Detecção automática de focos de incêndio utilizando visão computacional

Abstract

The increase in the frequency and intensity of forest fires has caused significant environmental damage, primarily due to the uncontrolled emission of greenhouse gases and illegal burnings. This highlights the urgent need for efficient methods for the early detection of these events. This work proposes an approach based on computer vision and convolutional neural networks (CNNs) for the classification of forest fire hotspots. Using the MobileNetV3 architecture, known for its performance efficiency and low resource consumption, the model was trained with the DeepFire dataset, which contains real images of forest fires, applying advanced image preprocessing techniques to augment the dataset and improve image quality. Additionally, the use of histograms in the HSV color space, either alone or in combination with features extracted from the images, was explored as a rich visual characteristic for fire detection, emphasizing the importance of the type of input data on the model’s performance. At the end of the study, five distinct models were developed, varying in their layer structure and type of input data. The best model achieved a high accuracy, but there is potential for the development of even more efficient models. This study demonstrates that the application of CNNs can be a powerful tool in the early detection of forest fires, contributing to the mitigation of environmental damage.