基于机器学习的深圳市坝光湿地园树种高光谱分类

高光谱遥感数据为树种的精细识别提供了可能。为探索高光谱数据在树种识别中的能力，本研究基于深圳市坝光古银叶树群落的8种主要树种叶片高光谱数据，比较了6种光谱预处理方式和2种分类方法对树种分类识别精度的影响，并基于随机森林算法对不同树种识别的特征波段进行了识别。研究结果表明，一阶导数预处理方法在分类识别中性能最好，平均分类精度为76.65%；随机森林回归方法较支持向量回归算法的性能好，模型平均分类识别精度为73.07%。从混淆矩阵可以看出，多毛茜草、银柴、阴香易错分为假萍婆，鸭脚木与银柴易错分，银叶树和细叶榕易错分。400 nm、495 nm、615～675 nm、835 nm、915～975 nm、1035～1065 nm、1085～1135 nm、1265～1275 nm、1425～1535 nm、2040 nm、2100～2270 nm、2430 nm附近的光谱数据与8个树种分类识别有密切关系。

Classification in Baguang Wetland Park in Shenzhen Based on Machine Learning and Hyperspectral Data

Hyperspectral remote sensing data provides the possibility for fine identification of tree species. In order to explore the ability of hyperspectral data in tree species identification, this study is based on the leaf hyperspectral data of eight major tree species in the heritiera littoralis community of Baguang, Shenzhen, and compared the performance of six spectral preprocessing methods and two classification methods to classify tree species. Then based on the random forest algorithm, the importance of the each band was evaluated. The results showed that the first derivative preprocessing method had the best performance in classification and identification, and the average classification accuracy was 76.65%. The random forest regression method had better performance than the support vector regression algorithm, and the model average classification recognition accuracy was 73.07%. It can be seen from the confusion matrix that Aidia pycnantha, Aporosa dioica, Cinnamomum burmanni were recoginized as Sterculia lanceolato. There were the misclassification between Scheffero octorphylla and aporosa diocia. And Heritiera littoralis was also misclassified as Ficus microcarpa. Spectral data near 400 nm, 495 nm, 615-675 nm, 835 nm, 915-975 nm, 1035-1065 nm, 1085-1135 nm, 1265-1275 nm, 1425-1535 nm, 2040 nm, 2100-2270 nm, and 2430 nm are identified as the spectral features, which are most important for the classification of eight tree species.