Predicting Seismic Death using Boosted Random Forest with Ancillary Data

High resolution, instant data on seismic casualty distribution are crucial for earthquake emergency response and further disaster relief planning. At present various approaches are used to disaggregate statistics data and predict seismic casualty for finer scale, gridded data sets. We propose a new semi-automated modeling method that incorporates detailed statistics and ancillary data in a flexible, “Boosted Random Forest (BRF)” estimation technique. By calculating the importance of the original 30-dimensional ancillary data, we integrate intensity-based seismic death statistics and a modified ancillary dataset into a BRF algorithm to make seismic death prediction, which produces a map for seismic death with 1 km spatial resolution. As a case study we evaluate the BRF algorithm and its products for four earthquake cases with other common seismic death data production methodologies. We discuss the advantages of the BRF method and improvements over the flexibility and accuracy of other machine learning approaches. Finally, we outline the limitations of the BRF algorithm and its continuation to other earthquake disaster relief tasks.