IDENTIFICAÇÃO DE EROSÕES EM CULTIVOS DE CANA-DE-AÇÚCAR UTILIZANDO INTELIGÊNCIA ARTIFICIAL.

Abstract

Soil erosion is a critical challenge for agriculture, impacting productivity and environmental sustainability. This study investigates the application of artificial intelligence models for the automatic segmentation of erosive features in sugarcane cultivation areas using drone imagery. Three semantic segmentation models (DeepLabV3+, InternImage, and SegFormer) were evaluated at different resolutions (1024x1024, 512x512, and 256x256), analyzing metrics such as IoU, Precision, Recall, and F1-Score. Results showed that InternImage 1024x1024 achieved the best performance, reaching 67.61% IoU and an F1-Score of 80.55%, with lower VRAM consumption and higher processing capacity. The qualitative analysis reinforced these findings, indicating that models trained at higher resolutions provide better accuracy. The study concludes that AI-based approaches for erosion detection are viable, scalable, and effective, with potential applications in agricultural monitoring and soil conservation programs.