Forest fire and smoke detection using deep learning-based learning without forgetting

Background: Forests are an essential natural resource to humankind, providing a myriad of direct and indirect benefits. Natural disasters like forest fires have a major impact on global warming and the continued existence of life on Earth. Automatic identification of forest fires is thus an important field to research in order to minimize disasters. Early fire detection can also help decision-makers plan mitigation methods and extinguishing tactics. This research looks at fire/smoke detection from images using AI-based computer vision techniques. Convolutional Neural Networks (CNN) are a type of Artificial Intelligence (AI) approach that have been shown to outperform state-of-the-art methods in image classification and other computer vision tasks, but their training time can be prohibitive. Further, a pretrained CNN may underperform when there is no sufficient dataset available. To address this issue, transfer learning is exercised on pre-trained models. However, the models may lose their classification abilities on the original datasets when transfer learning is applied. To solve this problem, we use learning without forgetting (LwF), which trains the network with a new task but keeps the network’s preexisting abilities intact. Results: In this study, we implement transfer learning on pre-trained models such as VGG16, InceptionV3, and Xception, which allow us to work with a smaller dataset and lessen the computational complexity without degrading accuracy. Of all the models, Xception excelled with 98.72% accuracy. We tested the performance of the proposed models with and without LwF. Without LwF, among all the proposed models, Xception gave an accuracy of 79.23% on a new task (BowFire dataset). While using LwF, Xception gave an accuracy of 91.41% for the BowFire dataset and 96.89% for the original dataset. We find that fine-tuning the new task with LwF performed comparatively well on the original dataset. Conclusion: Based on the experimental findings, it is found that the proposed models outperform the current state-of-the-art methods. We also show that LwF can successfully categorize novel and unseen datasets.