Repo for flood prediction using LSTMs and HAND

Abstract

Every year, floods cause billions of dollars’ worth of damages to life, crops, and property. With a proper early flood warning system in place, decision-makers can take the necessary steps to prevent or at least mitigate the damage caused by floods. Although various flood prediction models exist, a majority of them fail to be fast, reliable, and detailed simultaneously. Our proposed system presents a novel hybrid flood prediction model using Long Short Term Memory(LSTM) for multivariate time series forecasting of water depth based on meteorological conditions and Height Above Nearest Drainage(HAND) to predict river stage in real-time and map the inundated areas for the corresponding water depth using enhanced HAND. Unlike traditional flood forecasting models, this hybrid approach is resource efficient and easy to implement making it highly practicable for real-time flood inundation mapping.

Methodology

The proposed system prioritizes quick development and real-time predictions without compromising on the accuracy. A range of factors affect the occurrences of riverine floods. However, climatological conditions are the major driving force behind them. Factors such as land use/land change and deforestation, although important, only affect flooding in the watershed over a long period of time. Hence, the proposed system used only meteorological conditions and DEM rasters for predicting floods over the next few days.

The relation between weather conditions and flood inundation is simplified by breaking the system into two modules. The first module being estimation of river stage height and the second one being flood inundation mapping. The system uses LSTMs, a data-driven empirical approach, to model the dependence of stage height on meteorological data and HAND, a simplified conceptual approach, to generate flood inundation maps based on the terrain of the watershed and the river stage height predicted by the first module.

Modules :

1. Inundation Mapping - HAND algorithm to map inundated areas for a given stage height(as proposed in this paper).
2. River Stage Estimation - Recurring neural networks (LSTMs) to predict the maximum stage height based on weather conditions of the last 3 days.
3. Deforestation Analysis - Land use classification to identify the changing features of the area over time and identify the areas affected by deforestation.

Datasets

The proposed system uses different data for the three modules. Each of these are collected from different sources and processed separately. The module-wise requirements of data are as follows :

1. Inundation Mapping:
   1. Digital Elevation Maps from United States Geological Survey
2. River Stage Estimation:
   1. Meteorological data from National Climatic Data Center
   2. River stage height data from United States Army Corps of Engineers’ river gage data.
3. Deforestation Analysis:
   1. Satellite images - Landsat 8, Landsat 5 from USGS Earth Explorer

Results

Stage Height Estimation

We tested our proposed system for Cedar Rapids, Iowa. Our experiments showed that features such as vegetation and soil type have little effect on short term flooding and can be disregarded for the prediction module. Testing multiple models showed that single output LSTM models perform better than single shot models. These models are stable upto lead times of 4 days with a Nash-Sutcliffe Efficiency greater than 0.5.

Flood Mapping

Each pixel of the inundation map raster is compared with a reference map created by ground-truthing to identify how many points were incorrectly classified as not flooded. The red areas in the image depict false negatives generated by the proposed system.