Earth skin temperature long-term prediction using novel extended Kalman filter integrated with Artificial Intelligence models and information gain feature selection

Abstract

Predictions of Earth skin temperature (EST) can provide essential information for diverse engineering applications such as energy harvesting and agriculture activities. Several synoptic climate parameters influence EST, and its prediction and quantification is highly complex and challenging. The current research uses three different machine learning (ML) techniques—the integrated Extended Kalman Filter with Artificial Neural Network (EKF-ANN), standalone ANN, and Adaboost—to model EST at three locations with a tropical environment in the Malaysian region. Five predictors, including minimum and maximum air temperature, humidity, wind velocity at 10 m, and periodicity (month and day) information, are used for the modelling development. Different input combinations are constructed based on the statistical correlation and information gain (mutual information). The developed EKF-ANN model showed superior predictability performance compared to the ANN and Adaboost models. The superiority of the EKF-ANN model prediction was observed for the three investigated locations. In addition, the research findings confirmed that building the predictive models based on a limited climate dataset such as minimum and maximum air temperature can provide a substantial prediction matrix. Overall, the research offered insightful results on EST prediction for several locations of a tropical environment.