Aerial Forest Fire Detection based on Transfer Learning and Improved Faster RCNN

The timely detection and rapid disposal of forest fires are the key to protecting the ecological environment. However, in the detection of forest fires in the aerial photography scene of UAVs, due to factors such as small fire points in aerial photography, large influence of smoke, and complex aerial photography background, traditional deep learning fire detection algorithms often have the problem of low detection accuracy. To solve this problem, this paper proposes an aerial forest fire detection method based on transfer learning and improved Faster RCNN algorithm. First of all, in view of the lack of forest fire data from the existing UAV aerial photography, we use transfer learning technology to use the convolutional neural network parameters trained on the ImageNet dataset to initialize the convolutional layer of Faster RCNN. Second, Faster RCNN is improved based on feature fusion and attention mechanism. We introduced the feature fusion structure PPM to improve the receptive field of the model, thereby improving the detection accuracy of aerial forest fires. In addition, the attention mechanism is used to replace the original Flatten layer and FC1 and FC2 fully connected layers of Faster RCNN to further improve the learning efficiency and detection accuracy of the model. The experimental results show that our method achieves a detection accuracy of 93.7% when detecting forest fires in UAV aerial images, which is more than 18.0% higher than the traditional method, and meets the real-time detection requirements. This shows that our method can be effectively applied to forest fire detection in UAV aerial photography scenarios, and is expected to play an important role in actual forest fire detection.