U-High resolution network (U-HRNet): cloud detection with high-resolution representations for geostationary satellite imagery

Runzhe Tao,Yonghong Zhang,Lihua Wang,Qi Liu,Jiangeng Wang
DOI: https://doi.org/10.1080/01431161.2021.1875510
IF: 3.531
2021-02-11
International Journal of Remote Sensing
Abstract:<span>Cloud detection is one of the important tasks for remote sensing image preprocessing. In this work, multiple infrared observation channels with high time resolution are utilized to extract cloud from geostationary satellite imagery. Compared with detected cloud mask from polar satellite data with the high spatial resolution, the detection from geostationary satellite data has high timeliness and practicability, but it is challenging for the detailed segmentation ability of the network. Due to the fact that some small clouds only occupy a few pixels in geostationary satellite images with low spatial resolution, the existing network has poor segmentation ability for these small targets. To tackle this problem, a new neural network named U-High Resolution Network (U-HRNet) is proposed for cloud detection. Combing both multi-scale feature extraction using a High Resolution Network (HRNet) architecture and merging shallow information and deep information via the skip connection (SC) of a U-Net structure, the proposed U-HRNet produces strong high-resolution representations for accurate detection of details on cloud segmentation. The accuracy of the method is evaluated by manually labelled ground truth against different methods using objective evaluation indices. The results proved the proposed U-HRNet performs well on FengYun-4A (FY-4A) images and can effectively detect incorrect areas of the cloud mask products of the National Satellite Meteorological Centre (NSMC), depicting outperformance over existing methods.</span>
imaging science & photographic technology,remote sensing
What problem does this paper attempt to address?
### Problems Addressed by the Paper The paper aims to address the issue of cloud detection in geostationary satellite images. Specifically, it focuses on improving the detail segmentation capability in low-resolution geostationary satellite images, particularly for the precise segmentation of small cloud clusters and their edges. #### Background and Challenges - **High-Resolution Images vs. Geostationary Satellite Images**: Compared to the high spatial resolution images provided by polar-orbiting satellites, geostationary satellite data, although more timely and practical, pose challenges in detail segmentation. Since some small cloud clusters occupy only a few pixels in low-resolution geostationary satellite images, existing network structures perform poorly in handling these small targets. - **Limitations of Traditional Methods**: Traditional threshold methods are simple and efficient but are easily affected by complex surface environments, and the thresholds are difficult to adapt to changes in different times and weather conditions. - **Shortcomings of Machine Learning Methods**: Shallow machine learning methods (such as support vector machines, random forests, etc.) are more accurate than threshold methods, but the manually extracted features are difficult to capture the true characteristics of clouds in complex environments. #### Solution To address the above issues, the authors propose a new deep learning architecture—U-High Resolution Network (U-HRNet), which combines the advantages of High Resolution Network (HRNet) and U-Net: 1. **Multi-Scale Feature Extraction**: Extract multi-scale features through the HRNet architecture to maintain high-resolution representation. 2. **Reducing Errors Introduced by Linear Interpolation**: Use the Joint Pyramid Up-sampling (JPU) method for multi-scale feature fusion to extract rich high-resolution representations. 3. **Pixel-Level Semantic Segmentation**: Combine the decoder structure of U-Net to resample feature resolution and provide pixel-level semantic segmentation of cloud images. Additionally, considering the importance of shortwave infrared, water vapor channel, and longwave infrared channel in cloud detection and the unavailability of visible light channels at night, the paper uses high temporal resolution data from the Fengyun-4A (FY-4A) geostationary satellite and utilizes infrared observation channels as network input to train a general segmentation model suitable for both daytime and nighttime cloud images. #### Experimental Results Experimental results show that the proposed U-HRNet method performs excellently on Fengyun-4A satellite images, effectively detecting erroneous areas in the National Satellite Meteorological Center (NSMC) cloud mask products, and outperforms existing methods on multiple evaluation metrics.