Transfer Learning-Based Ensemble Approach for Rainfall Class Amount Prediction

Abstract

Predicting short-term precipitation amounts is challenging, especially due to meteorological data scarcity. While deep learning-based models have been shown to be more effective in predicting precipitation amounts, their performance heavily relies on the size of the training datasets. This paper presents a multi-station-based transfer learning ensemble approach to mitigate the data scarcity problem by transferring knowledge learned from multiple meteorological station datasets to a single target station. To achieve this, multi-layer perceptron, convolutional neural networks, and long-short-term memory (LSTM) systems were trained on weather station datasets from the Lake Victoria Basin (LVB). From the experiments, the LSTM model outperformed other state-of-the-art models achieving high F1 scores across individual stations. Fine-tuning pre-trained models for the target station demonstrated improved accuracy, with performance gains of up to 5%. Additionally, the ensemble of these models further enhanced performance, delivering highly accurate classification results. Summarily, the proposed ensemble approach demonstrates significant improvements in predicting rainfall class amounts, offering a robust solution for precipitation forecasting in data-scarce regions like the LVB.