High-Resolution Cloud Detection Network

Jingsheng Li,Tianxiang Xue,Jiayi Zhao,Jingmin Ge,Yufang Min,Wei Su,Kun Zhan

2024-07-10

Abstract:The complexity of clouds, particularly in terms of texture detail at high resolutions, has not been well explored by most existing cloud detection networks. This paper introduces the High-Resolution Cloud Detection Network (HR-cloud-Net), which utilizes a hierarchical high-resolution integration approach. HR-cloud-Net integrates a high-resolution representation module, layer-wise cascaded feature fusion module, and multi-resolution pyramid pooling module to effectively capture complex cloud features. This architecture preserves detailed cloud texture information while facilitating feature exchange across different resolutions, thereby enhancing overall performance in cloud detection. Additionally, a novel approach is introduced wherein a student view, trained on noisy augmented images, is supervised by a teacher view processing normal images. This setup enables the student to learn from cleaner supervisions provided by the teacher, leading to improved performance. Extensive evaluations on three optical satellite image cloud detection datasets validate the superior performance of HR-cloud-Net compared to existing methods.The source code is available at \url{<a class="link-external link-https" href="https://github.com/kunzhan/HR-cloud-Net" rel="external noopener nofollow">this https URL</a>}.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

The problem that this paper attempts to solve is: when dealing with high - resolution images, the existing cloud detection networks fail to fully explore and utilize the complex texture details of clouds. Specifically, traditional cloud detection methods are insufficient in capturing cloud texture information at high resolution, resulting in not - high - enough accuracy in cloud detection, especially performing poorly when dealing with small - scale cloud areas with irregular shapes and complex edges.

To overcome these challenges, the author proposes a new high - resolution cloud detection network (HR - cloud - Net), aiming to improve the performance of cloud detection in the following ways:

1. **High - Resolution Representation Module**: Keep the detailed texture information in high - resolution images through multi - level feature extraction and fusion.
2. **Inter - layer Cascade Feature Fusion Module**: Gradually fuse feature maps at different levels to ensure information transfer and enhancement from low - resolution to high - resolution.
3. **Multi - resolution Pyramid Pooling Module**: Integrate feature maps at different resolutions, capture global context information, and further improve the accuracy of cloud detection.
4. **Multi - view Supervision Strategy**: Introduce a teacher - student view mechanism, and the teacher view trained by normal images guides the student view trained by enhanced images, thereby improving the generalization ability of the model.

Through the above design, HR - cloud - Net can provide more accurate segmentation results in complex cloud detection tasks, especially when dealing with high - resolution remote sensing images. Experimental results show that HR - cloud - Net outperforms existing methods on multiple public datasets, verifying its effectiveness and superiority.

### Formula Summary

1. **Cross - Entropy Loss Function (Cross - Entropy Loss)**:
\[
L_{ce}=-\sum_{i\in P}\sum_{j\in C}t_{ij}\log y_{ij}
\]
where \( Y = [y_{ij}] \), \( T = [t_{ij}] \), \( C=\{1, 2\} \), \( P = \{1,\ldots, W\times H\} \).

2. **Augmented Image Cross - Entropy Loss Function (Augmented Cross - Entropy Loss)**:
\[
L_{aug_{ce}}=-\sum_{i\in P}\sum_{j\in C}I(y_{ij}>\tau)t_{ij}\log y_{aug_{ij}}
\]
where \( I(\cdot) \) is an indicator function, which outputs 1 when the condition is true and 0 otherwise; \( I(y_{ij}>\tau) \) returns a high - confidence mask.

3. **Total Loss Function (Total Loss Function)**:
\[
L=\lambda_1L_{ce}+\lambda_2L_{aug_{ce}}
\]
where \( \lambda_1 \) and \( \lambda_2 \) are trade - off hyperparameters.

Through the design of these formulas and modules, HR - cloud - Net can effectively capture and utilize high - resolution cloud texture information, thereby improving the accuracy and robustness of cloud detection.