Segmentação Semântica de Cultivos de Cana-de-Açúcar usando Deep Learning

Abstract

Sugarcane plays a strategic role in Brazil, requiring agricultural solutions to increase productivity and sustainability. This application uses semantic segmentation on aerial images of sugarcane obtained by drones, changing the precise identification of the Sugarcane class in challenging scenarios such as irregular morphology, leaf overlap, and class imbalance. Using the MMSegmentation framework, the study compared four Deep Learning architectures: FCN, PSPNet, DeepLabV3+, and SegFormer. The FCN, PSPNet, and DeepLabV3+ models acquired ResNet-50 as a backbone, and SegFormer used MiT-B0. The dataset consists of high-resolution orthophotos with a GSD of approximately 3 cm, accompanied by annotated segmentation masks. All models underwent fine-tuning with pre-trained weights under unified training conditions and were evaluated by standardized analyses. We discovered that PSPNet achieved the best overall performance in identifying the Sugarcane class, outperforming other architectures in all global and class-specific performance metrics, except for the Acc metric, where FCN and DeepLabV3+ performed slightly better. This result is attributed to the effectiveness of its Pyramid Pooling module in capturing context at multiple scales and in hierarchical feature aggregation, which strengthens the ability to discriminate between classes and the robustness against variations in images. Therefore, the results obtained can contribute to expanding the use of precision agriculture techniques in sugarcane cultivation, especially through automated monitoring, allowing for the early identification of problems that directly impact productivity.