Deep neural network regression for estimating land surface temperature at 10 meter spatial resolution using Landsat-8 and Sentinel-2 data

Land Surface Temperature (LST) is an important factor in geophysical parameters such as hydrological modeling, soil moisture, monitoring crop, etc. LST data with detailed resolution and the large-scale area is very helpful data in many research fields. Satellite imagery with thermal infrared sensors can be used to produce LST using a retrieval algorithm. Currently, Landsat 8 with TIRS sensor is freely available thermal infrared bands with the highest spatial resolution (resampled from 100m to 30m). Based on that situation, this study aims to build a model from optical bands of Landsat-8 as the input data and LST from Landsat-8 as the target data using Deep Neural Network regression (DNNr) architecture and then applied to Sentinel-2 to get LST at 10m resolution. The main difference of DNNr architecture with DNN for classification is we use linear activation function in the output layer. The study area is located in Yilan County, Taiwan. The input data from Landsat-8 and Sentinel-2 are optical bands (Blue, Green, Red, NIR), NDVI, and emissivity from NDVI. Both the input data have been standardized using the standardscaler function before feeding into the model. The input data were separated as 70 for training, 20 for validation, and the other 10 as testing data. We use air temperature data to calculate indirect validation with LST from Sentinel-2. The result shows, the mean absolute error and mean squared of testing data from DNNr are 0.581oC and 0.766oC. The correlation and maximum difference of air temperature with LST Sentinel-2 from DNNr are 0.92 and 2.94oC. Based on the experiments, our DNNr achieved a more good result than other regression architecture. Our DNNr architecture has been tested in other areas and also shows acceptable result. Based on that results, our LST product at 10m resolution can be used in others research fields. © 2021 SPIE.