Improving land surface temperature prediction using spatiotemporal factors through a genetic-based selection procedure (Case Study: Tehran, Iran)

Abstract

In recent decades, land surface temperature (LST) has been considered as an important satellite product for thermal monitoring of the environment. Despite the great development of LST retrieval technology from satellite data, especially from Landsat images, up-to-date thermal information cannot always be obtained. Predicting LST values for future times is a pressing issues to understand and mitigate the effects of climate change, especially in urban environments. This highlights the importance of LST forecasting methods. Accordingly, in this study a combination of Genetic Algorithm and Geographically and Temporally Weighted Regression (GTWR) is presented to predict LST based on nine selected spatiotemporal factors affecting LST in the urban areas. The genetic algorithm is used to determine weights of different factors in the LST prediction with respect to the GTWR as the regression model. Then, these weights are used to obtain predicted LST values for experimental points. Landsat-derived LST bands is used to train the genetic algorithm and extract the factors’ weights. Tehran city in the summer is selected as the case study in our work. Landsat LST data from 2013 to 2016 are used for training. To assess the performance of our scheme, we apply it in 2017. We find a correlation coefficient of 0.99 and a root-mean-square error of 0.5 K with respect to the Landsat data. The results of this study also indicate that LST prediction using weighted factors is superior to using all factors with equal contributions.