研究目的
To address the problem of accurate cloud detection in remote sensing images using a deep-learning based framework that combines a Fully Convolutional Neural Network (FCN) with a gradient-based identification approach for snow/ice regions.
研究成果
The proposed deep-learning based framework significantly improves cloud detection accuracy in Landsat 8 images by combining FCN with a gradient-based snow/ice removal approach. Future work will focus on expanding the network's field of view for broader cloud context identification.
研究不足
The study is limited to Landsat 8 images and requires correction of ground truths for snow/ice regions. The framework's effectiveness on images from other satellites or with different spectral bands is not explored.
1：Experimental Design and Method Selection:
The study employs a Fully Convolutional Neural Network (FCN) inspired by U-Net for pixel-level labeling of cloud regions in Landsat 8 images. A gradient-based approach is used to identify and exclude snow/ice regions from the training set ground truths.
2：Sample Selection and Data Sources:
The dataset includes 4600 training patches from 18 Landsat 8 images and 5100 test patches from 20 different images, covering various scene elements.
3：List of Experimental Equipment and Materials:
Landsat 8 multi-spectral data (bands 2 to 5) is used.
4：Experimental Procedures and Operational Workflow:
The network is trained on corrected ground truths (with snow/ice removed) and default ground truths, then evaluated on the test set.
5：Data Analysis Methods:
Performance is evaluated using Jaccard index, recall, precision, and overall accuracy metrics.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
