摘要
针对传统森林资源调查工作量大、时效性低的问题,开发具有较好普适性的森林蓄积量估测模型,以期为森林资源管理决策提供科学依据。以淳安县、临海市为研究区,运用2017年研究区Sentinel-2遥感数据、森林资源二类调查数据和数字高程模型数据,采用最小绝对收缩和选择算子(Lasso)特征选择方法,构建K最近邻算法(K-NN)、梯度提升迭代决策树(GBDT)、极端梯度提升(XGBoost)、梯度增强集成分类器(CatBoost)4种模型和基于单模型的堆叠法(Stacking)融合模型,通过10折交叉验证法检验模型精度,分析特征变量对于模型性能指标的影响。结果表明:(1)在淳安、临海两地的森林蓄积量估测中,CatBoost模型在4种单模型中综合表现最优,具有较好的普适性;(2)特征变量的加入极大提升了模型的决定系数(R^(2)),且均方误差、平均绝对误差和平均百分比误差等性能指标也显著优化;(3)融合模型的平均百分比误差最低为20.24%,较单模型有所提升。Lasso特征选择方法结合Stacking融合模型可以准确地估测森林蓄积量,具有较强实用性。
The forestry resources survey was complex and inefficient, a prediction model with good applicability was developed, which is suitable for forest stock volume prediction to provide a scientific basis for forest resource management and decision-making. Sentinel-2 was used as the source of remote sensing data, and combined with the diatomic data and digital elevation data to predict the forest unit volume of Chun’an County and Linhai City. The extracted factors use Lasso(Least Absolute Shrinkage and Selection Operator) for feature selection, K-NN(K-Nearest Neighbor), GBDT(Gradient Boosting Decision Tree), XGBoost(eXtreme Gradient Boosting), CatBoost(Categorical Boosting) and a single model based Stacking fusion model were used to construct the stock volume estimation model, and the data are verified through 10 folds cross-validation;by introducing feature variables, the impact of feature variables on model indicators is discussed. The results of the experiment indicate that:(1)The CatBoost model has the best overall performance among the four models, and performs well in the prediction of stock volume in Chun’an and Linhai indicating the model has good applicability;(2)The addition of characteristic variables greatly improved the model’s R^(2), while MSE(Mean Square Error), MAE(Mean Absolute Error), MAPE(Mean Absolute Percentage Error) were optimized significantly;(3)The minimum MAPE of the fusion model is 20.24%, which was significantly improved. Therefore, Stacking fusion model combined with the Lasso feature selection method has a high accuracy in estimating forest unit volume with a great practicality.
作者
李坤
吴达胜
方陆明
刘建超
Li Kun;Wu Dasheng;Fang Luming;Liu Jianchao(Forestry Intelligent Monitoring and Information Technology of Zhejiang Province,Zhejiang Agriculture and Forestry University,Hangzhou 311300,P.R.China;Jiyang College,Zhejiang Agriculture and Forestry University)
出处
《东北林业大学学报》
CAS
CSCD
北大核心
2021年第5期59-66,共8页
Journal of Northeast Forestry University
基金
浙江省科技重点研发计划项目(2018C02013)。
关键词
森林蓄积量
类别特征
集成模型
Forest stock volume
Category features
Fusion model