Missing Data Reconstruction in Land Surface Temperature Based on the Improved U-Net Framework

Abstract

As a key parameter for studying the movement of the earth, Land Surface Temperature (LST) data has always played an important role. However, since people began to collect LST data purposefully, they have been plagued by the problem of missing data. In recent years, with the development of satellite remote sensing technology, the quality of LST data has been continuously improved. However, due to the limited resolution of the sensor and the complex atmospheric environment, the problem of missing data still occurs in the LST data, and the accuracy and effectiveness of the data are therefore greatly reduced. In this article, we propose a LST data reconstruction method based on deep learning network, and use this method to reconstruct the 9-year (2000–2008) MODIS LST map. In the comparative experiment of selected regions, the ratio of “high-quality” pixels that can be collected in the reconstructed data is between 32% and 41.5%. Compared with traditional methods, the ratio of “high-quality” pixels that can be collected is increased by 10% to 20%, which greatly improves the use value of massive satellite data. Experiments also show that the use of different loss functions will also affect the reconstruction effect of the method.