allocation effectively. An overview of the entire methodological workflow is provided in Figure 1 for clarity and comprehensive understanding.

3.1. Deep Learning for Building Damage Assessment

High-resolution satellite imagery presents challenges for conventional convolutional neural network (CNN) models due to complex textures, multi-scale features, and heterogeneous object classes. While CNNs are effective at capturing local patterns, they often struggle with modeling long-range dependencies across large spatial extents. Recent advances in transformer-based methods have shown promise in addressing these limitations, as their self-attention mechanisms can capture global semantic relationships more effectively [44]. In this study, we propose a hybrid encoder–decoder framework that integrates convolutional and transformer-based components for improved segmentation performance. The encoder is built on the FastViT-MA36 variant [43], a hybrid vision transformer that combines large-kernel convolutions with efficient attention mechanisms to balance spatial detail extraction and contextual reasoning. This architecture was selected for its strong representational capacity and computational efficiency, which enhance segmentation robustness across complex post-disaster scenes. The decoder follows a U-Net design [45], progressively restoring spatial resolution via upsampling and incorporating skip connections from the encoder to recover fine-grained localization details. This design supports accurate pixel-level segmentation of building damage while preserving boundary precision across varied image conditions.

Specifically, the encoder structure begins with three MobileOne stems that downsample the input image from the original size (H × W) to a quarter resolution (H/4 × W/4) while expanding feature channels. This is followed by a FastViT stage that mixes features at constant resolution. Subsequent stages employ “PatchEmbed” convolutions, progressively halving spatial dimensions (H/4 → H/8 → H/16 → H/32) while maintaining or increasing the number of feature channels, resulting in a multi-scale feature pyramid. These hierarchical features provide rich spatial and semantic representations necessary for precise damage segmentation.

The decoder mirrors the encoder in reverse, sequentially upsampling feature maps to reconstruct the original spatial resolution. At each decoding stage, the upsampled features are concatenated with the corresponding encoder outputs via skip connections and then refined through convolutional layers. This design facilitates the recovery of fine-grained details that are often lost during downsampling, improving the model’s ability to delineate building boundaries and accurately classify damage severity. The final output layer generates dense, per-pixel class scores for five categories: background, no damage, minor damage, major damage, and destroyed.

A detailed summary of the architectural components, including input and output resolutions and channel dimensions across different network stages, is provided in

Table 1. Network architecture description.

Stage	Resolution	Channels In → Out	Block Type	Skip-Conn
Input	H × W	3 → –	–	–
Stem 0	H/2 × W/2	3 → 76	MobileOneBlock: 3 × 3 Conv (stride 2) → BN → GELU	–
Stem 1	H/4 × W/4	76 → 76	MobileOneBlock: Depthwise 3 × 3 Conv (s = 2) → BN → GELU	–
Stem 2	H/4 × W/4	76 → 76	MobileOneBlock: Pointwise 1 × 1 Conv→BN	–
Stage 0	H/4 × W/4	76 → 76	RepMixerBlocks	→ Decoder 2 (76 ch)
Stage 1	H/8 × W/8	76 → 152	PatchEmbed (s = 2) → RepMixerBlocks	→ Decoder 1 (152 ch)
Stage 2	H/16 × W/16	152 → 304	PatchEmbed (s = 2) → RepMixerBlocks	→ Decoder 0 (304 ch)
Stage 3	H/32 × W/32	304 → 608	PatchEmbed (s = 2) → AttentionBlocks	–
Decoder 0	H/16 × W/16 ↑	(608 + 304) = 912 → 256	Upsample × 2 → Concat → 2 × Conv2dReLU → Identity	← Stage 2 (304 ch)
Decoder 1	H/8 × W/8 ↑	(256 + 152) = 408 → 128	Upsample × 2 → Concat → 2 × Conv2dReLU → Identity	← Stage 1 (152 ch)
Decoder 2	H/4 × W/4 ↑	(128 + 76) = 204 → 64	Upsample × 2 → Concat → 2 × Conv2dReLU → Identity	← Stage 0 (76 ch)
Decoder 3	H/2 × W/2 ↑	64 → 32	Upsample × 2 → Concat → 2 × Conv2dReLU → Identity	–
Decoder 4	H × W ↑	32 → 16	Upsample × 2 → 2 × Conv2dReLU → Identity	–
Seg Head	H × W	16 → num_classes	Conv2d(k = 3, p = 1) → Activation	–

Note: “(↑)” means upsampled by 2×. “(→)” denotes the direction of feature flow or mapping. “(←)” represents a skip connection that receives features from the corresponding encoder stage.

