Characterizing Tree Species of a Tropical Wetland in Southern China at the Individual Tree Level Based on Convolutional Neural Network
Characterizing Tree Species of a Tropical Wetland in Southern China at the Individual Tree Level Based on Convolutional Neural Network
复制标题
基于卷积神经网络的中国南方热带湿地单树树种特征
DOI：
10.1109/jstars.2019.2950721
复制
发表时间：
2019-11-01
影响因子：
5.5
通讯作者：
Zhang, Hongsheng
中科院分区：
文献类型：
--
作者：
Sun, Ying;Xin, Qinchuan;Zhang, Hongsheng
在单株水平上进行物种分类将有利于森林景观可视化、森林管理和生物多样性监测等方面的应用。本文提出了一种基于卷积神经网络的树种个体分类算法。首先使用来自于光探测和测距（LiDAR）数据的冠层高度模型的局部最大值方法来检测单个树木。然后将检测到的树木个体裁剪成基于树顶的小块进行分类，并选择3个空间尺度图像小块进行分析和讨论。进一步采用改进的ResNet50深度网络进行裁剪后的树块分类。基于patch的方法考虑了树的上下文信息，不需要特征选择或特征约简过程。包括小无花果在内的1388个训练样本。f., Delonix regia, chisia speciosa a.st . hill。龙眼树。穆萨·娜娜·罗尔。其他树种（除上述6种树种外）通过野外调查和目测两种方法采集。利用航拍影像、LiDAR数据和Worldview影像进行树种分类。在362个测试树样本中，patch size为64的结果准确率最高，仅使用航空图像时，本文方法的总体准确率为89.06% + 0.58%，优于传统机器学习方法。对芦湖公园的可转移性研究也表明了该方法的可行性。尽管在单株树检测和多源数据融合方面仍然存在挑战，但该解决方案显示了利用遥感数据在单株树级别表征树种的潜力。
Classification of species at the individual tree level would be beneficial to many applications including forest landscape visualization, forest management, and biodiversity monitoring. This article develops a patch-based classification algorithm of individual tree species based on convolutional neural network. The individual trees are first detected using the local maximum method from the canopy height model, as derived from light detection and ranging (LiDAR) data. The detected individual trees are then cropped into patches for classification based on the tree apexes, and three spatial scale image patches are chosen for analysis and discussion. A modified ResNet50 deep network is further employed for the cropped individual tree patches classification. The patch-based method accounts for the contexture information of a tree and does not require the feature selection or the feature reduction processes. About 1388 training samples including Ficus microcarpa Linn. f., Delonix regia, Chorisia speciosa A.St.-Hil., Dimocarpus longan Lour., Musa nana Lour., Carica papaya, and Others (the other tree species except the above six) were collected from both field work and visual interpretation. Aerial images, LiDAR data, and Worldview images were used for the tree species classification. For 362 test tree samples, the results of patch size 64 achieve the best accuracies, and the proposed method outperforms the traditional machine learning method with the overall accuracy of 89.06% + 0.58% using aerial images only. Transferability Study to the Luhu Park also indicated the feasibility of our method. While challenges in individual tree detection and multisource data fusion remain, the solution shows the potential in characterizing tree species at the individual tree level using remote sensing data.