树种调查一直面临着成本高、效率低、精度不高等问题。利用遥感手段能大大提高树种类型调查的工作效率、节省成本;卷积神经网络(CNN)虽然已经在自然图像分类领域取得了许多突破,但是较少有人将CNN模型用于单木树种分类。基于上述考虑,搭建出CNN模型,并与高分遥感影像相结合,进行单木树种分类。在利用高分影像半自动化构建单木树种遥感影像样本集过程中,采用了影像冠层切片(CSI)圈定、人工标注、数据增强等方法;同时为了训练单木树种遥感影像样本集,对5个CNN模型进行针对性改写。通过对比分析发现:LeNet5_relu和AlexNet_mini都未取得最佳分类效果;GoogLeNet_mini56、ResNet_mini56和DenseNet_BC_mini56分别对不同的树种具有最佳分类效果;DenseNet_BC_mini56总体精度最高(94.14%),Kappa系数最高(0.90),是总体最佳分类模型。该研究证明了CNN在单木树种分类中的有效性,能为森林资源调查提供重要的解决方案。

Tree species investigation has been faced with problems such as high cost, low efficiency, and low precision. The use of remote sense can greatly increase the work efficiency of tree species investigation and save cost. Although convolutional neural network (CNN) has made many breakthroughs in natural image classification area, few people have used CNN model to carry out individual tree species classification. Based on the above considerations, this paper builds CNN models, and integrates them with high-resolution remote sensing imagery to classify individual tree species. In the course of semi-automatically constructing the sample set of remote sensing imagery of individual tree species with high-resolution imagery, the crown slices from imagery (CSI) delineation, manual annotation, and data augmentation are used. Meanwhile, in order to train the sample set of remote sensing imagery of individual tree species, five CNN models are adapted. Through comparative analysis, it is found that LeNet5_relu and AlexNet_mini cannot achieve the best classification effect. GoogLeNet_mini56, ResNet_mini56, and DenseNet_BC_mini56 have the best classification effect for different species respectively. DenseNet_BC_mini56 has the highest overall accuracy (94.14%) and the highest Kappa coefficient (0.90), making it the best classification model from all aspects. The research proves the effectiveness of CNN in the classification of individual tree species, which can provide a critical solution for forest resource investigation.