Detecção de erosões

Abstract

This work proposes the development of a system for detecting soil erosion in sugarcane plantations, using semantic segmentation techniques based on Artificial Intelligence (AI) and deep learning. Soil erosion represents a significant threat to agricultural productivity, resulting in nutrient loss, impaired plant development, and economic and environmental impacts. Therefore, the early identification of areas susceptible to erosion is essential to mitigate its effects and ensure the sustainability of plantations. To achieve this goal, aerial images captured by drones are processed to generate orthophotos, which are subsequently segmented into patches of multiple resolutions (256×256, 512×512, and 1024×1024 pixels). These patches are used for the training of semantic segmentation models, enabling neural networks to automatically identify erosion areas with high precision. The training process is structured into three main stages: generation and organization of the training, validation, and test sets, followed by the effective training of the models. Subsequently, quantitative metrics such as IoU (Intersection over Union) and F1-Score are applied to evaluate the models' performance. The methods developed in this project allow for a quick and precise analysis of soil conditions, reducing the need for manual inspections, which generally require more time and are subject to subjectivity. Furthermore, the application of these techniques contributes to the continuous monitoring of plantations, optimizing management strategies and reducing costs associated with soil maintenance, thus promoting greater sustainability in sugarcane cultivation.