Cloud-covered MODIS LST reconstruction by combining assimilation data and remote sensing data through a nonlocality-reinforced network

• Published in: International journal of applied earth observation and geoinformation Vol. 117; p. 103195
• Main Authors: Gong, Yuting, Li, Huifang, Shen, Huanfeng, Meng, Chunlei, Wu, Penghai
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.03.2023; Elsevier
• Subjects: Assimilation data; Land surface temperature reconstruction; Nonlocality-reinforced network; Remote sensing data; Land surface temperature reconstruction; Remote sensing data; Assimilation data; Nonlocality-reinforced network
• ISSN: 1569-8432; 1872-826X
• DOI: 10.1016/j.jag.2023.103195

•A nonlocality-reinforced network for cloud-covered MODIS LST reconstruction is proposed.•The combination of remote sensing data and assimilation data ensures the accuracy of the LST reconstruction under clouds.•The good validation results by comparing with the ground measurements verify the effectiveness of the proposed method.•This paper supplies a practical cloud-covered LST reconstruction way, in which four data combinations were optional. Reconstruction of cloud-covered thermal infrared land surface temperature (LST) is vital for the measurement of physical properties in land surface at regional and global scales. In this paper, a novel reconstruction method for Moderate Resolution Imaging Spectroradiometer (MODIS) LST data with a 1-km spatial resolution is proposed by combining assimilation data and remote sensing data through a nonlocality-reinforced network (NRN) model. Firstly, a data grading criterion is introduced to evaluate the importance of the various datasets, forming four combinations of multi-modal datasets for the training and testing of the NRN model. Secondly, the NRN model with a multiscale encoding–decoding structure considering the nonlocality-reinforced module is proposed for LST reconstruction. The results suggest that the proposed method can precisely reconstruct cloud-covered LST, with a mean absolute error (MAE) less than 0.8 K, even when no auxiliary remote sensing LST are used (Combination 1). The best result is the full combination (Combination 4), in which the coefficient of determination is 0.8956, the MAE is 0.5219 K, and the root-mean-square error is 0.7622 K. Compared with the traditional harmonic analysis of time series method, the improved enhanced spatial and temporal adaptive reflectance fusion method and the multiscale feature connected convolutional neural network method for LST reconstruction, the proposed method can achieve superior results. The proposed method with Combination 1 has been implemented to reconstruct the daily LST in the study area for 2019. Referring to the meteorological station observations, the reconstructed bias absolute value is less than 1 K, indicating that the proposed model is very effective and valid for regional cloud-covered LST reconstruction.