A Convolution and Attention Neural Network with MDTW Loss for Cross-Variable Reconstruction of Remote Sensing Image Series

Environmental images that are captured by satellites can provide significant information for weather forecasting, climate warning, and so on. This article introduces a novel deep neural network that integrates a convolutional attention feature extractor (CAFE) in a recurrent neural network frame and a multivariate dynamic time warping (MDTW) loss. The CAFE module is designed to capture the complicated and hidden dependencies within image series between the source variable and the target variable. The proposed method can reconstruct the image series across environmental variables. The performance of the proposed method is validated by experiments using a real-world remote sensing dataset and compared with several representative methods. Experimental results demonstrate the emerging performance of the proposed method for cross-variable image series reconstruction.