Developing an Original Computational Framework for Automated Plant Species Identification Using Deep Convolutional Neural Networks Trained on Nigerian Herbarium Specimen Images

Automated plant identification using machine learning offers transformative potential for accelerating botanical survey capacity in Nigeria, where expert taxonomic capacity is critically limited relative to the richness and complexity of the flora. This study develops an original computational framework for automated vascular plant species identification using deep convolutional neural networks trained and validated on digitised Nigerian herbarium specimen images. An original dataset of 47,840 digitised specimen images representing 620 Nigerian plant species was compiled from five university herbaria (University of Nigeria Nsukka, Ahmadu Bello University, University of Ibadan, University of Maiduguri, and Bayero University Kano) through a structured collaborative digitisation programme. Images were pre-processed, augmented, and split into training (70%), validation (15%), and test (15%) sets. Four convolutional neural network architectures were evaluated: ResNet-152, EfficientNet-B7, Vision Transformer (ViT-L), and an original hybrid architecture (BotaNet-NG) developed in this study that incorporates morphological feature embedding layers derived from botanical descriptor ontologies. Model performance was assessed by top-1 and top-5 accuracy, precision, recall, and F1-score. BotaNet-NG achieved the highest top-1 accuracy (87.4%) and top-5 accuracy (96.2%), outperforming all baseline architectures, attributed to the morphological feature embedding that encodes botanical expert knowledge into the network architecture. Performance was highest for families with distinctive morphology (Orchidaceae: 94.1%, Fabaceae: 91.3%) and lowest for morphologically cryptic families (Rubiaceae: 78.4%). The study provides an original and tested artificial intelligence framework for Nigerian plant identification and makes a methodological contribution to digital taxonomy by embedding morphological ontologies into neural network design.

Keywords: deep learning, plant identification, convolutional neural network, herbarium specimens, digital taxonomy