. Figure 2 illustrates the overall U-Net and FastViT integration, depicting the data flow from input through successive downsampling and upsampling stages to the final segmentation output.

3.2. Geolocation Information Derivation

Following the identification of damaged structures in satellite imagery, the next step involves accurately deriving their geographic coordinates for spatial analysis and visualization. High-resolution satellite images typically represent the Earth’s surface as raster datasets, where spatial information is stored in a grid of uniformly sized pixels. Each pixel holds numeric values corresponding to spectral reflectance or other radiometric measurements, allowing analysts to infer meaningful ground-level features such as buildings, vegetation, and infrastructure.

To relate pixel locations in raster images to real-world geographic coordinates, specialized file formats are employed to embed geospatial metadata. One of the most widely used formats is the GeoTIFF (Geographic Tagged Image File Format), an extension of the standard TIFF (Tagged Image File Format) file that includes information about coordinate reference systems, map projections, and transformation parameters. GeoTIFFs ensure that each pixel corresponds to a precise location on the Earth’s surface, enabling seamless integration with Geographic Information Systems (GISs) and other spatial analysis platforms. By embedding these metadata directly within the image file, GeoTIFFs facilitate accurate geolocation and simplify the process of managing and sharing spatial datasets across various applications, including disaster response and urban planning.

In this study, satellite imagery from the xBD dataset was utilized, providing high-resolution post-disaster images in the GeoTIFF format. The embedded georeferencing information allowed us to systematically extract real-world coordinates for the damaged structures identified through segmentation. The coordinate extraction process follows the standard affine transformation, Equations (1) and (2), proposed by ESRI [46]:

$$x_1=Ax+By+C$$

(1)

$$y_1=Dx+Ey+F$$

(2)

where the variables represent the following:

x₁, y₁: real-world coordinates;

x, y: pixel indices (column, row);

A and E: pixel size in x and y directions (E is often negative);

B and D: rotation/shear terms;

C and F: map coordinates of the upper-left pixel.

By applying these transformations, pixel indices corresponding to detected damaged structures were converted into precise latitude and longitude coordinates. These derived geospatial points were then aggregated and visualized to illustrate the spatial distribution of building damage across the affected regions. This geolocation step is critical for enabling subsequent mapping, analysis, and decision making in disaster response operations.

3.3. Web-Based Visualization

3.3.1. Web Server

We deployed our testing website on a cloud server hosted by Amazon Web Services (AWS), configured identically to our local workstation used for model training. This consistent setup ensures seamless integration between development and deployment environments, optimizing performance and maintaining platform consistency. The website is powered by Nginx, an open-source web server that also functions as a reverse proxy and load balancer. Renowned for its high efficiency, scalability, and ability to handle large numbers of concurrent connections, Nginx is particularly well suited for building damage assessment applications. Its lightweight, event-driven architecture supports fast data processing and response times, ensuring both reliability and speed in cloud-based environments where system stability is essential.

To enable interactive visualization of categorized building damage data, we developed the index.html page using Leaflet, a lightweight, open-source JavaScript library designed for web-based mapping applications. Leaflet allows for the dynamic and intuitive presentation of geospatial data through an interactive map interface. In our system, we use four distinct, color-coded markers to represent different damage severity levels. Markers in red, yellow, and green indicate buildings assessed as damaged, with the colors reflecting varying levels of severity. Blue markers are reserved for buildings identified as undamaged. This visual categorization enhances situational awareness by enabling users to quickly distinguish between damaged and undamaged structures.

