Macaúba – IA: Impulsionando a produção sustentável de biocombustível com identificação automatizada de plantas através de inteligência artificial

Abstract

The replacement of fossil fuels with renewable energy sources has gained significant attention in recent times. In this context, the macaúba palm (Acrocomia aculeata) emerges as an excellent alternative due to its high potential for biofuel production and its capacity to serve as a source of plant-based extractivism, given its widespread occurrence in natural environments. However, accurate mapping and quantification of macaúba populations for extractivism purposes remain challenging, hindering optimized economic utilization. Thus, this study aimed to identify and count macaúba plants in native environments using drone imagery combined with artificial intelligence models implemented through YOLOv8. Additionally, it sought to evaluate which version (nano, small, medium, large, and extralarge) is most effective in detecting macaúba plants. Images were collected from two regions with natural occurrences of macaúba in the municipality of Lavras-MG, Brazil, using an RGB camera mounted on a DJI Mavic 3 drone. A total of 1,921 images were captured and divided into training (70%), testing (20%), and validation (10%) datasets, with each network trained over 100 epochs. The results demonstrated high performance across all YOLOv8 versions applied to macaúba mapping, with accuracy exceeding 95%. Although training was set for 100 epochs, the nano version achieved significant learning within 52 epochs. All networks displayed high mAP values, reaching up to 98%, with the large version achieving this value earlier, at just 74 epochs. Regarding loss, the xlarge version was the first to reach minimum values but exhibited the poorest performance during testing, while the large version excelled with an accuracy rate of 94.71% on test images. These findings highlight the robust potential of YOLOv8 networks for mapping and detecting plants using drones, with the large and nano versions standing out.