Rapid post-disaster building damage assessment is essential for effective emergency response and recovery planning, yet traditional field-based evaluation and manual inspection of satellite imagery are labor intensive and inefficient for large, affected areas. This study develops an integrated workflow in which a Convolutional Neural Network (CNN) is used solely as a building damage classifier, while a Geographic Information System (GIS) serves as the primary tool for spatial analysis.
A ResNet-18 model was trained on 128×128 pre- and post-disaster building patches from the xBD dataset to perform binary classification of damaged and undamaged buildings for Hurricane Harvey (2017). The CNN predictions were then joined to building footprints and analyzed in GIS to identify and evaluate spatial damage patterns. Three GIS-based assessments were conducted: (1) a true versus predicted accuracy map to visualize spatial patterns of correct and incorrect classifications, (2) a distance-from-river analysis to examine the relationship between river proximity and building damage, and (3) kernel density estimation (KDE) to compare spatial clustering of damage.
Results show that the model achieved strong performance (88% overall accuracy) and the CNN predicted labels produced spatial damage patterns similar to the ground truth. These findings demonstrate that, even when the CNN functions only as a classifier, its output can be reliably used for spatial reasoning in GIS. Overall, the combined CNN–GIS workflow offers a faster and more efficient approach to large-area post-disaster damage mapping.
1. Introduction
Natural hazards such as hurricanes, floods, and earthquakes cause widespread structural damage, demanding rapid assessment to guide emergency responses, prioritize resources, and support early recovery. Traditional building damage assessment relies on field inspections, reports from local authorities, and manual interpretation of satellite imagery. Although these methods provide valuable insights, they are often time-consuming, difficult to scale, and sometimes unsafe for responders— particularly when the affected area is large and densely populated. These limitations highlight the need for approaches that can automate building-level damage classification while still enabling detailed spatial analysis of the disaster.
Recent advances in remote sensing and machine learning have introduced new capabilities for rapid damage detection. High-resolution satellite imagery provides extensive geographic coverage in the immediate aftermath of a disaster, while CNNs can learn visual patterns of damage and classify buildings accordingly. However, many existing studies end at the classification stage, producing a damage map without deeper spatial reasoning. Spatial analysis remains essential for understanding disaster patterns and supporting planning decisions.
This project addresses that gap by integrating a CNN-based building damage classifier with GIS to evaluate not only prediction accuracy but also spatial patterns. By comparing accuracy maps, damage–distance relationships, and kernel density surfaces, the study examines whether CNN outputs can be reliably used to reproduce key spatial trends within a flood-affected region.
2. Background
2.1 Geographic Information System (GIS)
A Geographic Information System (GIS) is a system or tool used to store, manage, analyze, and visualize geospatial data. In the context of damage assessment, GIS enables visualization of affected areas, analysis of spatial patterns, and the integration of multiple data sources such as imagery, building outlines, flood zones, and transportation networks. This integration provides a clearer understanding of how damage is distributed across a region.
In this project, GIS is used primarily for evaluation and interpretation: to determine whether the CNN generated labels are reliable for damage mapping. GIS therefore serves as the main environment for spatial reasoning, validation, and interpretation of CNN outputs.
2.2 Convolutional Neural Networks (CNNs)
Convolutional Neural Networks (CNNs) are deep learning models designed to learn spatial features from images, such as textures, edges, and complex patterns. For this study, a ResNet-18 architecture was selected due to its balance between accuracy and computational efficiency. Compared with deeper networks like ResNet-50 or ResNet-101, ResNet-18 uses significantly fewer parameters, enabling faster training and lower memory requirements—an important consideration in resource-limited environments.
ResNet incorporates residual “skip connections,” which help stabilize gradient flow and improve performance when learning complex visual cues such as debris, flood signatures, or shadowed structures. Previous studies have shown strong performance of ResNet models on remote sensing tasks, including damage detection, building footprint extraction, and land-use classification. In this project, the CNN is used strictly as a binary classifier (damaged vs. undamaged), not as a spatial analysis tool.
2.3 Dataset
The study uses the xBD dataset, a large-scale public dataset of high-resolution pre- and post-disaster satellite imagery coupled with building footprints and annotated damage labels. The dataset includes over 800,000 buildings from multiple global disaster events. For this project, only Hurricane Harvey (2017) data were used.
Building-centered image patches of size 128×128 pixels were extracted from pre- and post-disaster image tiles using the provided building polygons. The patch size was selected to capture each building and sufficient surrounding context (e.g., debris, water) while avoiding excessive background information. To ensure reliable model evaluation, the extracted patches were divided into training, validation, internal test, and external test datasets.
