11:00 AM - 1:00 PM
[AAS05-P02] Deep neural network based emulator of rainfall-runoff-inundation model with dimensionality reduction for avoiding overfitting
Keywords:d4PDF, rainfall-runoff-inundation, emulator, deep learning
Predicting spatial distributions of maximum inundation depth for individual rainfall events is important to mitigate hydrological disasters induced by extreme precipitation. Physics-based rainfall-runoff-inundation (RRI) models, a mainstream for predicting hydrological disasters, need massive computation resources to employ model simulations. This study aims to develop computationally-inexpensive deep learning (hereafter, Rain2Depth) by emulating an RRI model. This study focuses on the Omono river in Akita prefecture and emulates the prediction of the spatial distribution of maximum inundation depth from the spatial and temporal rainfall data for individual events.
The Rain2Depth is developed based on a deep convolutional neural network. We used hourly rainfall at 13 AMeDAS stations over seven days for input rainfall data, drawn from 50-ensemble of 30-years data from large-ensemble weather/climate predictions (d4PDF). The maximum inundation depth was simulated by Sayama et al. (2014)’s RRI model before the training. Firstly, we extracted the features in the input and output data with two dimensionality reduction techniques: principal component analysis and convolutional neural network (CNN) approach. This aims at avoiding the overfitting caused by lack of training data. The CNN approach is the optimal method for both input and output data. This is because this approach is useful for the extraction of the relationship between time series (input data) and spatial distribution (output data) with non-linear conversion by activation function. Next, the Rain2Depth is architected by connecting the features between input and output data with a neural network.
The prediction accuracy of the Rain2Depth was 18 cm for root mean square error (RMSE) and 0.95 for a coefficient of determination (γ) at a station with deep maximum inundation depth (hereafter, point A). The emulator provided more accurate predictions compared to our previous study (Kotsuki et al., 2020; RMSE = 30 cm, γ = 0.81) which used ensemble learning of multiple regularized regressions for a specific station. Kotsuki et al. (2020)’s approach needs to develop independent models for predicting the spatial distributions of maximum inundation depth. In contrast, this study enables to predict spatial distribution of maximum inundation depth only by training a single model, namely, Rain2Depth.
The Rain2Depth is developed based on a deep convolutional neural network. We used hourly rainfall at 13 AMeDAS stations over seven days for input rainfall data, drawn from 50-ensemble of 30-years data from large-ensemble weather/climate predictions (d4PDF). The maximum inundation depth was simulated by Sayama et al. (2014)’s RRI model before the training. Firstly, we extracted the features in the input and output data with two dimensionality reduction techniques: principal component analysis and convolutional neural network (CNN) approach. This aims at avoiding the overfitting caused by lack of training data. The CNN approach is the optimal method for both input and output data. This is because this approach is useful for the extraction of the relationship between time series (input data) and spatial distribution (output data) with non-linear conversion by activation function. Next, the Rain2Depth is architected by connecting the features between input and output data with a neural network.
The prediction accuracy of the Rain2Depth was 18 cm for root mean square error (RMSE) and 0.95 for a coefficient of determination (γ) at a station with deep maximum inundation depth (hereafter, point A). The emulator provided more accurate predictions compared to our previous study (Kotsuki et al., 2020; RMSE = 30 cm, γ = 0.81) which used ensemble learning of multiple regularized regressions for a specific station. Kotsuki et al. (2020)’s approach needs to develop independent models for predicting the spatial distributions of maximum inundation depth. In contrast, this study enables to predict spatial distribution of maximum inundation depth only by training a single model, namely, Rain2Depth.