A study on the influence of training parameters in CNN and Transformer architectures for semantic segmentation in vineyards

This study analyzes the performance of two semantic segmentation architectures, DeepLabV3+ (CNN-based) and SegFormer (Transformer-based), applied to precision viticulture, with the aim of understanding their strengths and limitations in real-world scenarios. The investigation focused on the influence of some key parameters (Atrous Rate, segmentation channels, batch size) on two datasets: San Donaci, large but low resolution, and Zeuli, smaller but high definition. The training results show how the size and resolution of the dataset significantly affect performance. SegFormer demonstrated superior efficiency with the MiT-B2 encoder, while DeepLabV3+ produced the best mIoU results with EfficientNet-B4. In testing, all models proved to be effective for the leaves and background classes, but encountered more difficulties with grapes bunches and trunks, often fragmented or poorly visible due to chromatic similarities and lighting conditions. In general, SegFormer with MiT b2 proved to be an interesting choice due to its precision in segmenting edges. In future, combining the two datasets and exploring hybrid CNN-Transformer architectures that exploit the strengths of each could improve the robustness of the segmentation.