Performance comparison of different algorithms for land surface temperature (LST) retrieval from Landsat-8 and the use of deep neural network regression to enhance the LST result (case study: Dallas, Texas and surrounding areas)

Land Surface Temperature (LST) is important data for various fields, especially for monitoring global warming. Landsat-8 satellite imagery provides thermal data with a spatial resolution of 30m (resampled from 100m) as the main data for LST retrieval. This research aims to compare several LST retrieval algorithms such as LST retrieval using band ten, Single Channel Model (SCM), Qin’s Split-Window Algorithm (Q-SWA), Sobrino’s Split-Window Algorithm (S-SWA), and LST from the analysis ready data of Landsat-8 Level two. The study focuses on Dallas, Texas, and surrounding areas in March 2023. We collected air temperature data from 20 U.S. Environmental Protection Agency (EPA) stations for indirect validation. Based on the results, the Q-SWA method outperformed other retrieval algorithms with Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) were 0.683°C and 0.842°C, respectively. Given that Landsat-8’s thermal band data is resampled from 100m to 30m, this research also explores enhancing the retrieved LST using Deep Neural Network Regression (DNNR). The best retrieved LST from the Q-SWA method served as the target data while several bands and spectral indices from Landsat-8 were used as the input for DNNR model. Due to the large scale of data, we randomly selected ten million pixels and divided into 80% of training and 20% of testing data. The DNNR model achieved MAE of 1.022°C on the testing dataset. The enhanced LST from the DNNR model was also validated with the same air temperature validation data and achieved the MAE score of 1.037°C. Based on the visual comparison result, the DNNR model successfully enhanced the retrieved LST by providing more detailed results at the same 30m resolution and showing promising performance based on error metrics. This finding suggests the potential for using deep learning regression in LST downscaling to achieve better spatial resolution.