In roses, the gray mold caused by the fungus Botrytis cinerea mainly produces rot of flower buds, which makes it a non-commercial product, causing important losses. Therefore, early detection of this disease in Ecuador is a priority for rose-producing farmers to sustainably increase productivity. However, only a few studies exist on the use of machine learning (ML) and deep learning (DL) techniques in roses so, this paper aimed at early detection of the Botrytis cinerea incidence in the floriculture sector in Ecuador, through a cost-effective, image processing prototype using convolutional neural networks (CNN), followed by in parallel molecular validation of the algorithm accuracy based on PCR.

With this strategy, the predictive capacity of the method was revalidated, and above all, it was ensured that petal spots detected by the CNN model really correspond to the disease-causing agent and not to another abiotic physiological alteration. To the best of researchers knowledge, there are no previous reports in which the graphical method based on CNN model has been validated in parallel with a molecular technique as an alternative way to demonstrate method accuracy and precision.

In this study, a binary classifier was developed using the pre-trained ResNet50 arquitecture and the EcuBotrytisCinerea dataset, called EcuRossCNN. Additionally, data augmentation techniques were employed to enhance the model's performance and generalization capabilities. Results The model demonstrated outstanding performance across key classification metrics, achieving an accuracy of 98% ± 1.0 and AUC of 98.76% ± 1.30. These results highlight its robustness and effectiveness in the classification task.

Finally, in parallel assayed plants with both EcuRossCNN and PCR, only 5 of 100 randomly assayed rose plants grown under greenhouse conditions were detected as false positives due to signals in petals detected by the EcuRossCNN quite similar to disease symptoms, but with no presence of the causal agent as judged by a specific PCR detection assay. Undoubtedly, this cost-effective method, together with human monitoring, will help in the accurate automated early detection of this terrible disease, even for non-special monitors.

Hernández, Diana & Flores-Calero, Marco & Landázuri, Pablo & Noceda, Carlos & Izquierdo Romero, Andrés & Rivera, Leonor & Trujillo, Luis. (2026). PCR-validated accuracy of a CNN algorithm for detection of gray mold caused by Botrytis cinerea with potential application in Ecuadorian rose bushes. Frontiers in Agronomy. 8. 10.3389/fagro.2026.1850785.
 

Source: www.floraldaily.com