Detecção de grãos de soja pós-colheita utilizando visão computacional e aprendizado profundo

Abstract

This work presents the development and validation of a deep learning-based computer vision system for automatic detection of soybean grain losses on the ground after harvest. Accurately monitoring these losses is a critical challenge in precision agriculture, tradi- tionally performed through manual methods that prove time-consuming, imprecise, and subjective. A solution was proposed using YOLOv5 and YOLOv11 architectures with the objective of automatically identifying and locating grains scattered on the soil under different field conditions. The methodology involved collecting 230 high-resolution images in a real-world environment, manual annotation of samples, application of a rigorous preprocessing and data augmentation pipeline, and training and optimization of multiple models varying architecture, resolution, and dataset. The results demonstrated that the YOLOv11s model trained with Dataset V2 achieved an mAP@0.5 of 92.7% and a Recall of 86.2%, significantly surpassing the YOLOv5 baseline. Qualitative validation in four challenging scenarios showed that the proposed model is robust under real field conditions, including low tonal contrast, complex backgrounds, and scale variations. This study contributes to the optimization of agricultural management, offering a viable and efficient tool for reducing losses and increasing sustainability in soybean cultivation.