Classificação de níveis de risco de incêndio em contextos ambíguos: uma avaliação de aprendizado supervisionado e auto-supervisionado

Abstract

Climate threats have been a global challenge for increasing extreme compound events, causing increased impacts and health risks, ecosystems, infrastructure, subsistence means and food. Among the events occurred, the seasons of forest fires are listed in various regions of the planet. A growing that reinforces the importance of implementing efficient strategies capable of detecting and classifying risks with accuracy and agility. In this context, the present study aims to analyze the capacity of different algorithms in classifying fire risk levels in order to identify more robust and generalizable architectures. The results obtained revealed a better performance of the self-suited Dinov2 model, both on the Firerisk reference dataset and in the Eurosat comparison dataset, as evidenced by the evaluated metrics. However, there is still room for improvements, such as the adoption of new techniques, methodologies and more robust architectures to more accurately evaluate Firerisk's complexity.