Forest fire detection in still images using image processing techniques

According to recent statistics [1], about 5 million hectares of land in the world and about 550 thousand hectares of land in Europe have been damaged every year by forest fire. Both efficient and effective detection methods for forest fire are therefore necessary for several aspects. Our solution for detecting and tracing forest fire is through unmanned air vehicles (UAV). Generally, UAVs fly over the forest area and automatically detect the forest fire and give the alarm. There is a camera and computer system on the mobile platform. At this thesis we proposed a solution for detecting the forest fire by using image processing techniques on the UAV. The major problem of detecting forest fire on mobile platforms is the unavailability of using motion information and low processing power. The purpose of the proposed study is to select the most suitable color space, features and classifiers for the fire classification. Works, respectively, database imaging fire and smoke region manually marking, 7 different color space conversions, creating histogram and create the lookup table, which denotes the likelihood of each pixel belonging to fire. Based on this information, we mark potential fire regions. Given a set of potential fire regions, we then generate the following features: the average likelihood values, number of pixels, boundary roughness, and variance. The feature set of the image then includes these 5 features: the average likelihood, the average boundary roughness and the average variance for all marked fire regions, the total number of fire pixels and the total number of potential fire regions in the image. Each image is represented by a 70-dimensional feature vector. Because of the curse of dimensionality we use some feature selection algorithm which are principal component analysis, K-means and relative entropy. After selecting the best features, they are applied to the classification process. Finally, the input image is classified as either fire or non-fire through various classification methods, such as SVM, K-Nearest Neighbor, and Neural Networks.