[Paper Review] Satellite imagery analysis for operational damage assessment in Emergency situations

Alexey Trekin, German Novikov+3 more|arXiv (Cornell University)|Feb 19, 2018

Remote-Sensing Image Classification4 references3 citations

TL;DR

This paper proposes a deep learning-based workflow for rapid, automated damage assessment in disaster response using pre- and post-event satellite imagery. By training a U-Net-like semantic segmentation model on labeled building damage data and fine-tuning it on new regions with minimal labeled data, the method reduces manual mapping time from days to hours while achieving high accuracy in identifying burned structures.

ABSTRACT

When major disaster occurs the questions are raised how to estimate the damage in time to support the decision making process and relief efforts by local authorities or humanitarian teams. In this paper we consider the use of Machine Learning and Computer Vision on remote sensing imagery to improve time efficiency of assessment of damaged buildings in disaster affected area. We propose a general workflow that can be useful in various disaster management applications, and demonstrate the use of the proposed workflow for the assessment of the damage caused by the wildfires in California in 2017.

Motivation & Objective

• Address the critical need for timely damage assessment in post-disaster emergency response.
• Reduce reliance on time-consuming manual mapping by humanitarian volunteers or experts.
• Develop a scalable, transferable method for automated building damage detection using remote sensing and deep learning.
• Demonstrate the effectiveness of fine-tuning a pre-trained model on limited new-region data to adapt to local conditions.
• Enable faster decision-making for relief operations through automated, accurate damage mapping.

Proposed method

• Use pre- and post-disaster satellite imagery pairs to train a deep learning model for change detection.
• Apply a U-Net-based convolutional neural network for semantic segmentation of damaged buildings.
• Create a training dataset by manually annotating damaged and undamaged buildings in pre- and post-event images.
• Fine-tune the pre-trained model on a small, region-specific dataset (146 burned, 137 unburned houses) to adapt to local data characteristics.
• Use the fine-tuned model to automatically segment damaged buildings across large areas with minimal human intervention.
• Validate results using a test subset of Santa Rosa, comparing performance before and after fine-tuning.

Experimental results

Research questions

• RQ1 Can a deep learning model trained on one disaster event be effectively adapted to a new geographic region with minimal retraining?
• RQ2 How does fine-tuning on a small, region-specific dataset improve damage segmentation accuracy compared to using a model trained on unrelated data?
• RQ3 To what extent can automated semantic segmentation reduce the time required for emergency damage mapping compared to manual methods?
• RQ4 How robust is the method to variations in image resolution, sensor type, and local building patterns across different regions?
• RQ5 Can the approach be generalized to other disaster types such as floods or earthquakes using similar training and fine-tuning strategies?

Key findings

• The proposed workflow reduced emergency mapping time from an estimated 1.5–2 days of manual work to under 3 hours, including model training and automatic processing.
• Fine-tuning the model on just 283 labeled houses (146 burned, 137 unburned) in Santa Rosa significantly improved segmentation accuracy, enabling clear separation of individual buildings.
• The model achieved high-quality segmentation results after fine-tuning, with distinct and accurate delineation of damaged buildings, outperforming the non-fine-tuned version which merged adjacent structures.
• The method demonstrated strong transferability: a model pre-trained on Ventura data was successfully adapted to Santa Rosa with minimal additional labeling effort.
• Semantic segmentation proved more robust than object detection for damage assessment, as it preserves spatial extent and enables area-based damage quantification.
• The approach is scalable and applicable to other disaster types, with potential for integration into GIS and real-time emergency mapping systems.