Research On The Estimation Method Of The Chinese Fir Volume Based On BP Neural Networks

Forest resource is a comprehensive ecological system, the growth of which is affected by both genetic information and the environment. As the forest volumn is the main quantitative indicator of the forest condition, so grasping forest resources dynamics in time is important to the monitoring and management of forest. Taking the artificial fir of state-owned forest farms in Baisha, Minhou County of Fuzhou as the study object, This study obtained estimated information from the digital elevation model and the high-resolution remote sensing image-QuickBird by use of3S technology. Then it combined with the field investigation data of Chinese fir volume prediction model based on the artificial neural network in order to provide the support of method and decision making about Chinese fir volume inversion and management.This study obtained the plot spectral information from remote sensing images and the terrain factors that affect the Chinese fir growth from digital elevation models. These information was combined with forest survey factors by field investigation and could be the estimated factors of Chinese fir volume prediction model. As the field sample plots in images became shadow of different degrees, both the shadow detection of normalized RGB color model and the shadow compensation method by analyze information of non-shaded area adjacent can be used to preprocess images. And then the band value of plots and the various types of vegetation indexes could be extracted. As the difficulty of identifying the canopy of forest with high coverage on the image, the range value of semivariogram could be an effective estimation of the average crown of stand according to the repetitive characteristics of spatial structure shown by the crown in images, that whose validity could be verified by field surveys. The site conditions of stand affect the growth of vegetation and biomass directly, so the topographic information of tree growth, such as elevation, slope, aspect, slope position and so on. They could be obtained from the digital elevation model as the terrain factors for stock volume estimation. As the factors with multicollinearity of different levels, this study used ridge trace analysis and stepwise regression analysis to optimize the selection of various factors to determine the best variables set.This paper, choosing BP neural network and partial least squares as the estimation method, major described the learning process and training strategies of neural network, and introduced the analysis process of network performance in detail. The specific content was that the46standard plots randomly selected to be used as training sample in the study area, then it was optimized by the BP network model by4aspects, including the basic structure of the network model, the influence of factor optimized performance, the number of neuron in hidden layer of training function and the transfer function, fitted the non-linear relationship between Chinese fir volume and various factors. Finaly the BP network prediction model of Chinese fir volume with the structure of10:3:1could be determined, which brought the overall prediction accuracy of88.5%obtained by testing the sample. Compared with the traditional method of partial least squares, it could lead to a good predicted result with high fitting accuracy and generalization ability, and reflect the superiority and application value of the non-linear approach.