应用局部样本最优K值KNN模型估测森林蓄积量

Forest Volume Estimation Using KNN Model with Local Sample Optimal K Value

为了探究KNN算法(K-最近邻法)的优化方法及Landsat8 OLI在森林蓄积量估测中的应用潜力。以陕西省留坝县为研究区,采用Landsat8 OLI为遥感数据源并结合同时期的森林资源调查数据,构建多元线性回归(MLR)、K-最近邻(KNN)、随机森林(RF)、距离加权K-最近邻(DW-KNN)和局部样本最优K值KNN(LSO-KNN)模型进行森林蓄积量的遥感估测。随机抽取总样本的2/3用于训练模型,1/3用于模型的检验,并以决定系数(R^(2))、均方根误差(R_(MSE))和相对均方根误差(R_(RMSE))作为精度的检验指标对模型进行评价。结果表明:(1)在构建的5种森林蓄积量反演模型中,4种机器学习模型均高于MLR模型;(2)基于局部样本最佳K值构建的LSO-KNN模型估测结果最佳,其决定系数为0.72,均方根误差为39.58 m^(3)/hm^(2),相对均方根误差为28.68%,均方根误差比MLR、KNN、RF和DW-KNN模型分别降低了30.89%、27.24%、24.23%和18.14%,说明LSO-KNN模型相比于其他模型更适用于森林蓄积量的估测。因此,根据Landsat8 OLI数据的LSO-KNN模型绘制的森林蓄积量空间分布符合实际,可以满足森林资源调查的要求和实现大尺度、长时间的森林资源动态监测。

In order to explore the optimization method of KNN algorithm and the application potential of Landsat 8 OLI in forest stock estimation,we took Liuba County in Shaanxi Province as the research area,using Landsat 8 OLI as the remote sensing data source and combining the forest resource survey data of the same period,a multi-source linear regression(MLR),K-nearest neighbor(KNN),random forest(RF),distance weighted KNN(DW-KNN)and local sample optimal K value KNN(LSO-KNN)models for remote sensing estimation of forest stock.The 2/3 of the total sample is randomly selected for training the model,1/3 is used for model testing,and the coefficient of determination(R^(2)),root mean square error(R_(MSE))and relative root mean square error(R_(R_(MSE)))are used as the accuracy test.The results show that:(1)Among the 5 forest stock inversion models constructed,the 4 machine learning models are all higher than the MLR model.(2)The LSO-KNN model constructed based on the best K value of local samples has achieved the results The best estimation results have R^(2) of 0.72,R_(MSE) of 39.58 m^(3)/ha,and R_(R_(MSE)) of 28.68%.Compared with MLR,KNN,RF and DW-KNN models,R_(MSE) is reduced by 30.89%,27.24%,24.23%and 18.14%,respectively.This research proves that the LSO-KNN model is more suitable for estimation of forest stock volume than other models.The spatial distribution of forest stock volume drawn based on the Landsat8 OLI and LSO-KNN model is in line with the actual situation,and can meet the requirements of forest resource surveys.