Evaluación de técnicas de visión artificial para clasificación automática de imágenes satelitales.

Abstract

Abstract: Forest fires that are not stopped from their initial phase spread easily and destroy everything in their path. The lack of systems that allow monitoring the national coverage of forest areas has undoubtedly been a disadvantage in the face of the adverse effects of climate change in our country. This project arose to respond to the need for a system that allows early detection of forest fires through the analysis and classification of satellite images using artificial vision technologies and deep learning models. Three pre-trained models (Yolov8, ViT and CNN) were evaluated at the level of precision, percentage of errors and execution time, which allowed identifying which is the most appropriate and optimal model for identifying forest fires in real time using satellite images. The results showed that the Yolov8 model is the most suitable for use in multi-objective early reaction systems, with results of 0.15 seconds average execution time and 49% multi-objective confidence.