Medicinal plant discrimination in Paramos and cloud forests of northern Peru: A study of deep learning techniques and a novel approach for data augmentation

Castro W.; Juarez M.; Gálvez-Paucar D.; Tirado V.; Gonzáles J.; Torres-Guevara F. · Remote Sensing Applications: Society and Environment (2026) · DOI: 10.1016/j.rsase.2026.102204

Resumen

Introduction: Medicinal plants in remote mountain ecosystems are threatened, requiring accurate identification for conservation and sustainable management. However, traditional methods require specialized expertise, and ground surveys in Peru’s Paramos and cloud forests are logistically challenging. This study develops and validates a methodology for discriminating six medicinal plant species using RGB images acquired from UAV and deep learning. Methodology: Individual plants from six species of medicinal plants were identified and georeferenced; from these, aerial RGB images were acquired at 3 to 5 m altitude. the rectangular ROIs of each image were extracted using the Segment Anything model (SAM), which constituted our image database. Then the image database was augmented through two methods (a) Geometric transformation and (b) a proposal called “Cartesian subsampling” Finally, five CNN architectures for classifying tasks were trained, validated, and their results statistically compared. Results: Geometric transformation yielded higher overall accuracy, with MobileNetV2 achieving 97.86% and ResNet-50 97.66%, while models trained from scratch performed lower (AlexNet-Deep: 90.14%, PhytoNet: 81.11%). Cartesian subsampling showed advantages when geometric invariance was critical for transfer-learning models. Confusion persisted between species of Myrcianthes that were morphologically similar, with F1 scores ranging from 0.76 to 0.97. Conclusion: classification precisions from 81% to 97% were achieved, with geometric transformation outperforming Cartesian methods in all models. ResNet-50 offers the best balance of accuracy, computational efficiency, and training stability for practical conservation applications in remote mountain ecosystems.

Instituciones: Universidad Nacional de Frontera