Influencing factors and growth state classification of a natural Metasequoia population

Abstract

By analyzing the importance of influencing factors and conducting a comparative study of the effects of different sorting algorithms, a new method is proposed that is suitable for classifying the growth state of a natural Metasequoia glyptostroboides Hu and W.C. Cheng population. We studied 2817 M. glyptostroboides trees over 100 years old and analyzed their growth state by measuring 15 factors from stumpage, site condition, and environmental data. The dimensionality of all factors were reduced using the random forest algorithm, and we classified the remaining factors using the following algorithms: random forest, back-propagation (BP) neural networks, and support vector machine (SVM). The applicability of each sorting algorithm was analyzed. When all the d factors are used for classification and modeling, the model’s overall accuracy, kappa coefficient and test accuracy were 85.5%, 0.739 and 85.8%, respectively. By reducing the dimensionality of the factors using the random forest algorithm, 11 factors most strongly influenced the classifications of the growth state of the Metasequoia population: diameter at breast height, height, crown width, age from stumpage data; longitude, latitude, elevation, slope aspect, gradient and slope position from the site condition data; and the edge of the field from the environmental data. For classifying the Metasequoia population, the random forest algorithm has the highest overall accuracy at 87.2%, which is 3.4 and 2.3% higher than the BP neural networks and SVM algorithms, respectively. The SVM algorithm is superior to the random forest algorithm with respect to classifying the state of mortality. The combination of the random forest and SVM algorithms and their combined information can be used to classify and predict the growth state of this natural M. glyptostroboides population to provide a scientific basis for its effective protection.