The interactive map provides a comprehensive overview of reported building damage, categorized by severity and type. Users can interact with the map by clicking on individual markers, which trigger pop-up windows displaying detailed information about each building. This functionality facilitates rapid situational assessment by the evaluation team, offering critical details such as geographic location and damage classification. The map’s interactive elements, clickable markers, and pop-up data support efficient and informed decision making.

3.3.2. Model Deployment and Visualization

The website server integrates the classification model for assessing building damage levels. The mapping interface displays markers stored in JSON (JavaScript Object Notation) format, generated by a backend Python script (Python 3.8). The model processes the input data, and the classification results for each building, along with its geographic coordinates, are encapsulated into structured JSON markers. These markers are then rendered on the interactive map, enabling users to visualize the spatial distribution of building damage.

Figure 3 presents the detailed architecture and workflow of our proposed building damage visualization system. The process begins with data loading to the model, which undergoes classification to determine the damage level. After classification, the geographic information is combined with the model’s results and formatted into JSON markers. These markers are stored on the server and accessed by the web interface for visualization.

The interactive map developed using Leaflet 2.0.0-alpha and JavaScript (ES2024), fetches and displays all markers. Each damage category is represented by a distinct color-coded marker, allowing the evaluation team to quickly interpret severity levels across different buildings. Users can click on individual markers to view detailed building information. Additionally, a filtering feature has been implemented, allowing users to selectively display buildings based on their damage classification level, thereby enhancing usability and decision-making efficiency.

4. Experiment and Results

4.1. Dataset Preparation

Effective deep learning-based post-disaster damage assessment requires datasets that encompass a wide range of disaster types, damage severities, geographic regions, and building structures. Such diversity enhances the model’s ability to generalize and accurately identify damaged buildings across varied scenarios. Publicly available datasets such as BRIGHT [47], BDD [48], and xBD [42] have become widely adopted for this purpose, offering extensive coverage of disaster-affected areas.

In this study, we selected the xBD dataset due to its size, diversity, and comprehensive annotations. The xBD dataset contains approximately 850,736 building annotations distributed across 45,362 km² of imagery, with each image having a spatial resolution of 1024 × 1024 pixels. Covering 19 major disaster events, the dataset provides both pre- and post-disaster imagery across multiple geographic regions, making it one of the most extensive and versatile resources for building damage assessment research. All imagery was sourced from the Maxar Open Data Program. For this research, we exclusively utilized the post-disaster imagery from the xBD dataset to align with our goal of developing a damage assessment framework that does not depend on pre-disaster data. Figure 4 presents examples of disaster scenes from the dataset, with annotated building damage levels.

The xBD dataset adopts a unified damage scale (

Table 2. Combined damage level proposed in the xBD dataset.

Damage Level	Description
No damage	No sign of water, structural, or shingle damage or burn marks.
Minor damage	Building partially burnt, water surrounding structure, volcanic flow nearby, roof elements missing, or visible cracks.
Major damage	Partial wall or roof collapse, encroaching volcanic flow, or surrounded by water/mud.
Destroyed	Scorched, completely collapsed, partially/completely covered with water/mud, or otherwise no longer present.

), classifying building impairments into four categories: no damage, minor damage, major damage, and destroyed. This standardized classification framework accommodates the variability inherent in multi-disaster contexts, enabling consistent model training across diverse scenarios.

For model development, we partitioned the post-disaster images into three subsets: 9168 images for training, 933 images for validation, and 933 images for testing. This balanced partitioning strategy exposes the model to a wide range of damage patterns during training while enabling rigorous evaluation of its generalization performance. By focusing exclusively on post-event imagery, the proposed framework is positioned to provide accurate, real-time damage segmentation and support rapid decision making in disaster response operations.

