摘要
以闽西北杉木人工林为研究对象,选取涵盖中龄林、近熟林、成熟林3个龄组的700个小班作为样地进行调查,以林龄、地位指数、林分密度、平均胸径作为输入变量,单位蓄积量为输出变量,运用BP神经网络和支持向量机2种机器学习方法建立林分收获模型,并采用遗传算法对模型参数进行优化。随机将样本数据分成350个训练样本和350个验证样本,对不同模型的拟合精度、预测精度进行对比分析,其中参数优化后的BP神经网络和支持向量机模型训练样本精度分别达到0.935 37和0.936 33,预测结果精度分别为0.921 30和0.926 97,训练样本和验证样本的总体拟合平均相对误差值均低于7%。分析结果表明,2种模型拟合精度高、预测性能好,为杉木人工林林分收获模拟和预测奠定了基础。为比较2种方法预测结果的差异性,将350个验证样本样地平均分为7组,分别用优化后的2种模型计算各组的预测精度,对预测精度与训练精度的差值进行t检验,结果表明,2种建模方法的预测结果不存在显著性差异,但模型精度的提高对森林资源的精确监测和森林生长动态预测具有重要的理论价值。同时,研究发现支持向量机模型的拟合精度和泛化能力均优于BP神经网络,该方法为收获模型研究提供了新思路。
Based on data from 700 sample plots of Cunninghamia lanceolata plantations in the northwestern area of Fujian Province, consisting of middle-aged, near-mature and mature trees, we established a yield model with stand age, site index, stand density and average diameter at breast height （DBH） as input variables and stand volume as the output variable. We used two machine learning methods, i. e. , a back propagation （BP） neural network and a support vector machine （SVM）. The parameters used in both modeling methods were optimized by a genetic algorithm. We randomly divided the plots into two halfs, i. e. , a 350 plot training set and a 350 plot test set and compared the fit and prediction accuracies of both models. After parameter optimization, the accuracies in fitting both models were 0. 935 37 for BP and 0. 936 33 for the SVM, with prediction accuracies of 0. 921 30 for BP and 0. 926 97 for the SVM. The average relative errors of both the training and test sets were less than 7% for both models. We conclude that both models established a basis for simulating and predicting stand yield of C. lanceolata plantations, given that the accuracy of both models was quite high with good performance of prediction. In order to analyze the differences in the results of both models, the 350 plot test set was evenly divided into seven groups. We calculated the prediction accuracy in each group with two optimization models and used a t-test to compare the absolute difference between the prediction and fitting accuracy. The results showed no significant difference between the two methods. All the same, the slight improvement in precision is important and valuable for monitoring forest resources as well as for predicting the dynamics of stand growth. We found that the accuracy of the fit and generalization ability of SVM were better than those of BP. Therefore, we recommend the SVM for providing new insights for research in forest yield models.
出处
《北京林业大学学报》
CAS
CSCD
北大核心
2015年第1期42-47,共6页
Journal of Beijing Forestry University
基金
福建省科技重大专项(2012NZ0001)
关键词
杉木人工林
收获模型
机器学习
预测
拟合
Cunninghamia lanceolata
yield model
machine learning
predicting
fitting
