Research On Support Vector Machine Algorithm And Its Application In Forest Health Monitoring System

Compared to traditional identification tools, such as neural networks, decision treesand Bayesian networks, support vector machine is established on a rigorous statisticaltheory and the basis of the convex programming, which makes it has better generalizationability and character to obtain the global optimal solution than other computationalintelligence algorithms. Compared with the normal linear machine learning method, in theprocess of solving nonlinear problems, it avoids the curse of dimensionality when mappingnon-linear data sets to a high dimensional linear feature space by a vector inner product,solution algorithm of the original problem is unsuitable for using in the data sets with alarge number and high dimensions due to excessive demands for memory-consuming,which makes it can’t meet the requirements for real-time and low error that the videoimage monitoring should have in the forest health monitoring system when the forest firehappens.This paper focuses on the boundary extraction algorithm, improvement ofdecomposition, chunking and parameter selection algorithm of support vector machines inthe kernel space in order to improve the processing performance of support vectormachines in forest health monitoring system. Traditional border extraction algorithm is inthe kernel space after the mapping, but the measure of the singularity makes it impossibleto access to good effect on data sets of different distribution, so in the kernel space newmeasurement model is proposed. Existing block decomposition algorithm could make thedecision-making functions deviate from the optimal solution to a large extent, fuse theimprovement of the existing decomposition algorithm and the chunking algorithm toimprove the efficiency of the algorithm and make the decision function can be more closeto the optimal solution. In addition, optimize parameters from the angle in which definesthe prediction accuracy of new algorithm by the ratio of singular points, avoid thegeneration of the middle training, improve the efficiency of searching parameters. Multipledata sets from UCI and Statlog, simulation results illustrate the effectiveness of the newalgorithm.In forest health monitoring system, the important part is the recognition of forest fires,which is the main application in this paper. Some is based on the smoke feature in themajor existing monitoring system, but at night, due to the bright contrast, smoke often cannot be recognized; Furthermore, the existing recognition system easily be confused withthe active light, according to the dynamic changes when the fire broke out and theirregularity of the fire zone, this system obtains the dynamic characteristics of the fireregion and irregular measure, which makes up parts of the features of the fire detection,itis designed with two layers of support vector machine, obtain a good accuracy andspeed in the recognition of the forest fires in experimental verification of many monitoringvideos.