Publication

Abstract : Early detection of fire is very critical in minimizing the damage caused to the environment and wildlife. Traditional methods for detecting fire and smoke use smoke sensors and lookout stations to monitor a specific portion of the forest cover by identifying the critical points. These methods are slow and inefficient because of the issues in identifying locations for placing sensors and lookout stations. Detection of smoke and fire in forest regions can be done efficiently with the use of image processing and deep learning techniques. In this paper, three different deep learning models are compared. First, a CNN model is designed to classify the images into four classes. The four classes are No Smoke, Low Smoke, High Smoke, and Fire. An inception architecture is chosen for its ability to extract features of an image with varying sizes. Although the performance of the Inception V3 model is better than the CNN model, it was designed and trained for ImageNet dataset. For the model to work optimally for the fire and smoke dataset, the fully connected layer is modified, and the model is retrained as part of the transfer learning process. The performance analysis among the CNN model, Inception V3 model, and the proposed modified Inception V3 model is done using accuracy, precision, recall, and f1-score for detecting smoke and fire in forest cover. The proposed model provides improved performance due to its targeted modifications, and it can be used for static surveillance of forest cover.