Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Distribution-aware Interactive Attention Network and Large-scale Cloud Recognition Benchmark on FY-4A Satellite Image

Jiaqing Zhang, Jie Lei, Weiying Xie, Kai Jiang, Mingxiang Cao, Yunsong Li

TL;DR

This work tackles cloud-type recognition in satellite imagery, a task underserved by prior detection-focused methods. It introduces the FYH benchmark, integrating FY-4A L1 data with H08 cloud-type labels across nine categories through domain adaptation to ensure consistent projection, resolution, and timing. The authors propose DIAnet, a high-resolution, multi-resolution network that uses a distribution-aware loss (DAL) and an Interactive Attention Module (IAM) to handle small targets and class imbalance, achieving state-of-the-art mean IoU on FYH (approximately 49.9%) and outperforming several baselines. The study provides a new dataset and a scalable, high-performance framework with practical implications for aviation support, weather forecasting, and climate research.

Abstract

Accurate cloud recognition and warning are crucial for various applications, including in-flight support, weather forecasting, and climate research. However, recent deep learning algorithms have predominantly focused on detecting cloud regions in satellite imagery, with insufficient attention to the specificity required for accurate cloud recognition. This limitation inspired us to develop the novel FY-4A-Himawari-8 (FYH) dataset, which includes nine distinct cloud categories and uses precise domain adaptation methods to align 70,419 image-label pairs in terms of projection, temporal resolution, and spatial resolution, thereby facilitating the training of supervised deep learning networks. Given the complexity and diversity of cloud formations, we have thoroughly analyzed the challenges inherent to cloud recognition tasks, examining the intricate characteristics and distribution of the data. To effectively address these challenges, we designed a Distribution-aware Interactive-Attention Network (DIAnet), which preserves pixel-level details through a high-resolution branch and a parallel multi-resolution cross-branch. We also integrated a distribution-aware loss (DAL) to mitigate the imbalance across cloud categories. An Interactive Attention Module (IAM) further enhances the robustness of feature extraction combined with spatial and channel information. Empirical evaluations on the FYH dataset demonstrate that our method outperforms other cloud recognition networks, achieving superior performance in terms of mean Intersection over Union (mIoU). The code for implementing DIAnet is available at https://github.com/icey-zhang/DIAnet.

Distribution-aware Interactive Attention Network and Large-scale Cloud Recognition Benchmark on FY-4A Satellite Image

TL;DR

This work tackles cloud-type recognition in satellite imagery, a task underserved by prior detection-focused methods. It introduces the FYH benchmark, integrating FY-4A L1 data with H08 cloud-type labels across nine categories through domain adaptation to ensure consistent projection, resolution, and timing. The authors propose DIAnet, a high-resolution, multi-resolution network that uses a distribution-aware loss (DAL) and an Interactive Attention Module (IAM) to handle small targets and class imbalance, achieving state-of-the-art mean IoU on FYH (approximately 49.9%) and outperforming several baselines. The study provides a new dataset and a scalable, high-performance framework with practical implications for aviation support, weather forecasting, and climate research.

Abstract

Accurate cloud recognition and warning are crucial for various applications, including in-flight support, weather forecasting, and climate research. However, recent deep learning algorithms have predominantly focused on detecting cloud regions in satellite imagery, with insufficient attention to the specificity required for accurate cloud recognition. This limitation inspired us to develop the novel FY-4A-Himawari-8 (FYH) dataset, which includes nine distinct cloud categories and uses precise domain adaptation methods to align 70,419 image-label pairs in terms of projection, temporal resolution, and spatial resolution, thereby facilitating the training of supervised deep learning networks. Given the complexity and diversity of cloud formations, we have thoroughly analyzed the challenges inherent to cloud recognition tasks, examining the intricate characteristics and distribution of the data. To effectively address these challenges, we designed a Distribution-aware Interactive-Attention Network (DIAnet), which preserves pixel-level details through a high-resolution branch and a parallel multi-resolution cross-branch. We also integrated a distribution-aware loss (DAL) to mitigate the imbalance across cloud categories. An Interactive Attention Module (IAM) further enhances the robustness of feature extraction combined with spatial and channel information. Empirical evaluations on the FYH dataset demonstrate that our method outperforms other cloud recognition networks, achieving superior performance in terms of mean Intersection over Union (mIoU). The code for implementing DIAnet is available at https://github.com/icey-zhang/DIAnet.
Paper Structure (12 sections, 6 figures, 4 tables)

This paper contains 12 sections, 6 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Benchmarks and Methodology
  3. Benchmarks
  4. Methodology
  5. Domain Adaptation
  6. Distribution of Sample Categories
  7. DIAnet
  8. Experience
  9. Experience Setting
  10. Ablation Studies
  11. Comparison with Other Methods
  12. Conclusion

Figures (6)

  • Figure 1: Comparison of different datasets: GF-1-WHU dataset li2017multi for satellite cloud detection, SkyCloud dataset gerhardt2023skycloud for ground-based cloud detection, and our proposed FYH dataset for satellite cloud recognition.
  • Figure 2: Projection Transformation: (a) is NOM projection and (b) is the result of projection transformation to EQR projection.
  • Figure 3: Domain Adaptive: (a) FY-4A L1 Data, (c) H08 Cloud-type Product, (b) and (d) the result of domain adaptive. The class unknown and clear are merged into class fill for display.
  • Figure 4: An illustration of cloud categories and the distribution of sample categories, (a) H08 Cloud-type Product, (b) The cloud categories distribution
  • Figure 5: An illustration of the proposed end-to-end DIAnet.
  • ...and 1 more figures