Research On Kernel Function And Bionic Intelligent Algorithm To Light Environment Evaluation System Of Planting Ginseng Under Forest

Support vector machine (SVM) is a kind of machine learning method based on statistical learning theory, and shows many unique advantages in solving small sample, nonlinear and high dimension data. It is applied to many fields, except in ginseng under forest light environment. Kernel is the key in SVM and many scholars in the world are interested in it. It is a hot topic on constructing a new kernel function. Bionic intelligent optimization algorithm is sourced from the study about some natural phenomena. It can solve global optimization problems. Moreover, it has adaptability. Parameters optimization of SVM affects the prediction accuracy and generalization ability. If bionic intelligent algorithm is used to optimize the parameters, the best predictive model will be established.Forest resources are valuable resources, and ginseng is the most important one of them. As a kind of plants living in the cold and humid environment, it is highly sensitive to optical environment. This paper researches of the field of ginseng under the forest planting, and study the question about natural light environment. In the paper, we design a kind of humanity, to facilitate future expansion to other features of the light intensity monitoring system. Based on physical and biological principle in ecological system, system analysis and machine learning method, this paper establishes the light environment impacting ginseng growing dynamic model.The main research contents and conclusions are as follows:(1)Constructing some new type of kernel functions. Research nuclear function principle, and be analysis of common kernel function, such as gauss kernel function polynomial kernel function and perceptron kernel function K(x, xi)=tanh(p(x,xi)+c). Construct these kernels and research their fuction, which arebased ongauss kernel function polynomial kernel function K2(x, xi)-(<x, xi>+c). Through analysising the characteristic curve, gauss kernel function has a strong local learning ability, and the parameters γ influencing the decision function judgement:If γ value is larger, the support vector number is more and the model generalization ability is more. But it can lead to overfit phenomenon while γ value is over a certain range. If γ value becomes smaller, the training set samples were misclassified probability. But it will cause less learning phenomenon if γ value is too small. Polynomial kernel function has a strong global learning ability. The larger d value will lead to enhance the ability of global learning ability. Perceptron nuclear balances local learning ability and global learning ability. Analysising of the special curve, the kernel has the stronger local learning ability, and the global learning ability is weaker. It can increase its global learning ability by decreasing the p value, but the local learning ability is weakened accordingly. The kernel function K(x,xi) sourced from the K(x,x')=K1(x,x')+K2(x,x') mode structure has local and global learning ability, the local learning ability is impacted completely by the gauss kernel function and global learning ability is impacted completely by the polynomial kernel function. If the kernel function K(x, xi) is sourced from the K(x,x')=λK*(x,x') mode structure, its performance depends solely on the baseline function property. If the baseline kernel function used K1(x,xi), K(x,xi) is only to improve local learning ability; If the baseline kernel function used K2(x,xi), K(x,xi) is only improve global learning ability. If the kernel function K(x,xi) is sourced from the K(x,x')=K1(x,x')K2(x,x') mode structure, its feature curve is similar with K1(x,xi)'s, which shows the strong local learning ability and the weak global learning ability. By adjusting the K1(x,xi) and K2(x,xi) parameters, it is improved to the local learning ability, but have little effect on the global learning ability. If the kernel function K(x,0.2) sourced from the model structure, it can adjust its local learning ability and learning ability of the global by adjusting the ai (i=1,2) value.(2)Constructing a new type of intelligent bionic algorithm--Tracing Target Algorithm. By using of Needle-in-a-haystack function and Schaffer function, it test the global optimization ability about genetic algorithm, particle swarm optimization algorithm and tracking algorithm. The Needle-in-a-haystack functionhas a global minimum f(0,0)=-3600in the range of-5≤x≤5,-5≤y≤5, and the Schaffer function has a global minimum f(0,0)=0in the range of-20≤x≤20,-20≤y≤20, i=1,2. When individual numbers are20, the maximum algebras are200, variable binary digits are25, the probability of crossover is0.9and mutation probability is0.08, genetic algorithm finds the best fitness value to global optimal solution-3600of Needle-in-a-haystack function through genetic algebra of40generations later, and finds the best fitness value into a local optimal solution0.085of Schaffer functions through genetic algebra of20generations later. When the evolution algebras are200, population sizes are20, particle swarm algorithm finds the best fitness into local optimal solution-2500of Needle-in-a-haystack function after a number of generation is10, and finds the best fitness tend to global optimal solution0of the Schaffer function after a number of generation is60. When population sizes are200, tracing the distance are20, Tracking algorithm finds the best fitness tend to global optimal solution-3600of needle-in-a-haystack functions after searching for algebra of20generations, and finds the best fitness tend to global optimal solution0of schaffer functions after searching for algebra of20generations later.(3)By using of TSL2561being sensitive to visible light characteristics and ATMega16L with I2C and SPI bus function, this paper designs the ginseng under forest light intensity real-time monitoring system, which is the host-from machine architecture and combined with the monitoring software of the host computer. The transmission distance is1000m between the master and slave in the system using a single host and multiple from machine SPI bus connection method, which is more convenient to increase the test unit monitoring point in the future. The method is simple and the light intensity transmission data have little interference with these outside factors, so the system is suitable for ginseng under forest light environment in which the light intensity is measured. It measures light intensity by putting these slaves and TNHY-9monitor in the standard illumination environment. In the process of data transmission, it ensures the reliability of data transmission by joining the check codes, and it distinguishs different test unit of forest bases by the serial slaves numbers. When the distance is500m between master and slave transmission, except for the measured light intensity value of341lux by the number5slaver, the measured light intensity value is340lux by the other slavers, and the variance is8.5. When the distance is1000m between master and slaver, the light intensity are different among all slavers, its variance is11.5. The performance of the system is decreased for the distance increasing between Master and slaver, but in the whole, the system of the light intensity data transmission is considerably reliability when the distance is not more than1000m between master and slaver. The real time monitoring system adjusts the sampling frequency based on requirements of measurement accuracy of light intensity in the forest light environment. It provides a new method for the measurement of light environmental data under the forest.(4)This papaer uses support vector machine to establish the forecasting model, and predicts individual net photosynthetic rate (Pn) by using of the visible light spectral composition proportion relation, and predicts Photosynthetic active radiation (PAR) by using of Direct radiation (PFDdir) and scattering radiation (PFDdif). It provides a kind of new method for light environmental forecasting and evaluation under the forest. This paper uses the epsilon-SVR formula, the formula of nu-SVR, linear kernel function (K1), polynomial kernel function (K2), radial basis function kernel function(K3), kernel function kernel function K1K2, kernel function K1K3, kernel function K2K3, kernel function K1K2K3, penalty parameters c and gamma optimized by using grid-search, genetic algorithm, particle swarm algorithm and tracking algorithm parameters optimization, and establishes different support vector model by the above combination. Mixed the other influencing factors called the ε particle, NRTA model for predicting Pn is the best optimal model. The fitting degree is90.903%when NRTA model predicts Pn sourced from August14, to August28,2011. EGSK(0.1,0,0.9) model for predicting PAR is the best optimal model and the fitting degree is86.897%when this model predicts PAR sourced from July21to July30,2010.