Improving accuracy of land surface temperature prediction model based on deep-learning

Land surface temperature (LST) data is essential for urban engineering as well as modeling the atmospheric phenomena. Such modeling efforts require accurate temperature prediction which is then used for predicting other meteorological phenomena such as urban heat island and fine dust air pollution. The automatic weather system (AWS) provides accurate temperature with high frequency but it cannot grasp spatially continuous distribution in detail because it is collected only at specific points. On the contrary, the LST data obtained from satellite imagery has a high spatial resolution and spatially continuous temperature can be grasped, but it is difficult to get high temporal frequency temperature data because of its revisit time. In this study, to solve this spatio-temporal tradeoff problem, a deep-learning method was used to create a spatially continuous temperature image using AWS data with a spatial resolution of 30 m. The seasonal temperature was predicted with accuracy of 3.6 °C for spring, 1.9 °C for summer, 3 °C for fall, and 1.4 °C for winter. The predicted temperature accuracy for spatial resolution of 30 m is better than other reported interpolation methods. In order to improve the prediction accuracy of the model, fine tuning procedures were applied to the deep learning model hyper parameters as well as the input feature data.