Growth Modeling For Chinese Fir Plantation Based On Artificial Neural Network

Recently, the mathematical model approach is receiving increasing attention fordescribing stand growth rule. However, the traditional theoretical growth model have somedisadvantages, for example, re-parameterization method result in systematic deviation easily.Artificial neural netword (ANN) models are becoming a very popular estimation tool for fittingnonlinear relationship, since they require no assumptions about the form of a fitting function,non-Gaussian distributions, multicollinearity, outliers and noise in the data. The problem ofback-propagation model are lack of ecological explain, and how to determine the structure ofthe network, over-learning and owe-learning problems, fell into the local minimum.According to remeasureed data of15permanent sample plots located at Dagangshanmountain in Jiangxi Province from1981to2008, ANN models for dominant height, averageDBH, and basal area growth were builded on the base of three kinds of predictor factors, i.e.,stand age, density, site index. Growth modeling at different plots by ANN provides scientificoperating decision for Chinese Fir [Cunninghamia lanceolata (Lamb.) Hook.] plantation. Theresults are listed as following:(1) Nilson density index (SD) based on the selt-thinning theory is more suitable as asingle measure for the Chinese fir plantations than the stand density index (SDI) and Curtisrelative density (RD). however, the measurement are deficient in assess all state stands becauseof the complexity of self-thinning theory. The actual value of stems density and averagediameter of basal area as a new density index is used for ANN model regardless of thecorrelation of input variables.(2) The paper mainly focuses on ANN model with back-propagation algorithm (BP) toreveal relationships of the relative importance between stand factors and its basal area, insteadof excessively pay attention to the result as before. In Neural Interpretation Diagram (NID), thethickness of lines connecting the units according to the relative values of their weights roughly showed that the influence of site index and stand age for the basal area are smaller than meanDBH and stems. Sensitivity analysis of basal area growth model demonstrated that the overallstand basal area increased smoothly with stand age, in detail, different site and stand eventuallytend to a different steady value; stand basal area existed in three kinds of performance statuswith stems density increased, i.e.,"S" shaped growth for all forest in first step, the followingstate occured differentiated phenomen, equilibrium, decline and rise; The effect curve ofaverage diameter at breast height on basal area grew along the lines of "S" shaped; Site indexwas of very weak effect on the basal area, the response curve was almost parallel toX-coordinate. Garson algorithm quantified the relative importance of predictor factors, theorder of contributed percentage was stems density(31.81%)> average DBH (30.95%)>standage (30.53%)>site index (6.7%). To a certain extent, the explanations can overcome the "blackbox" theory of ANN, has great potential to be applcable to the field of forestry.(3) Different from those traditional BP model based on empirical risk minimization, basalarea model established by support vector machine (SVM) for Chinese fir plantation employedstructural risk minimization rule. The SVM model overcome some problems of traditional BPby inner product of kernal function and Lagrange function dual conversion, such asdetermination of topologies and local minimum. And training the model by N-foldcross-validation also improve its performance. Practice parameter of SVM model is asfollowing: coefficient of determination(R2=0.993), Mean square error (4.6338),Sum of squaredresidual (SSE=1112.1), the corresponding value of BP is0.991,4.7144,1131.5respectively. themethod has not been reported in the forestry field.(4) Nonlinear time series BP model for basal area of Chinese fir plantation explored thepredicted values at future time by ANN. In the auto-regression BP model, the best model(R2=0.85) was built by lagged one step, the accuracy of the model is not increasing withlagging steps increases. multiple time series regression model could weakly advance theperformance of pridiction (R2=0.87). the improvement can be nelected relative to causalitysimulation (R2=0.993).