3. Study Area
Hurricane Harvey caused catastrophic flooding across the Greater Houston region in 2017, producing widespread structural damage across Harris County, Texas. Although the xBD dataset provides satellite imagery and building footprints scattered across multiple Harvey-affected locations, the spatial analyses in this study were conducted within a specific Area of Interest (AOI) located in western Harris County. The AOI lies near the Buffalo Bayou watershed and George Bush Park and includes a mix of residential neighborhoods, commercial structures, and open green areas.
This AOI is suitable for both distance-from-river analysis and cluster-based assessments such as kernel density estimation. While the CNN was trained using scattered Harvey tiles across the broader Houston region, the GIS analysis focused exclusively on this AOI to examine whether model predictions preserve local spatial damage patterns.
Overall, the AOI offers a well-defined spatial environment for evaluating how closely CNN predictions match real flood-related spatial trends and localized hotspots of damage.
4. Project Methodology
The workflow consists of two major phases:
- (a) CNN-based damage classification
- (b) GIS integration and spatial analysis
4.1 Phase 1 – CNN Classification
4.1.1 Patch Extraction
Using building footprints from the xBD dataset, 128×128 pixel image patches centered on each building were extracted. The centroid of each polygon footprint was used to compute a crop window on both pre- and post-disaster imagery tiles. This patch size ensured that the entire building and relevant surrounding context (debris, flood water) were captured while avoiding unnecessary background. The extracted patches were then split into training, validation, internal test, and external test datasets.
4.1.2 CNN Model Architecture
A ResNet-18 architecture was selected for binary classification, where 0 indicates an undamaged building and 1 indicates a damaged building. ResNet-18 is computationally efficient due to its smaller depth and reduced parameter count compared to deeper variants (e.g., ResNet-50), making it suitable for faster training while still providing strong feature extraction through residual skip connections.
The model was trained using:
- Optimizer: Adam
- Learning rate: 1 × 10-4
- Batch size: 32
- Epochs: 10
4.1.3 Model Evaluation Metrics
Performance was evaluated on an external test dataset using:
- Accuracy (overall correctness)
- Precision (correctness of predicted damaged buildings)
- Recall (percentage of actual damaged buildings detected)
- F1-score (balanced measure of precision and recall)
- Confusion matrix
These metrics provide a comprehensive understanding of classification performance.
4.1.4 CNN Output
The model outputs were stored in a CSV file containing:
uid– building IDpred_damage– predicted damage class (0 or 1)prob_damage– confidence score
This CSV serves as the primary input to Phase 2, where GIS is used for spatial reasoning.
4.2 Phase 2 – GIS Integration and Spatial Analysis
4.2.1 Data Integration
Building footprints from xBD (JSON format) were converted into a polygon feature class in ArcGIS
Pro and projected to WGS 1984 UTM Zone 15N to allow accurate distance calculations in meters. The
CSV file containing CNN predictions was imported and joined to the footprints using the
uid field, creating a spatial dataset that stores both true and predicted damage labels for
each building.
4.2.2 True vs. Predicted Accuracy Map
A comparison map was generated to visualize building-level damage based on true labels and CNN predictions, highlighting spatial patterns of agreement and error within the AOI.
4.2.3 Distance from River Damage Analysis
A nearby river (Bayou) within the study area was digitized in ArcGIS Pro. Euclidean distance from each building to the river was computed using the Near tool. Buildings were grouped into distance bands (0–200 m, 200–500 m, 500–1000 m, and >1000 m). For each band, the percentage of damaged buildings was calculated separately for true and predicted labels to examine whether CNN predictions capture the same flood-related spatial gradients as the ground truth.
4.2.4 Kernel Density Estimation (KDE)
KDE was generated for both true and predicted damaged buildings to identify hotspots of concentrated damage. KDE maps allow comparison of spatial clustering patterns. Similar KDE surfaces for true and predicted damage indicate that CNN predictions retain spatial information relevant for hotspot analysis, even when individual building classifications contain errors.
5. Results and Discussion
5.1 CNN Performance
The ResNet-18 model demonstrated strong performance in distinguishing damaged and undamaged buildings. On the external test set, the model achieved 88% overall accuracy with balanced precision and recall across both classes, indicating that the classifier can reliably detect building damage from post-disaster imagery.
