A Modified Transfer-Learning-Based Approach for Retrieving Land Surface Temperature From Landsat-8 TIRS Data

As a critical parameter of the land surface energy balance, land surface temperature (LST) has received extensive attention in various fields. Thermal infrared (TIR) remote sensing can efficiently obtain large-scale, long-time-series information on land surface thermal radiance. Multiple algorithms, including physics-based and deep learning algorithms, have been proposed for obtaining LST from the observations. In algorithm analysis, the theoretical performance is typically evaluated through simulated data representing atmospheric and land surface environmental conditions that vary globally before applying it to authentic TIR remote sensing images and validating the accuracy using ground-measured data. However, due to the complexity of the observation environment, errors in results obtained from validation using ground-measured data tend to be larger than theoretical errors obtained using simulated data, which limits the performance of the algorithm in practical applications. Obtaining globally covered, representative ground-measured data and synchronized observations of remote sensing images are costly, making it difficult to provide enough data to develop new algorithms. The transfer learning method can learn from the pre-trained deep learning model that applies to the source task, and fine-tuning using only a small number of samples from the target task can result in a well-performing model. This paper proposes a modified transfer-learning-based (TL) land surface temperature retrieval algorithm to pre-train a knowledge-driven deep neural network model using a simulation dataset and then use a small amount of ground-measured data for tuning to obtain the final LST retrieval model. The new algorithm was applied to the remote sensing data observed by the Landsat-8 Thermal Infrared Sensor (TIRS), and the validation results based on the ground-measured data and global Landsat-8 LST products showed that the RMSE of the fine-tuned model result was about 0.4 K lower than the Landsat-8 product, and about 0.3 K lower than the pre-trained model results, reaching 2.2 K. Moreover, the results are in good agreement with the Landsat LST product in multiple regions worldwide with different land cover types, which demonstrated the effectiveness and stability of the proposed TL algorithm.