应用SHAP可解释机器学习模型估测森林蓄积量  

Estimate the Forest Volume Using the SHAP Interpretable Machine Learning Model

作　　者：王元 王玥 周宇琛 陈伏生[1] 张绿水[1] 刘牧 Wang Yuan;Wang Yue;Zhou Yuchen;Chen Fusheng;Zhang Lvshui;Liu Mu(Jiangxi Provincial Key Laboratory of Conservation Biology(Jiangxi Agricultural University),Nanchang 330045,P.R.China)

机构地区：[1]保护生物学江西省重点实验室(江西农业大学),南昌330045

出　　处：《东北林业大学学报》2025年第5期66-73,共8页Journal of Northeast Forestry University

基　　金：江西省高校人文社会科学研究项目(JC22245)。

摘　　要：森林蓄积量是反映森林资源丰富程度的关键指标,精确估测森林蓄积量对于森林资源管理至关重要。以江西省林区为研究对象,运用谷歌地球引擎(Google Earth Engine)平台从Landsat 8遥感影像中提取多个植被指数、单波段及组合特征,并结合国家森林资源连续清查的地面实测数据,分析不同影像特征参数在森林蓄积量反演中的贡献率。结果表明:对比多元线性回归、神经网络、随机森林和XGBoost模型估测森林蓄积量的精度,随机森林模型估测精度为93.3%,决定系数(R^(2))为0.9337,均方根误差为2.2323,平均绝对误为2.3395;与BP神经网络模型(R^(2)=0.8219)和XGBoost模型(R^(2)=0.7916)相比,模型拟合度和预测效果更佳,比多元线性回归模型(R^(2)=0.688)处理非线性关系的稳定性和可靠性更高。通过解释特征参数的相对重要性,揭示出平均胸径、郁闭度等特征对森林蓄积量影响显著,且随机森林模型中各因子间存在相互作用。The forest volume is a key indicator reflecting the richness of forest resources,and accurate estimation of the forest volume is of great significance for forest resource management.Taking the forest areas in Jiangxi Province as the research object,the Google Earth Engine(GEE)platform was used to extract multiple vegetation indices,single bands,and combined features from Landsat 8 remote sensing images.Combined with the ground measurement data from the national continuous forest resources inventory,the contribution rates of different image characteristic parameters in the inversion of forest volume were analyzed.The results show that:when comparing the accuracies of the multiple linear regression,neural network,random forest,and XGBoost models in estimating the forest volume,the estimation accuracy of the random forest model is 93.3%,the coefficient of determination(R^(2))is 0.9337,the root mean square error is 2.2323,and the mean absolute error is 2.3395.Compared with the BP neural network model(R^(2)=0.8219)and the XGBoost model(R^(2)=0.7916),the random forest model has a better goodness of fit and prediction effect.It also has higher stability and reliability in handling nonlinear relationships compared with the multiple linear regression model(R^(2)=0.688).By explaining the relative importance of characteristic parameters,it is revealed that features such as the average diameter at breast height and the canopy density have a significant impact on the forest volume,and there are interactions among various factors in the random forest model.

关 键 词：SHAP解释模型 机器学习模型 森林蓄积量 

分 类 号：S758.4[农业科学—森林经理学]