基于PALSAR全极化数据的城市森林蓄积量估测

全极化雷达数据能够反映目标的全极化散射特征,在森林参数反演中具有较大的应用价值。笔者以南京紫金山国家森林公园为研究对象,以2011年的全极化雷达数据PALSAR和2012年120块野外调查样地为主要信息源,从Pauli和Cloude目标分解特征值、HH(horizontal-horizontal,水平)和HV(horizontal-vertical,水平垂直交互)两种极化状态的后向散射系数、比值植被指数、地形、人为干扰等方面,提取13个因子作为自变量,采用多元线性回归、人工神经网络、K最邻近分类算法、决策与回归树、装袋算法、随机森林6种方法建立遥感估测模型,进行森林蓄积量的估测。研究表明:①在6种遥感估测模型中,随机森林综合性能最高,装袋法次之,多元线性回归最低; ②海拔、坡向等地形因子,以及地物的雷达回波散射特征是影响研究区域森林蓄积量估测的重要变量; ③研究区单位面积蓄积量的空间分布呈现出由里向外逐渐降低的带状分布格局。

Estimation of stock volume of urban forest using fully polarimetric radar data of PALSAR

Forest stock volume is an important indicator for assessing the productivity of ecosystems, and also the basis for analysis of substance circulation in forest ecosystem. Forest stock volume of the different scales area can be estimated based on remote secsing technique, so the spatial distribution and dynamic monitoring of the forest stock volume are significant by using remote sensing techniques. Compared with the traditional optical remote sensing image, the fully polarimetric synthetic aperture radar(PALSAR)is almost unaffected by atmosphere and has the observation capabilities of the whole day and the ability to penetrate clouds, rain and snow. The fully polarimetric SAR image contains more abundant information. Since the fully PALSAR image has a great advantage of being able to obtain the fully polarized scattering attributes of the target object, it is widely used in estimation of forest parameters. In this paper, Zijinshan National Forest Park in Nanjing was chosen as the case study area, while PALSAR image in 2011 and 120 field plots in 2012 were collected as the main information source to estimate scenic forest stock volume. 13 factors including characteristic values extracted from Pauli and Cloude target decomposition, backscattering coefficients of HH and HV, ratio vegetation index, terrain and human disturbances were used to estimate forest parameters of unit stock volume using six remote sensing based models namely multivariate linear regression(MLR), artificial neural network(ANN), K nearest neighbor classification algorithm(KNN), classification and regression tree(CART), bagging(Bagging)and random forest(RF). Research results showed that: ① among the six models, the performance of random forest was the best, followed by bagging method, and multivariate linear regression was the worst; ② terrain factors of DEM(digital elevation model)and aspect, backscattering features of polarization radar echo were important environmental variables which affect unit stock volume; ③ spatial distribution of unit stock volume showed a zonal pattern descending from center to the periphery.