基于机器学习的杉木树高-胸径模型研建
Research and construction of a height diameter model for Cunninghamia lanceolata based on machine learning
[目的]基于TensorFlow和Sklearn平台利用深度神经网络(DNN)、支持向量回归(SVR)和随机森林(RF)3种机器学习算法建立杉木的树高-胸径模型,探求一种更加高效的杉木树高预测方法。[方法]以清镇国有林场 49 块样地中杉木的胸径、树高数据,分成不同比例的训练集和测试集组,分别为建模数据占比(20%、30%、40%、50%、60%、70%、80%)和检验数据占比(80%、70%、60%、50%、40%、30%、20%)。利用DNN、SVR和RF拟合树高-胸径模型,并将 3种算法拟合的模型与 11个传统基础模型进行比较,通过R2、RMSE和MAE的结果对比选出预测效果最好的模型。[结果]利用DNN和SVR建立的树高-胸径模型在训练集较少的情况下,其预测精度R2分别能达到 0.671和 0.65以上。在训练集占比 40%时,传统基础模型才能在精度上略高于机器学习的方法。在基础模型的预测中,精度最高的为在训练集占比20%情况下的DNN模型,其R2为 0.671,RMSE和MAE分别为 2.024和 1.599,在所有模型中,其预测精度最高。而SVR预测精度最高在训练集占比 30%时,其R2为 0.658,RMSE和MAE分别为 2.072和 1.652.对DNN中的Dropout、L2、Reduce和Relu四种模型进行其预测值与真实值进行残差分析,四种模型的拟合分布符合客观规律。[结论]使用深度神经网络、支持向量回归和随机森林构建杉木的树高-胸径模型,并且通过与传统模型的对比,得出深度神经网络和支持向量回归在训练集占比较少的情况下,其预测精度要高于传统基础模型。而利用DNN构建的模型比传统基础模型在某些特定情况下可以不用受到传统统计学的限制,能够更加简单的使数据收敛,为精准预测树高提供更好的可能。从而能够在数据量较少的情况下精准预测出杉木的树高。
[Objective]Based on TensorFlow and Sklearn platforms,three machine learning algorithms,namely deep neural network(DNN),support vector regression(SVR)and random forest(RF),were used to establish the tree height-DBH model of Chinese fir,and a more efficient tree height prediction method was explored.[Method]The DBH and tree height data of Chinese fir trees in 49 plots of Qingzhen State-owned forest farm were divided into training set and test set groups with different proportions,which were respectively the proportion of modeling data(20%,30%,40%,50%,60%,70%,80%)and the proportion of inspection data(80%,70%,60%,50%,40%,30%,20%).DNN,SVR and RF were used to fit the tree-height-DBH model,and the models fitted by the three algorithms were compared with 11 traditional basic models.The model with the best prediction effect was selected by comparing the results of R2and RMSE and MAE.[Result]The tree height diameter model established using DNN and SVR can achieve R2 prediction accuracy of 0.671 and 0.65,respectively,when the training set is small.Only when the training set accounts for 40%can traditional basic models have slightly higher accuracy than machine learning methods.In the prediction of basic models,the DNN model with the highest accuracy is the one with a training set proportion of 20%,with an R2 of 0.671,RMSE and MAE of 2.024 and 1.599,respectively.Among all models,its prediction accuracy is the highest.The highest prediction accuracy of SVR is achieved when the training set accounts for 30%,with an R2 of 0.658,RMSE and MAE of 2.072 and 1.652,respectively.Residual analysis was conducted on the predicted and true values of Dropout,L2,Reduce,and Relu models in DNN,and the fitting distribution of the four models followed objective laws.[Conclusion]Deep neural network,support vector regression and random forest were used to construct a height-DBH model of Chinese fir.Compared with the traditional model,it was concluded that the prediction accuracy of deep neural network and support vector regression was higher than that of the traditional basic model when the training set was relatively small.Compared with the traditional basic model,the model constructed by DNN can not be restricted by traditional statistics in some specific cases,and can more easily converge the data,providing a better possibility for accurate prediction of tree height.Thus,the tree height of fir can be accurately predicted in the case of a small amount of data.
王贵林;谭伟;
贵州大学林学院 贵州 520025##贵州大学林业信息工程研究中心 贵州 520025;贵州大学林学院 贵州 520025##贵州大学林业信息工程研究中心 贵州 520025;
S757
Cunninghamia lanceolata DNN SVR RF Height-diameter model
225-235 / 11
评论