基于CNN的吉林一号卫星城市土地覆被制图潜力评估

以吉林一号视频07B星高分遥感影像为基础，采用卷积神经网络（CNN）对城区土地覆被进行精细分类，设置多组光谱变量集合，并与最大似然法、多层感知机和支持向量机分类方法进行对比，全面评估分析各方法对城区土地覆被信息提取的适用性及波谱特征对分类精度的影响。结果表明：CNN模型的分类精度最高，总体精度高于90%，相比其他方法提高幅度达12%以上，能够显著降低“椒盐”噪音；红边波段对所有方法总体分类精度贡献十分有限，而近红外波段对分类精度的提升较为明显；总体而言，红边和近红外波段对CNN分类精度影响较为微弱。深度学习应用于吉林一号高分遥感数据能获取高精度城区土地覆被分类图，可为城市土地资源配置，城市规划与管理提供重要的支撑。

Evaluating the Potential of JL1 Remote Sensing Data in Urban Land Cover Classification Using Convolutional Neural Networks

This research classified urban land cover using the Convolutional Neural Network （CNN） model based on the fine spatial resolution remotely sensed imagery from recently launched JL1 07B satellite. We applied CNN to classify imagery using different combinations of spectral feature variables， and compared the performance of CNN with three other methods， namely maximum likelihood classification algorithm， multi-layer perceptron algorithm and support vector machine algorithm. The experimental results demonstrated that CNN consistently achieved the highest overall accuracy （>90%）， larger than that of other methods by above 12%， and reduced significantly the “salt-and-pepper” noise. The contribution of red-edge band to the classification accuracy was slight， while the near-infrared （NIR） band could increase the OA prominently. Overall， the effect of red-edge and near-infrared bands exerted a slightly impact on the OA of CNN， demonstrating the robustness and generalization of the CNN model. The high accuracy urban land cover classification map achieved using CNN based on JL1 satellite imagery can support the decision makings for land resource allocation， urban planning and regional administration.