For undamaged buildings, precision was 0.92 and recall was 0.86, indicating few false positive alarms. For damaged buildings, recall reached 0.92 with precision of 0.85, meaning most truly damaged buildings were correctly identified with a moderate rate of false positives.
The confusion matrix shows:
- 2,546 true undamaged buildings correctly classified
- 420 undamaged buildings incorrectly predicted as damaged
- 2,310 damaged buildings correctly identified
- 212 damaged buildings missed (predicted as undamaged)
Misclassifications often occurred in visually ambiguous areas, such as buildings partially obscured by trees or structures covered by shadows. Because the CNN relies solely on optical post-disaster imagery, damage modes like interior flooding or foundation failures that leave little visible exterior evidence cannot be detected. Variation in tile quality, illumination, and viewing angle may also introduce inconsistencies.
5.2 True vs. Predicted Accuracy Map
A representative tile from the AOI was used to visually compare true building damage labels with CNN predictions. The accuracy map displays pre-disaster imagery with building outlines, true labels, CNN predicted labels, and a prediction accuracy layer where correctly classified buildings appear in white and misclassified buildings in magenta.
The accuracy map demonstrates that CNN predictions align well with the true damage distribution across the neighborhood, with misclassifications concentrated in visually complex areas.
5.3 Distance from River Analysis
The AOI map with the digitized river network and damaged buildings reveals that both true and predicted damage trends exhibit a clear pattern: the percentage of damaged buildings is highest near the river and decreases with distance. This indicates that flooding near the river produced stronger concentrations of damage during Hurricane Harvey, and that the CNN successfully reproduces this flood-related gradient.
5.4 Kernel Density Estimation (KDE) Analysis
KDE maps provide smoothed visualizations of where damaged buildings are spatially concentrated in the AOI. Each damaged building contributes to a continuous density surface using a 200 m search radius. The output is expressed in kernel-weighted density units per square kilometer, representing relative clustering intensity rather than raw building counts.
KDE surfaces generated from ground truth labels and CNN predictions highlight several consistent hotspots. Minor differences appear where the CNN slightly over- or underestimates local damage intensity, and the smoothing process can spread classification errors over nearby areas. Nonetheless, the overall similarity between the two KDE surfaces demonstrates that CNN-generated building labels can reliably support GIS-based spatial analyses.
6. Conclusion
This project demonstrates an integrated workflow that uses a CNN purely as a building damage classifier and GIS as the main environment for spatial analysis. Using pre- and post-disaster building images from the xBD Hurricane Harvey dataset, a ResNet-18 model was trained for binary damage classification and achieved strong performance.
Although the CNN itself does not perform any spatial reasoning, its prediction outputs were successfully transferred into GIS for further interpretation. Across all analyses—accuracy mapping, distance-from-river assessment, and KDE—the CNN predictions closely matched the spatial patterns of the true labels. These results show that CNN predictions, despite being purely pixel-based classifications, are reliable enough to support GIS-based spatial assessments.
Overall, combining CNN classification with GIS provides an efficient and scalable approach for large- area disaster damage assessment. Even with a simple binary model and a limited AOI, the integrated workflow produced meaningful maps that reflect real flood-related damage patterns and can support post-disaster mapping and preliminary impact assessment.
7. Future Work
Future work may expand this study by:
- Incorporating multiclass damage levels.
- Integrating elevation or flood-depth rasters.
- Applying uncertainty mapping to better understand model confidence and guide interpretation of predictions.
- Extending the workflow to larger areas or multiple disasters to evaluate generalization potential.
- Testing more advanced CNN architectures or adding spatial context directly into the model to further improve prediction reliability before GIS integration.
References
- Blake, E. S., & Zelinsky, D. A. (2018). Tropical Cyclone Report: Hurricane Harvey (AL092017). National Hurricane Center, NOAA. Available at: National Hurricane Center technical report.
- Esri (2025). Euclidean Distance (Spatial Analyst) – ArcGIS Pro Documentation.
- Esri (2025). Kernel Density (Spatial Analyst) – ArcGIS Pro Documentation.
- Gupta, R., Hosfelt, R., Sajeev, S., et al. (2019). xBD: A Dataset for Assessing Building Damage from Satellite Imagery. arXiv:1911.09296.
- He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep Residual Learning for Image Recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 770–778).
- Yang, R., Meng, X., Yao, Y., Chen, B., You, Y., & Xiang, Z. (2018). An Analytical Approach to Evaluate Point Cloud Registration Error Utilizing Targets. ISPRS Journal of Photogrammetry and Remote Sensing, 143, 48–56.