Generative Adversarial Networks Enhanced Machine Learning for Climate Prediction

Abstract

Investigating climatology and predicting climate variables such as rainfall precipitation, temperature, humidity etc., plays a critical role in shaping policy decisions related to climate change mitigation and adaptation, and is key to developing effective strategies to address the challenges of a changing climate. In this study, four different machine learning regression models were used to predict rainfall precipitation in Lagos State, Nigeria. These models are Multiple Linear Regression, Decision Tree, Support Vector Machine (SVM), and Random Forest. The historical weather dataset used in this study, were collected from the Nigeria Meteorological Agency (NIMET), Lagos spanning 30 years (1991 2018), to generate more data, for better model accuracy and performance, this data was used to generate synthetic data, using the Generative Adversarial Networks (GANs). The inputs used in the proposed model were climate variables, minimum and maximum temperatures. Based on the R squared metric, the performance of the machine learning algorithms was evaluated, it was observed that the Random Forest algorithm performed best for both the real world dataset and the synthetically generated dataset with R square values of 0.873 and 0.7518 respectively. Multiple Linear Regression and Decision Tree algorithms also performed relatively well, while SVM performed poorly for both datasets. The results of this evaluation can be used to select the best machine learning algorithm for predicting rainfall precipitation in Lagos State, Nigeria.