|
Sea surface temperature (SST) is a key parameter for monitoring the ocean environment and understanding various ocean phenomena, and is a key indicator of climate change. Satellite remote sensing data is an important technical tool for SST research, but the availability of data is reduced due to the influence of clouds and aerosols, which generate a large amount of missing data. The data interpolation empirical orthogonal function (DINEOF) method has usability and accuracy in reconstructing missing grid points of remote sensing datasets. In this study, we use a convolutional self-encoder neural network, modified for model skip connection and fully connected layers, and introduce an attention mechanism to extract spatio-temporal features of SST data, called attention data interpolation convolutional autoencoder (A-DINCAE), to achieve the reconstruction of infrared radiometer SST data and compare A-DINCAE with DINCAE and DINEOF Reconstruction accuracy. The accuracy of the reconstruction results is quantitatively evaluated using cross-validation datasets and actual measurement data, and the study area is selected as the South China Sea with the boundaries of 103-121°E and 0-23°N. The validation results show that the reconstruction effect of the A-DINCAE model on the SST missing data is better than that of DINCAE, the accuracy of the reconstruction results is much higher than that of DINEOF, and the reconstruction results restore the main SST of the sea area physical features of the sea area. This paper confirms that the attention mechanism can improve the DINCAE spatio-temporal feature extraction ability, and the small-scale features of the missing data are restored under the same data reconstruction conditions, and the A-DINCAE is more efficient than DINEOF, and The accuracy of the improved model has been improved.
|
