Transfer Learning Approach for Rainfall Class Amount Prediction Using Uganda's Lake Victoria Basin Weather Dataset

Abstract

Predicting short-term precipitation amounts is challenging especially due to meteorological data scarcity. Although deep Learning-based models have proven to be more effective in predicting precipitation their performance heavily depends on the size of the training dataset. This paper presents a Multi-station-based Transfer Learning approach with the aim of mitigating the data scarcity problem by transferring knowledge learned from multiple meteorological stations to a target station. In order to achieve this, a Multi-layer Perceptron, Convolutional Neural Networks, and Long-Short Term Memory systems were trained to predict rainfall class amounts on individual weather stations. From the experiments LSTM model outperformed the other-state-of-the-art models with an F1-score of 93% for sample stations of Jinja and Mwanza, and 95% for Musoma respectively. Consequently, the pre-trained LSTM model on each station were used as base models for Transfer Learning on the target station of Kisumu with limited data. The results show that the performance of the resulting transfer learning model improved by 3% for Jinja, Mwanza, and Musoma after model fine-tuning.