不同特征筛选方法和估测模型对天然次生林郁闭度估测的影响
Effects of Different Feature Selection Methods and Estimation Models of Canopy Closure in Natural Secondary Forests
-
摘要: 为探讨不同特征筛选方法及估测模型对天然次生林郁闭度估测的影响,以帽儿山实验林场63块(0.09 hm2)样地为研究区域,使用机载激光雷达(Airborne Laser Scanning, ALS)数据提取特征,采用3种特征筛选方法(Pearson相关性分析、随机森林(Random Forest, RF)和Boruta算法)和3种模型(偏最小二乘回归(Partial Least Squares Regression, PLSR)、随机森林回归(Random Forest Regression, RFR)和支持向量回归(Support Vector Regression, SVR))估测郁闭度,以2种测定方法(鱼眼照片和样点法)进行精度评价及方差分析(analysis of variance, ANOVA)。研究结果表明,PLSR估测精度最高,SVR最低;以RF和Boruta算法筛选后的估测精度高于Pearson相关性分析,其中Boruta算法特征筛选后的PLSR模型估测精度最高(R2=0.451 1,RMSE=0.067 5);鱼眼照片和样点法的变量筛选和鱼眼照片法的模型对郁闭度均无显著影响,样点法中的估测模型对郁闭度有显著影响。此研究表明PLSR估测精度优于其他模型,样点法的估测模型对郁闭度有显著影响。ALS可有效地估测天然次生林的郁闭度,为大范围估测森林郁闭度提供依据。Abstract: In order to explore the influence of different feature selection methods and estimation models on canopy closure estimation of natural secondary forests, 63(0.09 hm~2) sample plots in Maoershan Experimental Forest Farm were taken as the research area, and Airborne Laser Scanning(ALS) data was used to extract features. Three feature selection methods(Pearson correlation analysis, Random Forest(RF) and Boruta)) and three models(Partial Least Squares Regression(PLSR), Random Forest Regression(RFR) and Support Vector Regression(SVR)) were employed to estimate canopy closure, and the accuracy evaluation and analysis of variance(ANOVA) of the model were carried out by fisheye photo method and sample method. The results showed that the accuracy of PLSR was the highest and SVR was the lowest. The estimation accuracy of RF and Boruta algorithm was higher than Pearson correlation analysis. The PLSR model after Boruta feature selection had the highest accuracy(R~2=0.451 1, RMSE=0.067 5). The feature selection of fisheye photo and sample method, model fitting method of the fisheye photo had no significant effect, yet the model of sample method had a significant effect on estimating canopy closure. The results indicated that the accuracy of PLSR was better than others. The model of the sample method had a significant effect on canopy closure. ALS can effectively estimate the canopy closure of natural secondary forests, which provided a basis for large-scale estimation.