A Deep Learning Approach for Natural Catastrophe Loss Assessment through Remote Sensing

Abstract:

Losses after natural catastrophe events bring challenges to society. A quick and accurate loss assessment improves social resilience. Financial relief can help rebuild damaged structures, help families bear necessary living expenses, and help commercial entities cope with business interruption. The timeliness of delivery of financial relief can be enhanced through improved insurance claim processing operations, which are aided by the availability of damage information. Remote sensing allows quick access to the damaged area to enable the provision of such information. Remote sensing data are often raster images, e.g. multi-temporal coherence-based data from C-band Sentinel-1 (SAR) and normalized burn ratio-based data from Sentinel-2 (multispectral) imagery. Conventionally, explicit feature engineering such as zonal statistics can be used to detect the damages. However, deep learning, a collection of neural network architectures, provides a more unified framework to tackle this problem. We experimented with multiple natural catastrophe events, including tropical cyclone, flood, wildfire, to study how to leverage deep learning to better estimate insured losses. For instance, convolution neural networks can act as a generalization of zonal statistics of different neighborhood sizes. A multi-channel input allows a flexibility to overlay multiple sources, such as SAR and optical images, to interact and learn patterns across layers. U-Nets encode multiple levels of details. Pooling operators, such as max pooling, or sequence representations, such as recurrent neural networks, can be added on top to allow the aggregation over time of a collection of building sites under a policy. The model can also accommodate joint prediction on multiple response variables, such as different coverages of a policy. The output can be used for insurance loss reserving, strategic dispatchment of claims adjusters, and potentially as an input to risk assessment for future pricing of policies. Additional adjustments may be required for different natural catastrophes. For example, there is a higher likelihood of partial loss from wind damages versus full loss from fire events. Practical issues such as extended coverage and policy endorsements are also considerations for the practical application of the loss estimates.