Study On The Identifying Method For Forest Fires Based On Fuzzy Support Vector Machine

Forests are regarded as important ecological resources. However, the forest fires can pose serious threats to the forests. Thus, when the fires occur, it is very important to find out or recognize the sources of fires since in this way, the losses are minimized as far as possible. However, the traditional method of fire detection that is based on the sensor, under the condition of complicated environment or the large space, has some limitations in the respect of accuracy and instantaneity. Instead, the non-contact fire detection technology that is based on the image processing can effectively overcome these limitations. In view of this, this paper aims to study the extraction and recognition methods of fire areas based on the improved fuzzy support vector machine.In the present thesis, the theory of fuzzy support vector machine is firstly studied. Under the framework of machine learning theory and based on the support vector machine, the ambiguity is added to the fuzzy support vector machine, which can reduce the effects of noise data or abnormal samples on normal samples. In this way, the better classification hyper-plane can be got. In this paper, the method of minimum radius of class hyper-sphere is adopted, which make the middle of samples can more conform to the regularities of distribution. In addition, in the present paper, the fuzzy membership degree is improved. When the training samples are close to the class-center samples, these training samples are normal samples, whose degree of membership should be 1. However, when the training samples are far from the class-center samples, they are more possible to become the noise, and the degree of membership of them decrease. In this way, the noise samples are excluded.In addition, the author tries to extract the fire areas based on the fuzzy support vector machine and regards the colors, grayscale mean, standard deviation, and correlation coefficient of fire samples and non-fire sample as the characteristic value of samples. Then, the improved fuzzy support vector machine is adopted to study and train these characteristic, which get the optimal hyper-planes. Afterwards, among the new pictures, the author of the present thesis extract the same characteristic as that mentioned above and classify them with the classification hyper-planes. The result of the experiment can indicate that with the way mentioned above, the fire areas are extracted better and thus the method is of great help to study the features of fires.Moreover, in the respect of fire recognition, the sample pictures in the paper are classified into fire sample pictures and non-fire sample pictures. In addition, the features of color, texture, and wavelet in these pictures are extracted. Then, the improved fuzzy support vector machine is adopted to train these features. In this way, the classification hyper-planes of fires and non-fires are achieved. Afterwards, in the new pictures, the same features as those mentioned above are extracted and these features are substituted into the classification hyper-planes in order to classify these new pictures. At last, the accuracy rate is 98 percent.