4.2. Implementation Details

Model training and evaluation were conducted on a high-performance workstation running Ubuntu 22.04, equipped with an AMD^® Ryzen Threadripper Pro 5955WX CPU (Advanced Micro Devices, Inc., Santa Clara, CA, USA), 128 GB of RAM, and dual NVIDIA RTX A6000 GPUs (NVIDIA Corporation, Santa Clara, CA, USA). This setup ensured efficient handling of the large-scale satellite imagery and complex model computations required for the study. The deep learning network was implemented using the PyTorch 2.1.2 framework. Model optimization was performed using the Adam optimizer, with an initial learning rate set to 0.0001. A StepLR scheduler was employed to dynamically adjust the learning rate during training, reducing it by a factor of 0.5 every 10 epochs to facilitate convergence. The loss function used for model training was the multi-class Dice Loss, selected to directly optimize for segmentation accuracy in the presence of class imbalance. By focusing on overlap between predicted and true segmentation masks, the Dice Loss effectively penalizes false positives and false negatives, which is critical for reliable damage classification.

4.3. Metrics

In this study, the performance of the proposed model is evaluated using the F1 score and Intersection-over-Union (IoU) metrics. Given that ground truth annotations are provided at the pixel level and instances of damaged buildings are relatively infrequent, overlap-based evaluation metrics are prioritized. These metrics emphasize spatial precision over tile-level classification accuracy, enabling a more accurate representation of the dispersed and irregular patterns associated with structural damage. Among these, the F1 score and IoU are particularly well suited for this task due to their robustness under class imbalance and their strong correspondence with expert evaluations of the spatial extent of damage, as evidenced in prior studies employing the xBD dataset [32,49].

The F1 score is the harmonic mean of precision and recall. It balances the proportion of correctly identified positive pixels (recall) with how many of the pixels the model labeled as positive are correct (precision), yielding a value between 0 (worst) and 1 (perfect). The mathematical expressions for precision, recall, and F1 score are provided in Equations (3)–(5):

$$Precision={\displaystyle \frac{TP}{TP+FP}}$$

(3)

$$Recall={\displaystyle \frac{TP}{TP+FN}}$$

(4)

$$F_1={\displaystyle \frac{2\ast Precision\ast Recall}{Precision+Recall}}$$

(5)

For multi-class evaluation, the macro-average F1 score is computed by taking the arithmetic mean across all the damage classes. This metric is calculated to provide a single, balanced measure of overall model performance, giving equal importance to each class. The mathematical expression for the macro-average F1 score is defined in Equation (6):

$$\mathrm{M}\mathrm{a}\mathrm{c}\mathrm{r}\mathrm{o}-\mathrm{a}\mathrm{v}\mathrm{e}\mathrm{r}\mathrm{a}\mathrm{g}\mathrm{e}\mathrm{F}1\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}={\displaystyle \frac{1}{N}}\sum _{i=1}^N{\mathrm{F}1}_i$$

(6)

where F1_i is the F1 score for class i, and N is the total number of damage classes.

The Intersection over Union (IoU), also called the Jaccard index, measures spatial overlap in segmentation: the area where prediction and ground truth intersect divided by the area encompassed by either of them. An IoU of 1 indicates identical masks; 0 means no overlap. The mathematical expression for Intersection over Union (IoU) is defined in Equation (7):

$$IoU={\displaystyle \frac{P\cap G}{P\cup G}}$$

(7)

where P is the predicted segmentation mask, and G is the ground truth segmentation mask.

4.4. Results

The quantitative performance of the proposed model across damage severity levels is summarized in

Table 3. Model performance on the validation and test sets.

Damage Level	Validation Set		Test Set
	IoU	F1	IoU	F1
No damage	0.603	0.752	0.608	0.756
Minor damage	0.276	0.432	0.299	0.460
Major damage	0.418	0.589	0.428	0.599
Destroyed	0.387	0.558	0.406	0.577
Macro-average	0.421	0.583	0.435	0.598

