Ferramenta de Detecção de Focos de Incêndios via Deep Learning em Imagens Orbitais

Abstract

Forest fires have caused great ecological and economic losses in Brazil. When there is low relative humidity, or even excessive heat, a small spark of fire, carried by the wind, can become a fuse for fires in open environments, such as environmental reserves and farms. This work uses computational resources and object recognition techniques in images acquired by sensors on board satellites with the aim of providing information in an efficient way that helps in the identification of areas under fire and, therefore, aid in decision-making. One field of computing that shows promise is research with Convolutional Neural Networks (CNNs), which has been increasing with the emergence of numerous object detection algorithms and which are used here for fire detection. In this way, this work developed a framework that use deep learning models and Application Programming Interface (API) to help identify fire and burning hotspots. These tools consist of software that collects images from three satellites and centralizes the data collected in the cloud and distributes them in an API for access to the collected data. We trained six CNN-based and Tranformer-based models using the MMDection framework, from which we obtained robust models for a variety of three sensors. Finally, a web tool with an interactive map of the Pantanal region indicating possible fire hotspots and their coordinates for consulting results. In this study, we present recent state-of-the-art trained models, previously chosen with the best result for the satellite group in fire identification. At the present time, we are working with the sensors of the CBERS04, CBERS04A and Amazonia 1 satellites. With this, we have developed a set of tools that we believe can be effective in helping to combat fires. We also provide a centralized and public database that can be consulted by other researchers. This study may indicate a direction for future research in this area, we can prove that deep learning models obtain good results for the chosen sensors in detecting fire in open fields, in addition to providing another tool for firefighting brigades, farmers and the community.