Plant Identification Using Machine Learning under Real-World Conditions

Abstract

This thesis aimed to research and develop a practical prototype application that can identify different types of plant life under different conditions using machine learning while enhancing the researcher’s experience in the field. The research questions investigated by the thesis were: How effective are trained machine learning models at detecting different species of plants under different conditions, how can we efficiently fine-tune models for better species recognition after the initial training and deployment, and what are any trade-offs regarding accuracy and computational efficiency when deploying the identification models on mobile devices. The dataset used to develop the created model was provided by the researchers at HAMK Lepaa.

The thesis approached the research problems using practical methods. The dataset provided from HAMK Lepaa was first combed through to separate images with good quality and bad quality, then they were sorted again by their corresponding species into 21 unique classes. This new dataset was then used for training a model using the You Only Look Once (YOLO) algorithm in a Google Colab environment. After multiple training sessions, the model is incorporated into a simple user interface created using Gradio and deployed into Hugging Face Spaces. The model and application’s performance are analysed using different confusion matrices and loss curve diagrams automatically generated during training.

The thesis was successful in the creation of a working prototype that utilizes a trained model to make predictions. The model itself had difficulty predicting classes that were either visually similar to other classes or belonged to the same family. The deployed prototype also had slower processing speeds due to the limitations of Hugging Face Spaces. Based on the analysis, it is recommended that a larger dataset of higher quality be used to avoid issues of misclassifications among different classes and overfitting during training. Alternative platforms and methods for deployment should also be considered to ensure that processing speeds are fast enough to remain practical for use.