Research method on the retrieval of land surface temperature based on transfer learning with a deep belief network

Land Surface Temperature (LST) is a key parameter for analysing urban climatology. The precise monitoring of spatiotemporal variations in LST is crucially important for research on regional climate change. However, with the increasing availability of thermal infrared satellite data, machine learning models developed using older sensors tend to perform poorly on data from newer sensors. In this study, we developed a transfer learning model based on Deep Belief Networks (DBN). The DBN was first pre-trained using 667,535 samples from the Landsat-8 and its LST products to obtain the pre-trained model. The pre-trained model was then fine-tuned using 1,842 samples from Landsat-9 and in situ LST data to get the fine-tuned model (DBNLST_TL). The fine-tuned model was able to retrieve LST from Landsat-9. The retrieval results were cross-compared with the Landsat-9 LST products, and both were compared with the in situ LST. The accuracy validation results indicate that the fine-tuned model (R2 = 0.88, root mean squared error (RMSE) = 5.22 K) was superior to the pre-trained model (R2 = 0.74, RMSE = 7.57 K) and the Landsat-9 LST products (R2 = 0.86, RMSE = 5.23 K). Compared with that of the Landsat-9 LST products, the fine-tuned model achieved an average RMSE of 1.58 K and demonstrated a 7.65% improvement in spatial coverage. This study illustrates that a fine-tuned model can mitigate a certain degree of the challenge of intersatellite sensor applicability during the transfer learning phase of LST retrieval. The fine-tuning approach expands the application scope of large-sample datasets and the suitability of the models across different sensors.