, reporting Intersection-over-Union (IoU) and F1 scores for both the validation and testing sets. The model demonstrates the highest accuracy in identifying undamaged buildings, achieving the best IoU and F1 scores among all categories. Specifically, for the validation set, it attained an IoU of 0.603 and an F1 score of 0.752 for the no-damage class. A similar trend is observed in the testing set, with an IoU of 0.608 and an F1 score of 0.756, indicating strong segmentation performance and reliable delineation of intact structures. Performance declines with increasing damage severity. The minor damage category yields the lowest scores across both datasets, with an IoU of 0.276 and an F1 score of 0.432 on the validation set and an IoU of 0.299 and an F1 score of 0.460 on the testing set. This outcome reflects the inherent difficulty of detecting minor structural damage, where visual cues are often subtle and less distinguishable from undamaged conditions. Such challenges are expected due to the limited visual degradation and greater intra-class variability. For the major damage and destroyed categories, the model achieves intermediate performance. On the validation set, it records an IoU of 0.418 and an F1 score of 0.589 for major damage, and an IoU of 0.387 with an F1 score of 0.558 for destroyed buildings. Similar results are observed in the testing set, with IoUs of 0.428 and 0.406 and F1 scores of 0.599 and 0.577 for major damage and destroyed categories, respectively. These results suggest that while the model is capable of identifying severely damaged structures, further improvements—such as targeted data augmentation, fine-tuning, or class-specific loss weighting—may enhance its sensitivity to nuanced damage levels, particularly for minor damage.

In addition to the quantitative results, qualitative assessments were conducted using representative post-disaster images from the xBD test set, as shown in Figure 5. The figure presents visual comparisons between ground truth annotations and the model’s predicted segmentation maps across different damage levels. The results demonstrate the model’s ability to accurately capture building morphology, location, and damage extent. Severely damaged and destroyed structures are consistently identified, which is critical for prioritizing search and rescue operations. The model also shows reasonable distinction between minor and major damage categories, reflecting its sensitivity to varying levels of degradation. However, some minor damage cases are missed, and occasional false positives arise in visually complex scenes, often due to textural similarities. These issues point to residual confusion at damage severity boundaries and occasional over-segmentation. Improving the model’s discrimination between subtle damage classes remains essential for real-world deployment, where nuanced distinctions directly influence emergency planning and response.

4.5. Case Study: Hurricane Ian

To further evaluate the performance of the proposed framework, we conducted a case study focused on Hurricane Ian, a major storm during the 2022 Atlantic hurricane season. Hurricane Ian originated from a tropical wave in the central Caribbean Sea on 23 September 2022 and dissipated on 1 October 2022 [50]. Tracking generally west–northwestward, Ian first made landfall near Pinar del Río Province, Cuba, on September 27 as a Category 3 hurricane on the Saffir–Simpson scale, causing widespread wind damage and flooding across the island. After reemerging over the Gulf of Mexico, Ian rapidly intensified to a Category 4 hurricane before making landfall in southwestern Florida near Cayo Costa on 28 September, with maximum sustained winds of 150 mph and a minimum central pressure of 935 mbar. The storm then moved northeastward across the Florida peninsula, passing south of Orlando and later re-entering the Atlantic Ocean near Daytona Beach before affecting coastal areas of Georgia, South Carolina, and North Carolina.

Hurricane Ian’s impact was most severe in Southwest Florida, particularly in Lee County, where storm surges exceeded 4 m on Sanibel Island and over 2.1 m in Cape Coral and Fort Myers. Coastal communities experienced widespread inundation and infrastructure destruction. The storm surge, amplified by high tides and a broad wind field, led to a “life-threatening and historic” inundation of 12 to 16 feet in areas such as Fort Myers Beach, resulting in mandatory evacuation orders affecting over 2 million residents [51]. In total, Hurricane Ian caused more than 150 direct and indirect fatalities, displaced tens of thousands of people, and damaged or destroyed over 60,000 homes. Economic losses were estimated at USD 112 billion, making Ian the costliest hurricane in Florida’s history and the third costliest in U.S. records [52]. Additionally, more than 2.5 million residents experienced sustained power outages, and extensive inland flooding severely impacted agricultural sectors.

