H197-0014
Prediction of flood claims over the contiguous United States (CONUS) by building a classification/regression-hybrid machine learning scheme

Wednesday, 16 December 2020
Poster
Qing Yang1,2, Xinyi Shen2, Feifei Yang3, Kang He3, Emmanouil N Anagnostou2, Hojjat Seyyedi4, Jack Eggleston5 and Albert Kettner6, (1)Guangxi University, College of Civil Engineering and Architecture, Nanning, China, (2)University of Connecticut, Civil and Environmental Engineering, Storrs, CT, United States, (3)University of Connecticut, Civil and Environmental Engineering, Groton, CT, United States, (4)Swiss Re America Holding Corporation, Schaumburg, IL, United States, (5)U.S. Geological Survey, Hydrologic Remote Sensing Branch, Leetown, WV, United States, (6)University of Colorado Boulder, INSTAAR, Boulder, CO, United States
Abstract:
Flood hazard causes an enormous number of house claims at billions-of-dollar scale payout each year over CONUS. The prediction of the claim number plays a critical role for promoting the preparedness in the changing climate. We propose a machine learning prediction scheme using event-wise flood predictors derived from Sentinel-1-derived inundation archive1-3, precipitation fields, as well as static topographic, land use, morphologic4,5 and building location. Flood events are automatically delineated by integrating USGS flooded stages, the IMERG precipitation6, and NOAA tidal water levels. To improve the performance in the presence of unevenly distributed flood claims in both the intensity and sample size, we build a hybrid scheme consisting of two steps, damage-level classification and claim number regression. To balance the sample sizes of the minority (rare-high damage samples) and majority (frequent-low or zero damage samples), we adopt the Borderline Synthetic Minority Over-sampling Technique (B-SMOTE), combined with the Iterative-Partitioning Filter (IPF) in the classification, and develop a Balance to the Target Level (BTL) resampling technique in the regression. Three random forest models and a generalized linear model (GLM) are trained for the low to high classes and the very high class respectively. We evaluate the model with 0.1° grids with a sample size of 410,002 in 1,760 flood events occurred since 2016 over the CONUS, which overlaps 286,891 claims out of the total 299,818 National Flood Insurance Program (NFIP) records. Our 50 times of 5-fold cross-validation yields acceptable performance with R2 at grid/event level, county/events level and county cumulative level of 0.54, 0.91 and 0.95 (Figure 1), respectively. We conclude that the proposed methods can facilitate various applications, including flood damage and risk assessment, and preparedness improvement.

Reference:

  1. Yang Q, Shen X et al. Bulletin of American Meteorological Society 2020; (under review).
  2. Shen X, Anagnostou EN et al. Remote Sens. of Environ. 2019; 221: 302-35.
  3. Shen X, Wang D, et al. . Remote Sens. 2019; 11(7): 879.
  4. Shen X, Vergara HJ, et al. Environ. Modell. Softw. 2016; 83: 212-23.
  5. Shen X, Anagnostou EN, et al. Sci. data 2017; 4: 160124.
  6. Huffman GJ, Bolvin DT, et al. IMERG ATBD, version 2015; 4: 30.

<< Previous Abstract | Next Abstract >>