Deep Learning-Based Classification of Andean Medicinal Plants Using Morphological Feature Extraction

Solis I.S.; Buleje C.Y.C. · Lecture Notes in Networks and Systems (2026) · DOI: 10.1007/978-3-032-26217-2_11

Resumen

Accurate identification of medicinal plants is crucial for safeguarding ethnobotanical heritage and ensuring the correct use of traditional remedies. Most studies rely on laboratory datasets, which fail to capture natural variability in illumination, background, and leaf morphology. This study proposes a multimodal deep learning framework for classifying ten Andean medicinal plant species based on leaf images collected under uncontrolled field conditions. The architecture integrates a visual branch based on EfficientNet-B0 for feature extraction and a tabular branch for six numerical descriptors describing leaf texture and geometry. Textural features were computed using the Gray-Level Co-occurrence Matrix (GLCM), while geometric descriptors were derived from segmented leaf masks. A late-fusion strategy combines both feature spaces before the final multilayer perceptron classification stage. Experimental results show that the hybrid model achieves an accuracy of 91% and a Macro-F1 score of 0.89, outperforming image-only (86%) and tabular-only (74%) models. Morphological descriptors contribute to reducing inter-class confusion, especially among visually similar species. The model demonstrates high robustness and generalization in non-controlled conditions. The results support the effectiveness of multimodal learning for improving plant species recognition and establishing automated field-based identification systems. In addition, they demonstrate the practical potential of the proposed framework for biodiversity monitoring and digital ethnobotany applications. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.

Instituciones: Universidad Nacional José María Arguedas