Detection of late blight in potato leaves using drone images and Deep Learning techniques

• Autores: Cairampoma, J. A., Gamarra-Gamarra, Delia, Dionisio, F. E.
• Formato: artículo
• Fecha de Publicación: 2026
• Institución: Universidad Nacional de Trujillo
• Repositorio: Revistas - Universidad Nacional de Trujillo
• Lenguaje: español
• OAI Identifier: oai:ojs.revistas.unitru.edu.pe:article/6482
• Enlace del recurso: https://revistas.unitru.edu.pe/index.php/scientiaagrop/article/view/6482
• Nivel de acceso: acceso abierto
• Materia: Detección del tizón tardío; Dron; Red neuronal convolucional; Mask R-CNN; Folíolos de papa; Late blight detection; Unmanned aerial vehicle (UAV); Convolutional neural network; Potato leaflets

Phytophthora infestans causes one of the most devastating diseases of the potato crop, also known as late blight. Since early identification of this pathogen is crucial for the effective control of the disease, this study aimed to propose an automated methodology for the identification of its lesions in potato leaflets, using convolutional neural networks called “Mask R-CNN”. The evaluations were carried out during the rainy season, in crops conducted by farmers in the locality of Huasahuasi, in the central Andes of Peru. One hundred photographs (5472 × 3078 pixels) were taken with a Phantom 4 Pro unmanned aerial vehicle (UAV) at an altitude of 3 m in crops with a late blight incidence between 2 and 3. The images were divided into four parts and then passed thorough quality control, resulting in 200 photos (1825 × 1369 pixels). Of the total, 75% was used for model training and 25% for model validation. The models were evaluated under real conditions, using metrics such as accuracy and recall. It was determined that the Mask R-CNN neural network, based on the ResNet 101 deep neural network architecture, offers acceptable accuracy and effectiveness (73.5%) in the identification of late blight lesions at the leaflet level. This methodology constitutes a significant contribution to precision agriculture in the Andes, validating a non-invasive tool capable of overcoming the topographical limitations of the area. Its practical application would optimize the use of fungicides through targeted detection, thereby promoting more sustainable and profitable potato production systems for local farmers.