[Paper Review] From Satellite Imagery to Disaster Insights

Jigar Doshi, Saikat Basu+1 more|arXiv (Cornell University)|Dec 17, 2018

Remote-Sensing Image Classification11 references49 citations

TL;DR

The paper introduces a Disaster Impact Index (DII) that quantifies disaster impact by measuring change in man-made features (roads, buildings) detected by pre-trained CNN segmentation on pre- and post-disaster satellite images, showing strong correlation with ground truth across hurricanes and fires.

ABSTRACT

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.

Motivation & Objective

• Motivate fast, scalable disaster mapping from satellite imagery to guide rescue and relief operations.
• Leverage time-series, high-level man-made features to detect area-wise change invariant to season and lighting variations.
• Develop a scalable method that avoids training CNNs specifically for damage detection by using pre-trained road/building models.
• Introduce a quantifiable Disaster Impact Index (DII) to aggregate change into actionable grids for prioritizing response.

Proposed method

• Use a pre-trained semantic segmentation model (Residual Inception Skip network) to extract roads and buildings from pre- and post-disaster images.
• Compute a pixel-wise change mask between pre- and post-disaster segmentation masks and apply dilation and de-noising to reduce noise.
• Partition the region into n×n grids (256×256 in experiments) and compute Disaster Impact Index (DII) as a normalized change demand over grids (Equation provided).
• Threshold DII at τ = 0.01 to identify regions of maximal impact and cluster these regions to guide relief efforts.
• Train on combined Spacenet and DeepGlobe datasets, and validate on Hurricane Harvey flood and Santa Rosa fire datasets using pixelwise and gridded ground-truths.

Experimental results

Research questions

• RQ1 Can change in high-level man-made features detected by pre-trained CNNs indicate disaster impact without training damage-specific detectors?
• RQ2 Does the proposed Disaster Impact Index (DII) correlate with actual disaster impact areas across different disaster types and resolutions?
• RQ3 Is the method scalable and generalizable across feature types and diverse satellite imagery datasets?

Key findings

• DII-based road change predictions achieve F1 of 81.2% and IoU of 68.3% on the Harvey flood gridded data.
• DII-based road change predictions achieve F1 of 83.5% and IoU of 71.7% on the Santa Rosa fire gridded buildings data.
• Pixelwise change predictions for roads and buildings yield lower F1 and IoU compared to DII-based gridded assessments, showing the value of aggregation.
• There is a positive correlation between predicted disaster areas (via DII) and ground-truth disaster impact areas for both flood and fire cases.