Fire Detection of Satellite Remote Sensing Images Based on VGG Ensemble Classifier

Abstract

It is an important application of satellite remote sensing image to detect fire spots timely. Traditional methods mostly used the threshold judgment on the mid-infrared or short-wave infrared bands, with a low accuracy. In order to improve the detection performance, we proposed a new fire detection method based on Landsat-8 images. The method included two steps. First, threshold judgment based on the normalized burning ratio Short-wave (NBRS) was used to determine the fire spot candidates, and a large number of non-fire pixels were quickly removed. After that, the block images formed by four kinds of band combinations were extracted at the candidate spot positions. Four detection models were constructed using adjusted VGG network in this paper. The final fire detection results were derived through weighted voting of the classification results of the four models. For it is difficult to obtain fire spot image samples, transfer learning was used in VGG model training. The proposed method was tested on six untrained remote sensing images. 83% of the detected fire spots were correct and the missing rate was 5%. The experimental results showed that our proposed method not only improved the detection accuracy, but the method was more robust than traditional methods.