The increasing frequency and intensity of droughts in Australia have significantly impacted the agricultural sector, economy, and rural communities, resulting in losses of billions of dollars and considerable social distress. Therefore, accurate rainfall prediction is crucial for sustainable water resource management, climate policy, and decision-making. Several studies have experimented with machine learning and deep learning models to accurately predict rainfall under diverse weather conditions. However, in places such as Australia, where the climate is uncertain, determining the most effective method for modeling complex rainfall processes is challenging. This study presents a geo-adaptive multi-layer stacking-based predictive architecture to support data-driven agricultural decision-making by providing accurate location-specific rainfall forecasts. The proposed model utilizes a dataset of daily weather observations over a decade from 49 locations across Australia. A rigorous data preprocessing pipeline was implemented that involved data cleaning, handling of missing values, and engineering of informative features. A consensus-based feature selection methodology validated across three independent importance measures identified Humidity3pm as a key predictor, with an importance score of 0.140 for rainfall in the region. The proposed model intelligently combines diverse machine learning models within an optimized framework. This framework comprises a base layer of classifiers, meta-learner, and geospatial module for location-specific prediction. To evaluate the model’s performance, we used an unseen test set from nine geographically distinct locations and several key metrics. The results demonstrated that GASB consistently outperformed traditional machine and deep learning models across various key metrics, including AUC, accuracy, precision, and Cohen’s kappa.
Mehla et al. (Mon,) studied this question.