深度学习对不同分辨率影像冬小麦识别的适用性研究

定量分析遥感影像尺度与分类精度之间的关系是进行土地覆盖分类的基础。深度学习具有从底层到高层特征非监督学习的能力,解决了传统分类模型中需要人工选择特征的问题。这种新型的分类方法的分类精度是否受到不同分辨率尺度影响,有待研究。利用深度卷积神经网络(Deep Convolutional Neural Network,DCNN)--金字塔场景分析网络(Pyramid Scene Parsing Network,PSPNet)进行4种分辨率(8、3.2、2和0.8 m)的米级、亚米级影像冬小麦分类。实验结果表明:PSPNet能够有效地进行大样本的学习训练,非监督提取出空间特征信息,实现"端-端"的冬小麦自动化分类。不同于传统分类器分类精度与分类尺度之间的关系,随着影像分辨率的逐步增高,地物表达特征越来越清晰,PSPNet识别的冬小麦精度会逐步增高,识别地块结果也越来越规整,不受分辨率尺度的影响。这对于选择甚高亚米级影像提高作物分类精度提供了实验基础。

Comparison Analysis on Wheat Mapping Using Deep Learning Algorithm from Different Satellite Data Source

Quantitative analysis on the relationships between the remote sensing scale and the land cover classification accuracy,which is the basis for making a decision on remote sensing resolution determination,is essential for mapping the concise land cover. Up to now,deep learning is an innovative algorithm to learn the hierarchical layer features without supervised control,which is different from the traditional classifiers that require man-made labels as input. Therefore,it is interesting to explore the inherent relationship between the classification accuracy and remote sensing image spatial scale from this algorithm. In this paper,we applied a Deep Convolutional Neural Network(DCNN)which is Pyramid Scene Parsing Network(PSPNet)on four scale remote sensing image(8 m,3.2 m,2 m,0.8 m)to map the wheat distribution. The experiment results showed that the PSPNet is good at learning the spatial feature without manual operations,then the wheat extent could be extracted automatically. Different from the conventional algorithm of determining the optimized spatial resolution,the PSPNet could identify the wheat better accompanying with the spatial resolution increased and more concise wheat results could be obtained. This conclusions represent that deep convolution neural network can take full use of the spatial information of the high remote sensing image to ensure the performance of wheat extent,which brings us a new idea of improving the accuracy of crop mapping adequately if we can get the super-high resolution remote sensing image.