Application of Kolmogorov-Arnold Networks in Drought Prediction: A Comparison with Artificial Neural Networks using SPEI

Abstract

The incidence of droughts has increased markedly in the recent years and their adverse impact has significantly disrupted the ecosystem. Therefore, there is a compelling need to address these alarming challenges. Drought prediction tools can aid in mitigating the effects of droughts. Artificial intelligence methods and data driven models have enhanced the predictive capabilities of the complex drought models. This study presents the application of Kolmogorov Arnold Network (KAN) model to predict the drought events using the Standardised Precipitation Evapotranspiration Index (SPEI). This model is compared to the Artificial Neural Network (ANN) model using the NASA POWER data for the drought prone regions of Bankura, Bagalkot, Mewat and Jaisalmer. The target variable SPEI is derived using the Pen Montieth equation for potential evapotranspiration and precipitation. A custom weighted Huber Loss function is used to emphasize the drought episodes in the prediction. Results demonstrates that both the models successfully capture the nonlinear relationships between the SPEI and the input variables. The developed models show a similar predictive capability with high RMSE (0.09-0.21) and R² (more than 0.94) values. The difference in the performances of the model is very negligible. This study highlights that KAN can be robust alternative to ANN for regional drought forecasting. The findings of the paper will help the agricultural community and the government agencies to monitor droughts and aid in mitigating its induced effects.