For this study, post-disaster satellite imagery for Hurricane Ian was obtained from the Maxar Open Data Program. Based on the observed extent of structural damage, we selected Fort Myers Beach, Florida, as the area of focus due to its significant concentration of heavily impacted buildings. Figure 6a presents the post-disaster conditions of the area, highlighting the widespread devastation. Figure 6b displays the corresponding building damage predictions generated by our model, effectively illustrating the spatial distribution and severity of structural impacts across the affected region. While the model’s predictions are not flawless, they provide substantial value in an emergency response scenario. Accurate identification of majorly damaged or destroyed structures can significantly enhance the efficiency of resource allocation and recovery planning following a disaster. Based on the model outputs, we applied the geolocation transformation equations (Equations (1) and (2)) provided by ESRI to calculate the geographic coordinates of the identified buildings across different damage categories. These coordinates were subsequently used to support web-based visualization, enabling rapid spatial assessment of damage severity across the affected region.

Figure 7 presents the interactive web-based visualization of building damage predictions for Hurricane Ian. Following the server workflow outlined in Section 3, the model’s outputs were transformed into a spatially explicit map, providing an intuitive overview of structural impacts across the affected coastline. Each building is represented by a color-coded marker corresponding to its predicted damage level: red for destroyed, yellow for major damage, green for minor damage, and blue for no damage. The clustering of markers by color enables users to rapidly identify the most severely impacted areas, particularly along the beachfront, while predominantly blue regions inland indicate minimal structural damage. A legend and layer-control widget located in the upper-right corner allow users to selectively display specific damage categories, facilitating targeted analysis of priority zones or identification of potential staging areas. Clicking on an individual marker opens a pop-up window displaying the building’s ID, precise geographic coordinates, and predicted damage classification, providing detailed contextual information without leaving the map interface. The visualization platform, developed using the lightweight Leaflet JavaScript library and served through Nginx on AWS, ensures fast loading times and compatibility across desktop, tablet, and mobile devices.

5. Discussion

5.1. Class-Wise Performance and Confusion Matrix Analysis

To evaluate the segmentation model’s performance and identify sources of misclassification, confusion matrices provide detailed insights by summarizing predictions across each class. Figure 8 presents the confusion matrices for both the validation and test sets, highlighting where the model correctly identifies damage levels and where confusion occurs. On the test set, the model accurately identifies background pixels 98% of the time and classifies buildings with no damage with 78% accuracy. In contrast, the model struggles with the “Minor” damage category, correctly classifying only 43% of instances. A significant portion of these are misclassified as “No Damage” (19%), “Major” (9%), or even “Background” (29%). The “Major” and “Destroyed” categories exhibit lower confusion, suggesting more distinctive visual features. These patterns reflect challenges stemming from class imbalance in the xBD dataset and the visual similarity between adjacent categories, particularly between “Minor” and “Major” damage. The confusion matrix thus underscores that buildings with minor damage are frequently misclassified, especially into the “No Damage” category, which reduces the overall reliability of predictions for this class.

Figure 9 illustrates examples of erroneous segmentations produced by the model. The ground truth masks indicate a mix of “Destroyed” and “Major Damage” classifications; however, the predicted masks reveal frequent confusion between these two categories. Specifically, several buildings labeled as “Destroyed” are misclassified as “Major Damage,” and vice versa. This confusion can be attributed, in part, to the model’s reliance solely on post-event imagery, which often lacks clear structural outlines, especially in the case of collapsed buildings. Without distinct object boundaries, the model tends to over-segment the scene, grouping adjacent pixels that may correspond to separate structures. This behavior significantly undermines the model’s ability to accurately assess the severity of damage, even when it successfully detects the presence of damaged buildings. Additionally, the class imbalance within the xBD dataset—particularly the under-representation of “Destroyed” structures—further exacerbates this issue. These misclassifications underscore the necessity for improved class delineation, more balanced loss functions, and targeted data augmentation strategies to enhance model sensitivity to severely damaged buildings.

