Predicting Land Surface Temperature in Tarkwa Nsuaem Municipality: A Comparison of Time Series Models

Comparison of Time Series Models in Predicting LST in Tarkwa Nsuem

Abstract

Land Surface Temperature (LST) prediction is a crucial environmental monitoring activity that has applications in both urban and rural areas for evaluating thermal conditions and humanenvironment interactions, especially in developing countries. This study compares the performance of three time-series models: Recurrent Neural Networks (RNN), Long Short-Term Memory networks (LSTM) and Multi-Layer Perceptrons (MLP) for LST prediction in the Tarkwa Nsuaem Municipal Assembly (TNMA), Ghana, where daily LST data (MODIS Terra MOD11A1 product) from 2014 to 2023 was aggregated to weekly time steps and used to train and test the models. The performance of the models was evaluated based on Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE) and Symmetric Mean Absolute Percentage Error (SMAPE). The results indicate, the LSTM model performed better (RMSE = 1.3533, MAE = 0.9991, MAPE = 3.8037%, SMAPE = 3.7166%) compared with RNN (RMSE = 1.3863, MAPE = 3.9078%) and MLP (RMSE = 1.9674, MAPE = 5.9421%). In comparison with the MLP, the LSTM model reduced the RMSE by 31.2% and MAPE by 36.0%, while comparing with the RNN model, the RMSE and MAPE were reduced by 2.4% and 2.7%, respectively. The superiority of the LSTM model is due to its gated memory structure, which captures long-range temporal dependencies in the LST time series. Hence, the LSTM is suggested as the best approach for the prediction of LST in TNMA. Therefore, further studies should be conducted to explore extended hyperparameter tuning, advanced preprocessing methods and additional explanatory variables.