Attention-driven and multi-scale feature integrated approach for earth surface temperature data reconstruction

Abstract

High-resolution observation is crucial for studying surface temperatures characterized by complex variations, particularly surface air temperatures in oceanic regions, which serve as significant indicators of sea-air coupling changes. Due to the scarcity of conventional observations of surface atmospheric temperatures in these areas, high-resolution surface atmospheric temperature data derived from satellite inversion has become the primary source of information. However, missing data resulting from factors such as the orbital spacing of polar satellites, cloud cover, sensor errors, and other disruptions poses substantial challenges to Earth Surface Temperature (EST) estimation. In this paper, we introduce ESTD-Net, a novel deep learning-based model designed for surface temperature data inpainting. ESTD-Net incorporates an enhanced multi-head context attention mechanism and a modified transformer block to capture long-range pixel dependencies, thereby improving the model's ability to focus on boundary regions. Additionally,the Stage Two employs a convolutional U-Net in an autoregressive manner to refine the coarse output from the Stage One, enhancing local spatial continuity and smoothing boundaries. In addition, we adapt two loss components-weighted reconstruction loss and gradient consistency regularization-to the specific demands of Earth surface temperature interpolation. Our ablation studies confirm that their integration significantly improves spatial consistency and accuracy, particularly in textureless regions and in maintaining physically meaningful gradients. Evaluation results demonstrate that ESTD-Net outperforms existing methods in both pixel-level accuracy and perceptual quality. Our approach offers a robust and reliable solution for restoring earth surface temperature data.