Forward and reverse retrieval algorithm between surface temperature and near surface air temperature through deep neural network model

Abstract

This thesis examined two way retrieval of Air Temperature (Ta) and Land Surface Temperature (LST) which can affect the rates of biotic processes, including phonologies, growth, carbon, fixation and respiration. For the first, we tried to retrieve LST of Landsat-8 satellite to use reference data. To retrieve LST which are most suitable for South Korea, we tested 4 split window algorithms which were referred in previous studies. As a test result, we confirmed the algorithm from Prata et al in 1991 could be used to retrieve most suitable LST. And then we examined the Ta retrieval method using Landsat-8 data throughout Deep Neural Network (DNN) model. To make accurate DNN model, we tested to select suitable input variables and model conditions. As a result, we selected Band 10 and Band 11 of Landsat-8 data as basic input data to the DNN model because Landsat-8 did not serve LST data officially. And Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI) are also selected as input variables and the Automated Synoptic Observing System (ASOS) Ta data was used as dependent variable. With these data and selected epoch and batch size condition, the Ta retrieval DNN model showed that the correlation coefficient was 0.9752 and Root Mean Square Error was 2.192 K. This result was more accurate than previous Ta retrieval studies. Then We examined the LST retrieval using Ta data throughout DNN model. In this step, we used 2 type of Ta data which are ASOS Ta and mixed ASOS Ta data and computational fluid dynamics (CFD) Ta data. With ASOS Ta data, we made the DNN model which showed 0.9801 correlation coefficient and 2.5408 K RMSE value. However, if we used CFD Ta data, we could make LST in satellite observation area without missing from cloud. So, we applied CFD Ta data to the DNN model using ASOS Ta, the result showed too low accuracy to use retrieved LST. And then we made the LST retrieval DNN model with mixed Ta data. The result of the DNN model showed 0.9806 correlation coefficient and 2.656 K RMSE value which are enough to use retrieved LST. And we finally made LST without missing pixel and it could be retrieved daily and hourly LST continuously.