Abstract
Accurate plant species identification is essential for biodiversity monitoring, ecological analysis and automated environmental assessment. However, conventional deep learning models often struggle to distinguish visually similar plant species due to subtle morphological variations and high inter-class similarity. To address this limitation, this study presents a hierarchical deep learning architecture integrating ResNet50 feature extraction, k-means clustering and cluster-specific Convolutional Neural Network (CNN) classifiers for fine-grained plant species classification. ResNet50 is employed to extract discriminative morphological representations, while k-means clustering organizes visually related species into compact feature groups before classification, thereby reducing inter-class confusion and enabling specialized CNN learning within each cluster. The proposed methodology was evaluated using the PlantCLEF 2015 dataset containing 113,205 images representing 1,000 plant species, including trees, herbs and ferns from Western Europe. Experimental evaluation achieved 0.9800 classification accuracy with 0.0615 loss, outperforming conventional architectures including ResNet50, InceptionV3, Xception, EfficientNet and MobileNet. In addition, the proposed architecture maintained efficient computational performance with an average inference latency of 18 ms per image. These findings demonstrate the suitability of the architecture for automated biodiversity assessment, ecological monitoring and large-scale botanical image analysis. Future work will focus on lightweight attention-enhanced architectures and cross-dataset generalization for deployment in resource-constrained environments.
Keywords: Convolutional Neural Networks, Deep Learning Methodology, K-means Clustering, Plant Species Classification, ResNet50.