Benchmark Data Development to Classify Damage for Natural Disaster Relief Efforts
Abstract
In the aftermath of natural disasters, there is a need to assess damage across the impacted regions to inform stakeholders in estimating loss and planning response and recovery. The objective of this research is to improve the automated classification of multiple damage classes from overhead imagery to inform natural disaster response and recovery. To reach this objective, we propose to first build and optimize a benchmark image library of damage features for testing and training automated supervised classification algorithms and second to evaluate the use of heterogenous data fusion for feature extraction and object classification using automated supervised classification algorithms. Supervised classification requires benchmark training data and to date, the majority of benchmark datasets are event-specific. The proposed effort will build a benchmark image library of damage for use across events and available for future events. NGA’s recent release of the xView dataset (Lam et al., 2018) which includes a collection of satellite images with multiple labels including “demolished building” and “intact building” is one such benchmark image library. Our own research in liquefaction evaluation (Zhu et al., 2017; Baise and Rashidian, 2018) has resulted in a temporal and geospatial database of liquefaction occurrence which if matched with imagery could be a parallel image library for liquefaction. Similar temporal and geospatial databases exist for other earthquake related damage (e.g. landslides: Nowicki et al. 2014; 2018). In the proposed effort, we will focus on evaluation of existing benchmark image libraries (buildings: xView), and creation of new benchmark image libraries (e.g. ground failures; lifeline infrastructure) optimized for training data for the supervised classification of building damage, infrastructure damage, and ground failure due to earthquakes across events. Heterogeneous data fusion will combine very high resolution (VHR) optical imagery, VHR SAR, and geospatial data layers such as those provided by the OpenStreetMap (OSM) Initiative and digital elevation models and use these fused layers as input into supervised classification to improve classification accuracy. The combination of VHR optical and SAR have shown higher classification accuracy; however, the addition of geospatial data with imagery has been even more promising. Once we have evaluated and defined the benchmark image libraries, we will use heterogeneous data fusion across VHR optical imagery, VHR SAR, and geospatial datasets to improve classification accuracy in automated supervised classification. We will use our content knowledge on the causes of earthquake-related damage to pair appropriate geospatial datasets with imagery and improve classification accuracy. Access to the proposed benchmark imagery library will increase the rapid aspect of automated damage detection. The proposed supervised classification approach with data fusion across VHR optical, SAR, and geospatial will inform future efforts in classification by providing insight into desired input features and ultimately increase classification accuracy.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 28, 2020
- Source ID
- HM04762010006
Entities
People
- Laurie Baise
Organizations
- National Geospatial-Intelligence Agency
- Tufts University