Source-linked AI summary
Cloud-Net: An end-to-end Cloud Detection Algorithm for Landsat 8 Imagery
Sorour Mohajerani, Parvaneh Saeedi
TL;DR
Cloud detection is important for satellite-image analysis but is challenging with limited spectral bands and varied cloud features. The paper introduces Cloud-Net, an end-to-end FCN trained on Landsat 8 patches to learn local and global cloud features without complicated preprocessing. On the 38-Cloud benchmark, it improves FCN’s Jaccard Index by 8.7% and exceeds Fmask’s performance.
Problem
Cloud detection is an important preprocessing task in satellite imagery, where limited available spectral bands make the problem challenging.
Method
Cloud-Net is an end-to-end fully convolutional network that uses convolution blocks to learn local and global cloud features from Landsat 8 image patches.
Results
8.7%: Cloud-Net improves FCN’s Jaccard Index on the 38-Cloud test set and exceeds Fmask’s performance.
Takeaways & Limitations
Cloud-Net provides end-to-end pixel-level cloud labeling without complicated preprocessing and delivers superior performance to the compared methods.
Abstract
from arXiv · showhide
Cloud detection in satellite images is an important first-step in many remote sensing applications. This problem is more challenging when only a limited number of spectral bands are available. To address this problem, a deep learning-based algorithm is proposed in this paper. This algorithm consists of a Fully Convolutional Network (FCN) that is trained by multiple patches of Landsat 8 images. This network, which is called Cloud-Net, is capable of capturing global and local cloud features in an image using its convolutional blocks. Since the proposed method is an end-to-end solution, no complicated pre-processing step is required. Our experimental results prove that the proposed method outperforms the state-of-the-art method over a benchmark dataset by 8.7\% in Jaccard Index.
1. INTRODUCTION
Cloud detection is an important preprocessing task in satellite-image analysis, while existing approaches include threshold-based, handcrafted, and deep learning methods. Cloud-Net is introduced as an end-to-end network designed to capture local and global cloud features without complicated preprocessing.
- Motivation: Cloud identification supports remote-sensing tasks by locating coverage that can occlude land objects and interfere with change detection, parameter retrieval, and tracking.Cloud coverage can also provide information about climate parameters and natural disasters.
- Related work: Existing cloud-detection methods comprise threshold-based, handcrafted, and deep learning-based approaches.Examples include Fmask, Haze Optimized Transformation, CNN classification, and FCN segmentation.
- Contribution: Cloud-Net performs end-to-end pixel-level labeling while learning local and global cloud features from the entire scene.Its convolution blocks are designed to capture complicated semantic cloud features.
- Contribution: Cloud-Net does not require complicated preprocessing such as super-pixel segmentation.The paper contrasts this with Fmask and the method in.
- Contribution: The paper also modifies the dataset from by replacing images with inaccurate or uncertain ground truths and makes the revised dataset available by request.
2. METHODOLOGY
Cloud-Net uses contracting and expanding network arms to transform Landsat 8 spectral patches into cloud probability masks. Its design uses shortcut connections and soft Jaccard-loss optimization for cloud segmentation.
- Architecture: Cloud-Net uses contracting and expanding arms to extract image features, recover cloud attributes, and produce a pixelwise cloud probability map.The contracting arm extracts deep low-level features, while the expanding arm reconstructs the cloud mask.
- Data and dimensions: The network uses four Landsat 8 bands, processes 384 × 384 patches resized to 192 × 192, and outputs a 192 × 192 × 1 cloud mask.The input size is 192 × 192 × 4.
- Architecture: Shortcut connections combine layers within blocks and connect the contracting and expanding arms.These connections help the expanding arm generate a more accurate cloud mask and prevent vanishing gradients during backpropagation.
- Optimization: Training applies flipping, rotation, and zooming augmentation, uses ReLU activations and a final sigmoid layer, and optimizes the model with Adam.
- Optimization: The soft Jaccard loss uses ground-truth and output pixel arrays, with ϵ = 10^-7 added to numerator and denominator to avoid division by zero.N denotes the total number of ground-truth pixels.
3. EXPERIMENTAL SETTING
The experimental setting uses a modified 38-Cloud dataset, patch-based inference, and pixel-level comparisons between predicted masks and ground truths. Performance is assessed with accuracy, recall, precision, specificity, and Jaccard Index.
- Dataset: The modified dataset contains 18 Landsat 8 training images and 20 testing images, with five inaccurate or uncertain images replaced.Training ground truths were manually annotated instead of automatically generated.
- Test phase: For an unseen test image, Cloud-Net splits the scene into non-overlapping 384 × 384 patches, resizes them to 192, and predicts a cloud probability map for each patch.
- Test phase: Each patch probability map is binarized with a global threshold of 0.047, resized to 384 × 384, and stitched into the final scene mask.
- Evaluation: Predicted cloud masks are compared with corresponding ground truths for each pixel, using cloud and clear classes.
- Evaluation: Performance is measured with Overall Accuracy, Recall, Precision, Specificity, and Jaccard Index.The metrics are defined from true-positive, true-negative, false-positive, and false-negative pixel counts.
4. EXPERIMENTAL RESULTS
On 20 test images from the 38-Cloud dataset, Cloud-Net outperforms the compared FCN and Fmask methods. The reported FCN comparison shows an 8.7% improvement in Jaccard Index under the same training set.
- Quantitative results: 8.7%: Cloud-Net improves FCN’s Jaccard Index on the 38-Cloud test set.The comparison uses the same training set, with FCN trained using the setting from and more accurate ground truths.
- Quantitative results: Cloud-Net exceeds Fmask’s performance on the reported 38-Cloud evaluation.
- Evaluation: The quantitative results are reported over 20 test images from the 38-Cloud dataset.
5. CONCLUSION
The paper develops Cloud-Net, a deep learning algorithm for cloud segmentation that uses sophisticated convolution blocks. It also reports superior performance to competing state-of-the-art algorithms and provides a more reliable dataset for future analysis.
- Cloud-Net is a deep learning-based algorithm for segmenting clouds in remote sensing images.
- Its specialized architecture benefits from sophisticated convolution blocks.
- Cloud-Net performance is superior to competing state-of-the-art algorithms.
- The modified dataset provides researchers with a more reliable and accurate resource for future analysis.