[Paper Review] Prediction of Rainfall in Rajasthan, India using Deep and Wide Neural Network

Vikas Bajpai, Anukriti Bansal+2 more|arXiv (Cornell University)|Oct 22, 2020

Hydrological Forecasting Using AI4 citations

TL;DR

This paper proposes a Deep and Wide Rainfall Prediction Model (DWRPM) that combines convolutional feature extraction with a multi-layer perceptron (MLP) to predict daily rainfall in Rajasthan, India using 71 years of historical rainfall data. The model achieves superior performance over MLP, LSTM, and 1D-CNN, with a mean absolute error (MAE) of 0.7765 and root mean square error (RMSE) of 2.1716 across 158 rain-gauge stations, demonstrating strong generalization across diverse geographical zones.

ABSTRACT

Rainfall is a natural process which is of utmost importance in various areas including water cycle, ground water recharging, disaster management and economic cycle. Accurate prediction of rainfall intensity is a challenging task and its exact prediction helps in every aspect. In this paper, we propose a deep and wide rainfall prediction model (DWRPM) and evaluate its effectiveness to predict rainfall in Indian state of Rajasthan using historical time-series data. For wide network, instead of using rainfall intensity values directly, we are using features obtained after applying a convolutional layer. For deep part, a multi-layer perceptron (MLP) is used. Information of geographical parameters (latitude and longitude) are included in a unique way. It gives the model a generalization ability, which helps a single model to make rainfall predictions in different geographical conditions. We compare our results with various deep-learning approaches like MLP, LSTM and CNN, which are observed to work well in sequence-based predictions. Experimental analysis and comparison shows the applicability of our proposed method for rainfall prediction in Rajasthan.

Motivation & Objective

• To develop a robust rainfall prediction model for Rajasthan, an arid region highly dependent on monsoon rains and vulnerable to droughts and floods.
• To improve upon existing deep learning models by integrating wide and deep learning architectures for better generalization across diverse geographical zones.
• To evaluate the model’s performance on historical daily rainfall data spanning 1957–2017 from 33 districts across Rajasthan.
• To compare the proposed DWRPM with standard deep learning baselines such as MLP, LSTM, and 1D-CNN for sequence-based rainfall prediction.
• To assess the model’s ability to generalize across different atmospheric zones, including plains, plateaus, and desert regions.

Proposed method

• The DWRPM uses a wide component that applies a 1D convolutional layer to extract temporal patterns from raw daily rainfall sequences, replacing direct input of rainfall intensity values.
• The deep component employs a multi-layer perceptron (MLP) to model complex non-linear relationships in the extracted features.
• Geographical coordinates (latitude and longitude) are embedded and integrated into the model in a unique way to enhance generalization across different regions.
• The model is trained on 71 years (1957–2017) of daily rainfall data from 158 rain-gauge stations across Rajasthan, with data split into training and test sets.
• Performance is evaluated using MAE and RMSE on both zone-wise and station-wise predictions to assess generalization and regional accuracy.
• The model is compared against baseline deep learning models: MLP, 1D-CNN, and LSTM, all trained under the same experimental setup.

Experimental results

Research questions

• RQ1 Can a deep and wide neural network architecture effectively predict daily rainfall in Rajasthan using only 71 years of historical rainfall data?
• RQ2 How does the integration of geographical coordinates improve the model’s generalization across diverse climatic zones in Rajasthan?
• RQ3 Does the DWRPM outperform standard deep learning models like MLP, LSTM, and 1D-CNN in terms of prediction accuracy for daily rainfall?
• RQ4 To what extent can a single trained model generalize across different atmospheric zones—such as desert, hill, plain, and plateau regions—in Rajasthan?
• RQ5 What is the model’s performance on extreme rainfall events, particularly heavy and very heavy rainfall?

Key findings

• The DWRPM achieved a mean absolute error (MAE) of 0.7765 and root mean square error (RMSE) of 2.1716 across all 158 rain-gauge stations in Rajasthan, outperforming MLP, LSTM, and 1D-CNN.
• The model demonstrated strong generalization, achieving MAE values of 0.43–0.67 and RMSE values of 0.60–1.59 across four randomly selected rain-gauge stations from different atmospheric zones.
• Zone-wise performance showed the lowest error in the North-West Desert region (MAE=0.679, RMSE=1.512), with slightly higher errors in the South-Eastern Plateau (MAE=0.903, RMSE=2.613).
• The DWRPM outperformed all baseline models, with the next best being 1D-CNN (MAE=0.8406, RMSE=2.2894), followed by LSTM (MAE=0.8750, RMSE=2.3095), and MLP (MAE=1.3137, RMSE=2.7808).
• The model’s performance was consistent across diverse geographical conditions, indicating effective generalization due to the integration of latitude and longitude as input features.
• While effective for light and moderate rainfall, the model shows room for improvement in predicting extreme rainfall events such as heavy and very heavy rainfall.