February 3, 2025, 2:30pm, ENR2 S215
When
Title: Improving Inundation Estimates with Deep Learning and Satellite data
Abstract: Floods affect more people than any other hazard, and the frequency and
magnitude of exposure is growing with demographic, urban growth, and climatic
changes. Improved spatial extent of inundation for both historical records and
forecasting has multiple applications, from promoting underwriting for pricing new
insurance instruments, to providing evidence of flood exposure in the courtroom, and to
improve calibration of models for near real time and forecast applications.
This talk will give an overview of work in the Social Pixel Lab to advance the science
and application of inundation mapping with deep learning by combining diverse data
sources together, including satellites, physics-based models, and even human
experiences. I’ll first show our large labeled datasets across multiple sensors to support
deep learning flood mapping efforts. We demonstrate that using labels from commercial
data (Planetscope) can be leverage to train public sensors (Sentinel-1/2) and increase
accuracy by 15.6% compared to using only public sensor data for labeling. Second, Ill
show how we fused Sentinel-1 and MODIS sensors to generate an improved weekly
inundation history in Bangladesh (Giezendanner et al 2023) using a LSTM-CNN (Long
Short-Term Memory Network and Convolutional Neural network). The model results in
an R2 of 0.66 and consistently outperforms a baseline CNN only model. Third, Ill show
how adding in human experiences of inundation can increase inundation map accuracy
(up to 10% in our experiments in Rio Grande Valley, Texas) when introduced in deep
learning model training. Finally, I’ll give an example from my company, Floodbase,
which uses AI to fuse physically based inundation models with satellite observations
(Frame et al 2024). Training physical model inputs (e.g. soil moisture, terrain router)
from the National Water Model on high quality satellite-based maps using a CNN is
another way to produce gap-free inundation maps independent of satellite overpass or
cloud cover. Floodbase developed this CNN model, which produces 250m resolution
maps every three hours since 1979 across CONUS, and outperformed the National
Water Model’s HAND flood map approach when tested in the 2023 atmospheric river
floods in California (25% to 36% RMSE, respectively) (Frame et al 2024).
Relevant Papers: