DEEP LEARNING FOR PLANT SPECIES CLASSIFICATION: A COMPREHENSIVE CNN-BASED APPROACH WITH FEATURE VISUALIZATION AND PERFORMANCE BENCHMARKING

Abstract

Plant species classification is a critical problem in precision agriculture, in the conservation of biodiversity and ecological monitoring. In this paper, a thorough deep learning framework is proposed using Convolutional Neural Networks (CNNs) for high accuracy plant species classification using the V2 Plant Seedlings Dataset of 12 plant species. Multiple CNN architectures were systematically evaluated on a variety of standard metrics such as accuracy, precision, recall, and F1-score across the benchmark datasets.Various CNN architectures such as custom CNN, ResNet-50, DenseNet-121, EfficientNet-B3 and hybrid CNN-Transformer models were benchmarked using various metrics like accuracy, precision, recall, and F1-score across the benchmark datasets. The proposed pipeline includes three key steps: preprocessing the data (standardized resizing, normalization, and background segmentation), data augmentation, and visualization of the features using UMAP. The best model (CNN-Transformer) outperforms the current baselines by 2–5% with 98.9% classification accuracy. The separation between classes is robust as confirmed by feature analysis using UMAP projections, where the overlapping clusters are morphologically similar species. The model's discriminative power is additionally confirmed by the evaluation carried out with confusion matrices and precision-recall curves. This work sets a reproducible standard for plant classification under controlled conditions and lays the groundwork for implementation in the actual agricultural system.