Source-linked AI summary
From Satellite Imagery to Disaster Insights
Jigar Doshi, Saikat Basu, Guan Pang
TL;DR
The paper targets slow and error-prone manual disaster mapping and the need to identify areas of maximum damage efficiently. It proposes CNN-based change detection on pre- and post-disaster satellite imagery, followed by gridding and Disaster Impact Index computation. The framework shows high correlation with ground-truth impact areas across Hurricane Harvey flooding and the Santa Rosa fire, while reporting top F1 scores of 81.2% and 83.5% in the abstract.
Problem
Manual disaster mapping is time-consuming and can produce erroneous results, motivating faster identification of severely affected areas.
Source: §1 · p. 1Method
The framework compares CNN-derived man-made-feature masks before and after disasters, denoises and grids the changes, then computes and thresholds the Disaster Impact Index.
Source: §2.1 · p. 2, §2.3 · p. 3Results
The method shows high correlation between predicted impact areas and ground-truth labels for the Hurricane Harvey flood and Santa Rosa fire.
Source: §3.3 · p. 4, §4 · p. 5Takeaways & Limitations
The approach identifies regions of priority for coordinating disaster relief efforts using changes in satellite-derived man-made features.
Source: §1 · p. 1, §4 · p. 5The study focuses on roads and buildings, although the framework could be extended to other natural and man-made features.
Source: §4 · p. 5
Abstract
from arXiv · showhide
The use of satellite imagery has become increasingly popular for disaster monitoring and response. After a disaster, it is important to prioritize rescue operations, disaster response and coordinate relief efforts. These have to be carried out in a fast and efficient manner since resources are often limited in disaster-affected areas and it's extremely important to identify the areas of maximum damage. However, most of the existing disaster mapping efforts are manual which is time-consuming and often leads to erroneous results. In order to address these issues, we propose a framework for change detection using Convolutional Neural Networks (CNN) on satellite images which can then be thresholded and clustered together into grids to find areas which have been most severely affected by a disaster. We also present a novel metric called Disaster Impact Index (DII) and use it to quantify the impact of two natural disasters - the Hurricane Harvey flood and the Santa Rosa fire. Our framework achieves a top F1 score of 81.2% on the gridded flood dataset and 83.5% on the gridded fire dataset.
1 Introduction
The paper addresses the limits of manual and snapshot-based disaster mapping by using change in satellite-derived man-made features to locate areas of maximal damage and prioritize response.
- CNN-based change detection addresses costly, non-scalable damaged-building annotation requirements in prior disaster-mapping approaches.Existing methods also face seasonal, lighting, and noise differences that make direct RGB comparison unreliable.Source: §1 · p. 1
- The framework compares pre- and post-disaster changes in roads and buildings to identify areas of maximal damage for relief prioritization.It uses models trained on general feature datasets rather than disaster-specific damaged-feature datasets.Source: §1 · pp. 1–2
- Using high-level man-made features makes the approach invariant to season, lighting, and noise differences in time-series imagery.Source: §1 · pp. 1–2
- The approach was evaluated on human-annotated and state-provided disaster-impact datasets and showed a strong positive correlation with actual impacted areas.Source: §1 · p. 2
2 Proposed Approach
The proposed approach extracts man-made features before and after a disaster, computes normalized change on grids, and thresholds the resulting Disaster Impact Index to identify severely affected regions.
2.1 Overview
- The pipeline compares pre- and post-disaster semantic-segmentation masks, denoises their change mask, and computes DII over n×n grids.It dilates the pre-disaster mask by 5 pixels and removes connected components smaller than 1000 pixels before grid aggregation.Source: §2.1 · p. 2, Figure 1 · p. 2
2.2 Model Architecture
- A Residual Inception Skip network provides the semantic segmentation model for extracting roads and buildings from satellite imagery.The encoder-decoder uses inception modules, asymmetric convolutions, batch normalization, leaky ReLUs, and continuous Dice loss.Source: §2.2 · p. 2
2.3 Disaster Impact Index (DII)
- DII counts pre-disaster feature pixels absent after the disaster within each grid, normalized by the region-wide average pre-disaster feature-pixel count.The formulation focuses on pre-disaster detections that disappear post-disaster because the reverse change is usually not disaster-caused.Source: §2.3 · p. 3
- Normalization makes DII independent of region and feature, allowing comparison across scenarios and use of a common threshold.The analysis uses 256×256 grids and τ = 0.01 to infer regions of maximal impact.Source: §2.3 · p. 3
3 Experiments
Experiments train on public satellite-segmentation datasets, validate across flood and fire disasters, and evaluate pixelwise and gridded impact detection against ground truth.
3.1 Training Datasets
- The models were trained on SpaceNet and DeepGlobe imagery, with Planet Labs imagery additionally used to train the roads model.The datasets include building footprints, roads, buildings, and land-cover labels at high and lower resolutions.Source: §3.1 · p. 3
3.2 Validation Datasets
- Validation covers Hurricane Harvey flooding and the Santa Rosa fire, pairing roads with floods and buildings with fires across DigitalGlobe and Planet Labs imagery.The areas span approximately 143 km² near Sugar Land and 120 km² near Santa Rosa.Source: §3.2 · p. 3
- Ground truth combines pixel-wise and grid-based human annotations, with FRAP fire-impact labels used instead of human annotations for Santa Rosa.The grid annotations target affected areas rather than relying only on pixel metrics.Source: §3.2 · p. 3
3.3 Evaluation Results
- Evaluation compares pixelwise change detection, thresholded DII on 256 × 256 grids, and DII against impact-area labels using τ = 0.01.All three settings are evaluated for both disasters against corresponding ground-truth labels.Source: §3.3 · p. 4
- F1 scores show high correlation between CNN-based change masks and ground truth for both pixelwise and DII-based approaches in the flood and fire evaluations.The corresponding quantitative results are reported in Table 1 for Hurricane Harvey and Table 2 for Santa Rosa.Source: §3.3 · p. 4, Table 1 · p. 4, Table 2 · p. 4
4 Conclusion and Future Work
The paper uses CNN-based change in man-made features as a proxy for disaster impact and derives DII-based priority regions, while noting its current focus on roads and buildings.
4 Conclusion and Future Work
- CNN segmentation across before-and-after imagery detects structural change in man-made features, which is clustered and thresholded into the Disaster Impact Index for prioritizing relief efforts.The experiments report high correlation between predicted impact areas and ground-truth labels.Source: §4 · p. 5
- The current approach focuses on roads and buildings, although the authors state that it can extend to other natural and man-made features.This defines the present feature scope and the proposed direction for future extension.Source: §4 · p. 5