AI-Driven Approach Outperforms Traditional Models in Monsoon Forecasting

By Samudrapom DamReviewed by Susha Cheriyedath, M.Sc.Jan 25 2024

In an article recently published in the journal Scientific Reports, researchers proposed an artificial intelligence (AI)-based approach to predict normal Indian summer monsoon rainfall in 2023.

Limitations of physical models

An accurate all Indian summer monsoon rainfall (AISMR) forecast is critical for decision-and policy-making due to its applications in several socioeconomic sectors, such as energy, agriculture, and water resources. Physical models are conventionally used to develop seasonal forecasts, which provide a probabilistic prediction with an ensemble spread indicating the uncertainty level associated with the event.

However, physical models possess limited flexibility in input parameters and are computationally expensive and highly sensitive to parameter initialization. Moreover, the physical models’ forecast skill depends on simulating the relationship between predictands and predictors and the ocean-atmospheric coupling and feedback processes.

The potential of AI-based approaches

Recently, machine learning (ML) and AI have gained significant attention as promising and effective techniques for predicting atmospheric and ocean variations. These techniques can analyze large amounts of model and observational data and identify emerging patterns more efficiently compared to conventional physical models.

Data-driven methods, specifically ML and neural models, have been utilized in several studies for forecasting rainfall owing to their ability to capture complex and non-linear relationships, flexibility in input parameters, and better forecast ability than physical models.

The slowly varying ocean-atmosphere conditions related to tropical climate phenomena, including the Indian Ocean Dipole (IOD) and El Niño-Southern Oscillation (ENSO), play a critical role in ISMR’s interannual variability and are commonly considered the key predictability sources for the seasonal forecast. Thus, a more actionable and reliable seasonal forecast for AISMR/AISMR forecasts with improved accuracy can be developed by leveraging advanced ML and AI techniques that utilize these predictability sources.

The proposed approach

In this study, researchers proposed a data-driven approach by incorporating the major drivers of AISMR’s interannual variability, including IOD and ENSO, and employing multiple empirical models to reliably forecast the AISMR of 2023. The study's objective was to develop a suitable alternative to predictions using physical models.

A set of successive AISMR forecast experiments were performed by employing several deep learning models, including long short-term memory (LSTM) and convolutional network networks (CNN), and statistical ML models, including autoregressive integrated moving average (ARIMA), seasonal-ARIMA (SARIMA), extreme gradient boosting (XGBoost), and support vector regression (SVR), using three datasets, including Niño3.4 index, historical AISMR, and categorical IOD data (AISMRNiñoIOD), Niño3.4 index and historical AISMR (AISMRNiño), and only historical AISMR (AISMR), for a range of lookback windows.

In every set of experiments, the model performance was validated using Spearman correlation and root mean squared error (RMSE) percentage for the 2002 to 2022 test period by comparing predictions of the model and observed AISMR values from the Indian Meteorological Department (IMD).

Research findings

Models trained using only the AISMR dataset underperformed with a lower Spearman correlation of 0.2 to 0.4 and a higher RMSE percentage of 0.8 to 1.1 when compared with the models trained using the AISMRNiño dataset, which showed a higher Spearman correlation of 0.67 and lower RMSE percentage.

However, the models trained using the AISMRNiñoIOD dataset demonstrated the best results with 0.71 Spearman correlation and 0.64 RMSE percentage. A 5-year lookback period yielded better outcomes for most of the evaluated models in comparison to other lookback periods.

Moreover, LSTM produced the best results among the different models trained in this study. The lowest RMSE percentage and highest Spearman correlation were realized using the LSTM model as it could capture the most relevant characteristics of a series and disregard irrelevant parts selectively.

The incorporation of IOD and ENSO-related information into the AISMR dataset improved the model performance, which indicated that data-driven models could effectively identify the non-linear interaction of various drivers and their combined teleconnections.

Thus, the LSTM model trained using the AISMRNiñoIOD dataset with a five-year lookback period was utilized for AISMR prediction in 2023. During the test period from 2002–2022, the LSTM model demonstrated a successful forecast rate of 61.9%, which was substantially higher compared to the 28.57% successful forecast rate of physical models used by IMD. This significant increase in success rate realized using a data-driven model could assist the rainfall forecasting systems transition from conventional models to novel data-driven models.

The LSTM model predicted that India will receive a normal 790 mm rainfall/AISMR in 2023, typical of a normal monsoon year. However, this predicted rainfall was close to the normal and below-normal category threshold, necessitating cautious measures to prevent adverse impacts of a below-normal monsoon season.

In conclusion, the findings of this study effectively demonstrated the significant forecast skill displayed by data-driven models for seasonal AISMR prediction.