普洱季风常绿阔叶林乔木物种多样性高分辨率遥感估测

Applying high resolution remote sensing to assess tree species diversity of monsoonal broad-leaved evergreen forest in Pu′er City

季风常绿阔叶林是我国南亚热带典型的地带性植被,也是云南省普洱地区重要森林类型。季风常绿阔叶林乔木物种多样性遥感估测对研究区域尺度生物多样性格局及其规律具有重要作用。根据光谱异质性假说和环境异质性假说,首先使用1m空间分辨率的机载高光谱数据和激光雷达数据提取了光谱多样性特征和垂直结构特征。然后利用基于随机森林算法的递归特征消除方法选择对研究区森林乔木物种多样性指数具有较好解释能力的遥感特征,并对Shannon-Winner物种多样性指数进行建模、制图。研究结果表明:(1)基于机载LiDAR数据提取的垂直结构特征和机载高光谱数据提取的光谱多样性特征均对研究区森林乔木物种多样性具有较好的解释能力,随机森林模型估测结果分别为R^(2)=0.48,RMSE=0.46和R^(2)=0.5,RMSE=0.45;两种数据源融合可以进一步提高遥感数据的森林乔木物种多样性估测精度,随机森林估测模型R^(2)和RMSE分别为0.69和0.37。(2)机载激光雷达数据对研究区针阔混交林乔木物种多样性的估测能力优于机载高光谱数据。(3)机器学习方法有助于从高维遥感数据特征中选择适合于森林乔木物种多样性建模的少量特征。该研究在云南普洱开展对季风常绿阔叶林的遥感估测研究,可为森林生物多样性调查提供补充手段,有助于森林生物多样性大尺度、长期动态监测。

The monsoonal broad-leaved evergreen forest is an important vegetation type of the Pu′er area which is located in the southern Yunnan Province, China. Tree species diversity mapping of monsoon evergreen broad-leaved forest plays an important role in studying the patterns of biodiversity at the regional scale. Remote sensing is an efficient alternative to the traditional field work to map tree species diversity over large areas. We tested the utility of using spatial heterogeneity in the airborne hyperspectral reflectance spectrum and structure heterogeneity in airborne LiDAR point cloud to model tree species diversity of monsoonal broad-leaved evergreen forest in Pu′er area. Shannon-Winner Index of tree species was calculated from field measurement. According to the spectral heterogeneity hypothesis and environmental heterogeneity hypothesis, the spectral diversity and structural diversity features were firstly extracted using airborne hyperspectral data and Lidar data. Then, the Random Forest based Recursive Feature Elimination(RF-RFE) was used to select the valuable airborne remote sensing features for forest tree species diversity modeling. At last, Random Forest regression model was used for modeling and mapping. Results showed that the spectral and structural diversity variables extracted from airborne hyperspectral, and Lidar remote sensing data explained 48% and 50% variance in tree species diversity. However, combining both variables explained 69% variance in tree species diversity. The Lidar variables had a better performance for coniferous and broad-leaved mixed forest tree species estimation than the hyperspectral variables. According to the correlation analysis, the result showed that not all RF-RFE selected remote sensing variables had a significant correlation with field measured Shannon-Winner diversity index. Machine learning such as Random Forest model is helpful to select a few valuable features for forest tree species diversity modeling from massive remote sensing data. This study shows the high potential of airborne LiDAR and hyperspectral data for monsoonal broad-leaved evergreen forest tree species diversity modelling and monitoring. The selected spatial and structure variables also highlight the potential of remotely sensible essential biodiversity variables for mapping and monitoring landscape floristic diversity from air-and space-borne platforms.