基于光学-ALS变量组合和非参数模型的天然次生林地上生物量估算

【目的】通过组合机载激光雷达(airborne laser scanning, ALS)数据和Sentinel-2A数据提取特征变量,探讨估算天然次生林地上生物量(aboveground biomass, AGB)最佳的变量组合方式和估算方法。【方法】以2015年ALS数据、2016年Sentinel-2A数据和黑龙江帽儿山林场森林资源连续清查固定样地数据为数据源,通过ALS数据提取高度特征变量(all the LiDAR variables, 记为A_L),Sentinel-2A数据提取若干植被指数变量(all the optical variables, 记为A_O),然后将光学-ALS结合变量(combined optical and LiDAR index, COLI,记为I_COL)结合成为新的变量 I_{\mathrm{CO}{\mathrm{L}}_1} 和 I_{\mathrm{CO}{\mathrm{L}}_2} ,以6组特征变量组合方式(A_O+A_L、 I_{\mathrm{CO}{\mathrm{L}}_1} 、 I_{\mathrm{CO}{\mathrm{L}}_2} 、 I_{\mathrm{CO}{\mathrm{L}}_1} +A_O+A_L、 I_{\mathrm{CO}{\mathrm{L}}_2} +A_O+A_L、 I_{\mathrm{CO}{\mathrm{L}}_1} + I_{\mathrm{CO}{\mathrm{L}}_2} +A_O+A_L)作为输入变量,分别使用多元线性逐步回归(stepwise multiple linear regression,SMLR)、K-最近邻法(K-nearest neighbor,K-NN)、支持向量回归(support vector regression,SVR)、随机森林(random forest, RF)和堆叠稀疏自编码器(stack sparse auto-encoder,SSAE)共5种方法构建了天然次生林AGB估算模型,探讨I_COLs变量以及不同模型对生物量估测精度的影响。【结果】结合变量I_COLs对于森林AGB的估算十分有效,加入I_COLs变量能够很大提高森林AGB模型的估算精度;与其他4种模型相比,无论使用哪些变量作为输入数据,SSAE模型的精度最高;当使用SSAE模型,以光学和ALS变量组合 ( I_{\mathrm{CO}{\mathrm{L}}_1} + I_{\mathrm{CO}{\mathrm{L}}_2} +A_O+A_L)作为输入特征变量时,模型的准确性最高:R²=0.83,均方根误差为11.06 t/hm²,相对均方根误差为8.23%。【结论】结合变量COLIs能够有效地提高天然次生林AGB的估算精度,而且深度学习模型(SSAE)在估算天然次生林AGB方面优于其他预测模型。总体而言,利用ALS和Sentinel-2A数据组合变量的SSAE模型可以较准确地估算森林AGB,为天然次生林地上生物量的估算和碳储量评估提供技术支持。

Estimation of aboveground biomass of natural secondary forests based on optical-ALS variable combination and non-parametric models

【Objective】 We explored feature variables extracted through a combination of airborne laser scanning (ALS) and Sentinel-2A data, and investigated with the aim of identifying the best variable combination mode and estimation method for estimating the forest aboveground biomass (AGB) of natural secondary forests. 【Method】 Based on ALS data from 2015, Sentinel-2A data, and the fixed sample plots of the continuous inventory of secondary forest resources from Maoershan Forest Farm in 2016, this study extracted height features from ALS data (all the LiDAR variables, AL), several vegetation indices from Sentinel-2A (all the optical variables, AO), and then combined the two kinds of variables into new variables (COLI₁ and COLI₂). Finally, five models were constructed, including the stepwise multiple linear regression (SMLR), K-nearest neighbor (K-NN), support vector regression (SVR), random forest (RF), and stack sparse auto-encoder (SSAE) of AGB for natural secondary forests using six feature combinations (AO+AL, COLI₁, COLI₂, COLI₁+AO+AL, COLI₂+AO+AL and COLI₁+ COLI₂+AO+AL). The influence of the COLIs variable, and that of different models, on the accuracy of the AGB was investigated. 【Result】The COLIs variable could efficiently improve the accuracy of AGB estimates and, compared to the other four models, SSAE had the highest accuracy regardless of the variables. The SSAE model with the combination of optical and ALS features (COLI₁+ COLI₂+AO+AL) had the best model performance of R² which was 0.83, RMSE was 11.06 t/hm², rRMSE was 8.23%. 【Conclusion】 The combined variable COLIs can effectively improve the estimation accuracy of natural secondary forest AGB, and one way of the deep learning model (SSAE) is superior to the other prediction models in estimating the AGB of a natural secondary forest. This conclusion will help to further apply the deep learning model to draw a large-area AGB spatial distribution map and estimate the other forest parameters. In general, the SSAE model with the combination of ALS and Sentinel-2A data could estimate AGB more accurately than the other models. This finding provides a technical support for AGB estimation and carbon evaluation of natural secondary forests.