Machine learning-driven analysis of agro-climatic data for temperature modeling and forecasting in Philippine urban areas

Abstract

The effects of climate change in the Philippines call for effective strategies to improve resilience, especially in urban areas. Machine learning models are now being used to provide data-driven insights for climate action, in particular, to address urban overheating. In this context, this paper developed machine learning models by using agro-climatological data to predict the maximum temperature at 2 m (in °C) in Manila and Dagupan, Philippines, with 32 predictors. A pipeline of standard scaling, principal component analysis, regression models, and time-series models were used for forecasting. It was found that the multilayer perceptron (MLP) regressor had the best test forecast performance in the Manila data, with an $R^2$ of 0.8128 and MSE of 0.9334, even without autoregressive information. Meanwhile, Long Short-Term Memory was found to have comparatively decent performance with a test $R^2$ of 0.6452 for the case of univariate autoregressive forecasting. We also prove that the models are location-specific since the model trained at Manila data yields inaccurate results when transferred to the Dagupan data. With more accurate forecasts of maximum temperature, policymakers can make more informed decisions toward a more sustainable living in Philippine cities.