摘要:基于深度学习的高分辨率光学影像云检测过程中，云和云阴影及其边缘细节丢失较为严重，主要原因在于不同尺度空间语义信息特征融合存在不足。针对该问题，本文构建一种基于深度学习的多尺度特征融合网络（Multi-scale Feature Fusion Network, MFFN）的云和云阴影检测方法，该算法结合防止网络退化的残差神经网络模块（Res.block）、扩大网络感受野的多尺度卷积模块（MCM）和提取并融合不同尺度信息的多尺度特征模块（MFM）。试验表明，本算法能提取丰富的空间信息与语义信息，可取得较为精细的云与云阴影掩模，具有较高检测精度，其中云检测准确率达0.9796，云阴影检测准确率达0.8307。同时，该工作可为深度学习技术应用于业务云检测提供理论支持及技术储备。
关键词:云检测/
云阴影检测/
残差模块(Res.block)/
多尺度卷积/
多尺度特征模块
Abstract:Cloud detection based on high-resolution optical images combined with deep learning methodology cannot provide adequate and accurate information about the cloud, cloud shadows, or their edge details. The main reason for this problem is the insufficient fusion of semantic information in different scales of classification techniques. To address this problem, this study combines the Res.block (Residual block) module that can prevent network degradation, multiscale convolution module that can increase the receptive field of the network, and multiscale feature module that can extract and integrate information from different scales. In addition, this study proposes a detection algorithm based on the multiscale feature fusion network and deep learning. The experimental results showed that rich spatial and semantic information could be extracted by the algorithm. Cloud and cloud shadow masks with a higher level of accuracy can also be acquired. The accuracy of cloud and cloud shadow detection is 0.9351 and 0.8103, respectively. This study provides theoretical support and technical reserve for the application of deep learning techniques to operational cloud detection.
Key words:Cloud detection/
Cloud shadow detection/
Res.block (Residual block)/
Multi-scale convolution/
Multi-scale feature module
PDF全文下载地址:
http://www.iapjournals.ac.cn/dqkx/article/exportPdf?id=357a2c30-9718-44ab-8114-389298a861b2
