Evaluation of Convolutional Neural Network Models for Leaf-Based Endangered Tree Classification

The conservation of endemic tree species in the Philippines requires lightweight and localized technological solutions. Therefore, we benchmark three Convolutional Neural Network (CNN) architectures: MobileNetV2, EfficientNetB0, and NASNetMobile, using 1800 leaf images from five endangered Dipterocarpaceae species and an unknown class. The models were trained with transfer learning and deployed on a Raspberry Pi prototype. Validation accuracies exceeded 97.00%, with NASNetMobile reaching 99.44%. Field testing produced lower accuracies of 85.00% for NASNetMobile, 77.78% for MobileNetV2, and 66.11% for EfficientNetB0. One-way analysis of variance with Tukey’s Honestly Significant Difference showed no statistically significant differences at the 0.05 level. These findings highlight the importance of field validation and statistical testing in CNN benchmarking for embedded biodiversity monitoring.