Phenological prediction system for crops using RGB images

The prediction of phenological stages is key in precision agriculture, given that current sampling is still carried out manually and in the field, which remains costly and not very scalable. Therefore, this thesis studies a computer vision pipeline to estimate phenological stages according to the BBCH scale from images of hazelnut (Corylus avellana) and artichoke (Cynara cardunculus) in the industrial context of 3A s.r.l. Multi-source datasets (3A, GBIF, Roboflow) were curated with license traceability and labeling assisted by agronomists; lightweight CNNs (MobileNetV2/EfficientNetB0, 224×224) were trained with stratified validation, and UMAP/K-Means were used to diagnose label quality and separability. The results obtained for some models show that for artichoke cultivation (with a division into three classes), test accuracy varies between 0.69 and 0.75, Macro-F1 is 0.66 with fine-tuning and 0.72 without fine-tuning. Based on these results, it was concluded that errors are concentrated between adjacent BBCH stages. For hazelnut cultivation, the models were divided into two phases, reproductive and vegetative. For the first types of phases (with a division into four classes), an accuracy of 0.90 and a Macro-F1 of 0.85 were achieved. while for the vegetative phases (with a division into 8 classes), an accuracy of 0.87 and a Macro-F1 of 0.81 were achieved. Given that the hazelnut data comes mainly from a single site, there are still risks of overfitting and domain shift. More robust architectures are being evaluated, ablation studies are being conducted, and training is being scaled to the complete combined dataset (3A + open sources) with validation by groups.