首页 >  , Vol. , Issue () : 1993-2002

摘要

引用本文:

DOI:

收稿日期:

2018-12-24

修改日期:

2019-03-13

PDF Free   EndNote   BibTeX
基于Sentinel-2数据的区域林龄信息反演研究——以落叶松为例
唐少飞, 田庆久, 徐凯健, 徐念旭, 岳继博
南京大学国际地球系统科学研究所
摘要:

林龄结构信息能够有效反映区域森林群落不同生长阶段的固碳能力,对于评估森林生态系统的健康状况具有重要意义。目前,针对区域优势树种林龄结构反演存在技术不成熟、模型参量单一化以及林龄组反演不完善等问题亟待解决。本研究以中国温带典型优势树种落叶松林为研究对象,分别选择其展叶初期、展叶末期和成熟期时段的Sentinel-2影像,采用多元线性回归(MLR)、随机森林(RF)、支持向量机回归(SVR)、前馈反向传播神经网络(BP)以及多元自适应回归样条(MARS)等5种方法依次构建落叶松林龄反演模型。通过相关性分析首先确定最佳遥感反演季相,并在此基础上根据相关性差异筛选出5个最优特征变量用于模型反演,分别为冠层含水量(CWC),归一化水体指数(NDWI),叶面积指数(LAI),光合有效辐射吸收率(FAPAR)和植被覆盖度(FVC)。研究结果表明,展叶末期为落叶松林最佳遥感反演季相。除植被衰减指数(PSRI)以及成熟期的NDVI、RVI外,落叶松林龄与各指标之间均呈负相关关系,其中与冠层含水量(CWC)的相关性最高,pearson相关系数达到-0.74(p<0.01)。此外,不同模型反演结果表明,随机森林模型(RF)为最佳落叶松林龄估测模型,其训练样本和验证样本预测误差绝对值∣ε∣≤5的累积百分比分别达到0.98和0.71;多元线性回归模型(MLR)的林龄估测结果最差,其训练样本和验证样本预测误差绝对值∣ε∣≤5的累积百分比仅为0.64和0.52,因此非线性模型能更好的解释林龄与建模变量之间的关系。

Research on regional forest age information retrieval using Sentinel-2 data: a case study of Larix gmelinii
Abstract:

The information of forest age structure can effectively reflect the carbon sequestration capacity of regional forest communities at different growth stages, which is of great significance for assessing the health status of forest ecosystems. Currently, there are some problems to be solved urgently, such as immature technology, simplified model parameters and imperfect division of stand age groups in the age structure retrieval of regional dominant tree species. According to the seasonal rhythm of Larix gmelinii forest, which is a typical temperate dominant species in China, the Sentinel-2 images of the early of leaf expansion, end of leaf expansion and mature stages are selected in this study. The retrieval model of Larix gmelinii stand age is constructed by using multiple linear regression (MLR), random forest (RF), support vector regression (SVR), feedforward back propagation neural network (BP) and multiple adaptive regression spline (MARS). Through correlation analysis, the optimal seasonal phase of remote sensing retrieval is firstly determined, and on this basis, five optimal characteristic variables are selected for model retrieval according to the difference of correlation, which are canopy water content (CWC), normalized difference water index (NDWI), leaf area index (LAI), fraction of absorbed photosynthetically active radiatio (FAPAR) and fractional vegetation cover (FVC), respectively. The results show that the end of leaf expansion is the optimal remote sensing retrieval phase. Except for the plant senescence reflectance index (PSRI) and NDVI and RVI in mature stage, there is a negative correlation between the stand age of Larix gmelinii and each index, among which the correlation between the stand age and canopy water content (CWC) is the closest, and the correlation coefficient of pearson reached -0.74 (p<0.01). In addition, The results of different retrieval models show that random forest model (RF) is the best model for estimating stand age, and the cumulative percentage of the absolute value of prediction error (∣ε∣≤5) of training samples and verification samples is 0.98 and 0.71, respectively. Multivariate linear regression model (MLR) is the worst for estimating Larix gmelinii forest age, and the cumulative percentage of the absolute value of prediction error (∣ε∣≤5) between training sample and verification sample is 0.64 and 0.52, respectively. Therefore, nonlinear model can better explain the relationship between stand age and modeling variables.

本文暂时没有被引用!

欢迎关注学报微信

遥感学报交流群 分享按钮