Development of AI-based Image Recognition for Plant Species Identification

Thesis Abstract

The accurate and efficient identification of plant species remains a critical challenge for botanists, environmentalists, and agricultural professionals, particularly given the increasing demand for biodiversity monitoring, conservation efforts, and sustainable resource management. Traditional identification methods, often reliant on expert knowledge and manual morphological analysis, are labor-intensive, time-consuming, and prone to human error, thereby limiting their applicability in large-scale ecological studies and rapid field assessments. This study aims to develop a robust, AI-based image recognition system capable of accurately classifying diverse plant species through automated analysis of visual features. The specific objectives are to construct a comprehensive image database of representative plant species, design and train deep learning models—specifically convolutional neural networks (CNNs)—for species classification, evaluate the performance of different CNN architectures, and propose an integrated framework for practical deployment in field conditions. The research adopted a quantitative, experimental design, centered on the collection and analysis of plant images captured from botanical gardens, herbariums, and field sites within a designated ecological region. The population comprised publicly available and newly captured high-resolution images of fifty plant species, with a total of 10,000 images. A stratified sampling technique was employed to ensure balanced representation across species, resulting in a sample of 8,000 images allocated for training, validation, and testing purposes. Data collection instruments included digital cameras with standardized settings to ensure image quality, along with metadata annotations documenting species labels and environmental context. The images were preprocessed to normalize size, background, and lighting conditions to enhance model training. To ensure validity and reliability, domain experts verified species labels, and the dataset underwent iterative validation during model development. Multiple CNN architectures—such as ResNet50, InceptionV3, and DenseNet121—were trained using transfer learning techniques and optimized through hyperparameter tuning. Model performance was evaluated based on accuracy, precision, recall, F1-score, and the area under the receiver operating characteristic (ROC) curve. Analytical techniques included confusion matrix analysis, receiver operating characteristic analysis, and comparative performance assessment via ANOVA testing to determine statistically significant differences among models. The anticipated results are that the CNN-based models will achieve classification accuracies exceeding 90% across the test dataset, with DenseNet121 predicted to outperform other architectures due to its feature propagation efficiency. The system is expected to demonstrate high robustness against variations in image quality and environmental conditions, supporting its deployment in real-world scenarios. The findings will contribute to the existing body of knowledge by empirically validating the effectiveness of deep learning techniques for botanical classification tasks and by providing a scalable, automated methodology for plant identification that surpasses traditional approaches in speed and accuracy. The main conclusion asserts that AI-driven image recognition systems, particularly convolutional neural networks, provide a promising tool for rapid, accurate, and scalable plant species identification, with significant implications for biodiversity conservation, ecological research, and precision agriculture. The study recommends ongoing development of the model to incorporate multispectral and hyperspectral imaging modalities, integration with mobile applications for field use, and expansion of the dataset to include rare and cryptic species. Future research should explore hybrid approaches combining spectral data with morphological features and investigate the potential for real-time identification in diverse environmental conditions, thereby enhancing the system’s applicability and resilience across different ecological contexts.