TransGA-Net: Integration Transformer With Gradient-Aware Feature Aggregation for Accurate Cloud Detection in Remote Sensing Imagery

Kai Xu,Wenxin Wang,Xiaoyuan Deng,Anling Wang,Bocai Wu,Zhaohong Jia

DOI: https://doi.org/10.1109/lgrs.2024.3353784

IF: 5.343

2024-02-02

IEEE Geoscience and Remote Sensing Letters

Abstract:Significant progress has been achieved in cloud detection using deep convolutional neural networks (CNNs), leading to a substantial increase in detection accuracy. However, discriminating between actual clouds and cloud-like objects, as well as accurately identifying thin clouds and cloud boundaries, have been persistent challenges in the cloud detection. To address these issues, an Integration Transformer with gradient-aware feature aggregation (TransGA-Net) is proposed in this letter. This framework employs Transformers as encoders, significantly enhancing the modeling capabilities for global features and long-term dependencies. Additionally, a novel gradient-aware feature aggregation module (GAFAM) is proposed, focused on gradient variations within cloud boundary regions and utilizing gradient information to guide feature aggregation. The performance of TransGA-Net is validated on Landsat-8, GF-2, and Sentinel-2 datasets. In general, experiments show TransGA-Net surpasses the highest performance achieved by the other methods by 0.6%, 1.3%, and 0.7% in OA on the three datasets, respectively, demonstrating that proposed method are particularly effective in identifying areas with thin cloud cover and distinguishing between clouds and cloud-like objects. The source code is available at https://github.com/phaeton2017/transga-net.

imaging science & photographic technology,remote sensing,engineering, electrical & electronic,geochemistry & geophysics