Deep Learning for Unmonitored Water Level Prediction and Risk Assessment

This proposed research is a match project designed to run in tandem between the Mid-America Transportation Center (MATC) and the Missouri Department of Transportation (MoDOT). It uses deep learning and other computational intelligence methods to leverage public geospatial data and historical NOAA data to develop forecasting tools to create virtual water level monitors. These tools inform existing models developed in previous MATC/MoDOT projects for flood prediction and models develop by the USGS, FEMA, NOAA, and others and are used to reduce the errors from these models due to sparse data for prediction. The project scope includes a survey instrument to gather data from first responders who are required to travel during these hazardous events. These data are then used to determine the water levels and rate of change at unmonitored sites based on projected rainfall totals based on drainage basin information and recent weather patterns. The data from these virtual monitors is then used for flood event prediction to improve accuracy. The results of these virtual monitors will be validated by manual testing at prediction locations. In addition, the data from the virtual monitors and the validation readings will be used to determine the sources of uncertainty in the predictions and recommend where physical monitors should be placed to improve future predictions. This provides the transportation safety or disaster planner increased accuracy to better plan for flooding events.

The lack of available data is a significant barrier to accurate and timely predictions of flooding and particularly flash flooding events. This makes it impossible for the current alert systems in place to provide localized updates, requiring the issuing of warnings to general areas spanning large time periods to insure that some warning is provided. This lack of real-time rate of water rise information to support accurate reporting of these flooding events can impede first responders and endanger the travelling public. Deep learning methods can provide decision analytics and tools capable of ingesting historic flood pattern data and information in existing hydrographic models, but it was found that the lack of monitoring sites to provide current data decreases their efficacy. Hybrid computational intelligence techniques will be used to create virtual monitors to provide the water levels at identified locations. These estimates will be verified at selected locations by manual measurements taken to coincide with the water level predictions. The locations for the water level prediction will be determined from input solicited from Missouri Highway Patrol Officers and other first responders through a survey instrument. Combined, this will provide a tool to give more precise and timely warnings for flood events.

This proposed project will result in a series of tools and protocols based on computational intelligence and deep learning methods. These materials will promote safety and economic viability of the surface roadways by providing improved real-time processes for flash flood prediction that will be coupled with alerts and driver rerouting schemas. These processes are data driven and use geospatial data and information to fully model areas categorized as high risk for flooding.

Download the Final Report