基于CNN和农作物光谱纹理特征进行作物分布制图

Crop Mapping Using Remotely Sensed Spectral and Context Features based on CNN

以卷积神经网络(Convolutional Neural Network,CNN)为代表的深度学习技术,在农作物遥感分类制图领域具有广阔的应用前景。以多时相Landsat 8多光谱遥感影像为数据源,搭建CNN模型对农作物进行光谱特征提取与分类,并与支撑向量机(SVM)常规分类方法进行对比。进一步引入影像纹理信息,利用CNN对农作物光谱和纹理特征进行提取,优化作物分布提取结果。实验表明:①基于光谱特征的农作物分布提取,验证结果对比显示,CNN对应各类别精度、总体精度均优于SVM,其中二者总体精度分别为95.14%和91.77%;②引入影像纹理信息后,基于光谱和纹理特征的CNN农作物分类总体精度提高至96.43%,Kappa系数0.952,且分类结果的空间分布更为合理,可有效区分花生、道路等精细地物,说明纹理特征可用于识别不同作物。基于光谱和纹理信息的CNN特征提取,可面向种植结构复杂区域实现农作物精准分类与分布制图。

Deep learning algorithms such as Convolutional Neural Network(CNN)can learn the representative and discriminative features in a hierarchical manner from the remote sensing data. Considering the low-level features as the bottom level,the output feature representation from the top level of the network can be directly fed into a subsequent classifier for pixel-based classification,the CNN has a broad application prospect in the field of agricultural remote sensing. The advantage of CNN in feature extraction can obtain the crop classification in complex planting structure area from multi-spectral remote sensing data,which is difficult in conventional methods. In this paper,a crop mapping method using remotely sensed spectral and context features based on CNN from Landsat OLI data is proposed and applied in Yuanyang county.The architecture of the proposed CNN classifier contains eight layers with weights which are the input layer,two convolution layers,two max pooling layers,two full connection layers and output layer. These eight layers are implemented on spectral and context signatures from 4 different phase Landsat OLI images to discriminate different crops against others. Experimental results demonstrate that the proposed CNN classifier can achieve better classification performance than support vector machines in spectral domain. The context features calculated by the gray level co-occurrence matrix method from Landsat OLI data can enhance the proposed CNN method to achieve the best results.In terms of verification accuracy,the proposed CNN classifier is superior than SVM in spectral domain. The overall accuracy of the two methods is 95.14% and 91.77%,respectively. The accuracy of the proposed classifier is further improved by adding spatial context features on the basis of spectral information. The overall accuracy and Kappa coefficient of the proposed method is 96.43% and 0.952.Furthermore,the crop mapping using spectral and context features based on CNN achieves better spatial representation especially for peanut and roads which is easy to form mixed-pixel. The context features can be extracted by the CNN to enhance the feature representation of these small objects.The CNN-based method from remotely sensed spectral and context features for crop mapping can achieve outstanding performance especially for the fine ground objects in complex planting structure area such as peanuts and roads.