Automated Classification and Explainability of Cedar (Cedrela montana) and Cinchona (Cinchona pubescens) Using Deep Learning and Grad-CAM: A Case Study in the Amazon Region of Northern Peru

Accurate identification of timber species is essential for sustainable forest management and the prevention of illegal logging, particularly in regions such as the Peruvian Amazon. This study evaluated explainable deep learning models for the classification of cedar (Cedrela montana) and cinchona (Cinchona pubescens) using macro-images of transverse wood sections collected in the Amazonas region of northern Peru. A dataset of wood images was expanded through a sliding-window patch extraction strategy and used to train and compare three architectures: a custom Convolutional Neural Network (CNN), MobileNetV2, and EfficientNetB0. Model performance was assessed using accuracy, precision, recall, and F1-score, while Gradient-weighted Class Activation Mapping (Grad-CAM) was applied to interpret the visual regions influencing predictions. All models achieved high classification performance. EfficientNetB0 obtained the best results, reaching 100% accuracy, precision, recall, and F1-score, while MobileNetV2 and the custom CNN achieved near-perfect performance. Training and validation curves indicated stable convergence without significant overfitting. Grad-CAM analysis showed that the custom CNN generated more localized and interpretable activation regions, whereas MobileNetV2 and EfficientNetB0 focused on broader textural patterns, revealing different feature-learning strategies. These findings demonstrate the potential of explainable deep learning for automated wood identification and support applications in forest monitoring, biodiversity conservation, and illegal logging control.