Segmentação semântica de ervas daninhas em plantações de cana de açúcar

Abstract

This study presents a systematic comparison of five semantic segmentation architectures U-Net, DeepLabV3+, PSPNet, Swin Transformer, and SegFormer - applied to weed segmentation in sugarcane fields using UAV orthoimages from Mato Grosso do Sul, Brazil. A dataset comprising 30,076 georeferenced patches of 256x256 pixels was created, focusing on the segmentation of castor bean (Ricinus communis). Among the evaluated models, SegFormer achieved the best performance, reaching 82.11% mIoU, 89.77% mAcc, and 99.38% aAcc on the test set. The analysis identified severe class imbalance as a major challenge, with background pixels dominating the dataset. The results indicate that architectures incorporating multi-scale context aggregation mechanisms are better suited to this agricultural domain. The results contribute to advancing precision agriculture in Brazil, enabling the optimization of agrochemical usage through intelligent management systems.