5.2. Performance Comparison

To assess the effectiveness of the proposed model, we conducted a comparative evaluation against widely adopted baseline segmentation architectures. Figure 10 presents the Intersection over Union (IoU) and F1 scores for U-Net, U-Net with ResNet50 backbone [53], and U-Net with FastViT backbone across the validation and test sets, disaggregated by damage severity categories in the xBD dataset. The results clearly demonstrate that the U-Net + FastViT model consistently outperforms the other configurations across all classes, namely “No Damage,” “Minor Damage,” “Major Damage,” and “Destroyed.” For instance, in the “No Damage” category, U-Net + FastViT achieves approximately 0.75 IoU and 0.76 F1 scores on the test set, outperforming U-Net (0.69 IoU, 0.69 F1) and U-Net + ResNet50 (0.73 IoU, 0.73 F1). Similar performance margins are observed in the other categories, underscoring the robustness and generalizability of the FastViT-enhanced model. This consistent performance gain across varying levels of structural damage indicates that U-Net + FastViT delivers superior segmentation accuracy and reliability in post-disaster damage assessment tasks.

To expand our analysis of backbone architectures, we further evaluated the performance of ResNet50 integrated with DeepLabV3+ [54] and Feature Pyramid Network (FPN) [55] architectures, comparing them with the U-Net + FastViT model under consistent class definitions and dataset splits. Figure 11 presents the IoU and F1 scores for these configurations across the validation and test sets, segmented by damage severity levels. The results indicate that U-Net + FastViT consistently achieves superior performance across all classes—most notably in the more challenging “Minor Damage” and “Major Damage” categories—surpassing both DeepLabV3+ and FPN-based variants. This performance advantage highlights the effectiveness of the FastViT backbone, whose hybrid CNN–transformer architecture facilitates richer contextual reasoning and enhanced feature representation. Such capabilities enable the model to better differentiate nuanced damage patterns, leading to more precise segmentation outcomes, particularly in classes where fine-grained visual cues are critical.

5.3. Limitations and Future Research Directions

While this study presents a practical framework for post-disaster building damage assessment, several limitations should be acknowledged, providing opportunities for future research and system improvement. First, the segmentation model exhibits varying performance across damage severity classes. Although the model reliably identifies undamaged and severely damaged structures, its accuracy diminishes when distinguishing between minor and moderate damage levels. This reflects broader challenges in damage assessment from satellite imagery, where subtle structural changes are difficult to capture. Future work could explore advanced data augmentation strategies, damage-aware loss functions, or multi-task learning frameworks to enhance the model’s sensitivity to less severe damage categories. Second, the current framework operates exclusively on satellite imagery at a fixed spatial resolution. While satellite images offer broad coverage, they may miss fine-grained structural details compared to higher-resolution aerial or drone imagery. Integrating multisource remote sensing data, including UAV or ground-based imagery, could improve the detection of smaller-scale damages and provide richer context for model predictions.

Third, the geolocation extraction process assumes accurate and consistent georeferencing metadata within the input satellite imagery. However, minor misalignments or distortions, especially in heavily damaged urban environments, could affect the precision of coordinate calculations. Incorporating automated post-processing steps, such as map-matching or spatial correction algorithms, could further enhance the spatial accuracy of identified damage locations. Finally, while the developed web visualization platform demonstrates rapid deployment and accessibility, it currently functions as a static post-processing tool. Future developments could focus on real-time integration of model outputs, allowing dynamic updates as new imagery becomes available. Enhancements such as user-driven annotation, collaborative decision-making tools, or predictive analytics modules could further expand the platform’s utility for disaster management agencies.

6. Conclusions

This study presents a practical framework for post-disaster building damage assessment using deep learning and satellite imagery. The proposed system operates solely on post-disaster images, addressing a critical gap where pre-disaster reference data are often unavailable. By integrating a transformer-enhanced encoder for damage classification, geolocation extraction for spatial mapping, and a lightweight web-based visualization platform, the framework offers an end-to-end solution for rapid damage assessment in disaster response operations. Quantitative evaluations on the xBD dataset demonstrate that the model effectively identifies and segments damaged structures, particularly in cases of major and destroyed buildings. Qualitative assessments, including a case study of Hurricane Ian, further illustrate the framework’s ability to provide detailed, actionable insights into the spatial distribution of building damages. The results demonstrate the potential of post-disaster-only deep learning approaches to support emergency response efforts, enhancing situational awareness and facilitating informed decision making. This study highlights the practical integration of artificial intelligence into disaster management workflows, contributing toward more efficient and timely disaster response capabilities.