Spatio-Temporal Modelling of Temperature Data using Machine Learning

Abstract

The scarcity of weather stations in many developing countries poses a significant obstacle to obtaining precise temperature data. Akwa Ibom State in Nigeria faces this challenge leading to reliance on satellite weather data (NASA POWER) and ground-truth measurements for temperature predictions. The study helps to predict temperatures in locations without direct measurements by developing and evaluating a machine learning framework that integrates satellite-driven data from NASA POWER and ground-truth data from Internet of Things (IoT) sensors. The study used measurements recorded every 15 minutes for 8 months with temporal coverage from June 2023 to February 2024. Our investigation explored five distinct modeling approaches: Extreme Gradient Boosting (XGBoost), Random Forest (RF), Ridge Regression (RR), Stacked Hybrid, and Weighted Hybrid models (WHM). The results indicate that the WHM method achieved the best balance between precision and generalization, with a coefficient of determination (R²) of 0.801, a root mean square error (RMSE) of 2.006 °C and a mean absolute error (MAE) of 0.850 on the validation data, attaining an accuracy of 95%. In comparison, XGBoost demonstrated strong baseline performance at 90%. This means that the WHM method improved performance by 5% compared to the best baseline, which was XGBoost. This carefully engineered combination of base models demonstrates how strategic ensemble techniques can mitigate individual model weaknesses while maintaining predictive power. In general, this study contributes to the development of a model capable of estimating environmental temperatures in locations where weather stations are not available.