Improving Meteorological Drought Prediction in Tamil Nadu Through Weighted Dataset Construction and Multi-Objective Optimization

Abstract

Droughts typically develop gradually, and early prediction is crucial for the government to formulate effective mitigation plans. Our approach does not involve predicting specific drought index values. Instead, we forecast whether a particular year will experience drought. Insufficient investigation has been carried out regarding variations in additional climatic indicators like shortwave radiation, wind speed, sea level, and pollution in the context of droughts in the state of Tamil Nadu, India. In the study period taken from 1995 to 2020, only three years (2002, 2009, and 2017) experienced drought occurrences, resulting in an imbalanced dataset. To enhance the classification performance of this imbalanced dataset, a weighted dataset is constructed using a feature weighting approach known as the Single Objective Scorer (SOS) based Multi-objective PSO(MPSO) in conjunction with the Gradient Boosting Classifier. The proposed model facilitates objective-based multi-population formation and neighborhood learning. Precision and recall are crucial metrics, particularly in measuring imbalanced dataset classification performance. The application of multi-objective optimization techniques helps to strike a suitable balance between precision and recall. In addition to the Standardized Precipitation Index (SPI) and Standardized Precipitation Evapotranspiration Index (SPEI), 14 climatic indicators based on land, atmosphere, and sea are utilized. By employing the weighted dataset created with SOS-based MPSO, a significant improvement in recall value of 0.81 is achieved. Based on the weights assigned to the features, it is identified that the Mean Sea Level of the Arabian Sea and CO2 are significant indicators for predicting meteorological drought. The Explainable AI techniques SHAP and LIME are employed for interpreting the drought prediction model, providing insights into its workings.

Keywords

• Particle swarm optimization algorithm
• multi-population
• climatic indicators
• imbalanced dataset
• pollution
• mean sea level