Multi-label deep learning for plant leaf disease classification
The advancement of IT technology, particularly the emergence of deep learning, has resulted in significant changes in various industries, including agriculture. Agricultural produce quality and quantity are crucial guarantees for human quality of life and directly impact public welfare. Leaf diseases are a major threat to crops and directly affect crop productivity and survival. Previously, manual means were used for this process, e.g., routine inspection of crops by trained labour and experts, which were time or cost-consuming and often not timely. Fortunately, Deep Learning and Computer Vision methods can now detect and classify leaf diseases. However, the simultaneous classification of plant species and disease types remains challenging., e.g., 1. Various diseases may have similar symptoms; 2. One plant may have multiple diseases; 3. Various plant species may have the same disease. Thus, an effective classification model should accurately classify both plant species and disease types simultaneously. This study aims to address the issue of detecting and classifying plant leaf diseases in agriculture using deep learning and computer vision technology while further enhancing performance.
This study is divided into four phases. In the first phase, a literature review is conducted to identify current research gaps in plant leaf disease classification, select suitable benchmark datasets, and research common Deep Neural network models and traditional DL multi-label approaches. In the second phase, common traditional DL multi-label approaches’ performance is tested and collected as baselines for the next stages. This phase aims to investigate whether using multiple labels (plant types and disease types) simultaneously is useful for leaf disease prediction and how to incorporate these two labels’ information into deep learning models to improve performance. The third phase proposes a model that predicts both plant and disease species and explores how to enhance the performance of the multi-label deep learning approach model. The fourth phase investigates the impact of different loss balance weights and deep supervision methods to enhance the proposed model’s performance. The study’s primary objective is to develop a computer-vision-based approach that aids farmers in the early detection and classification of plant leaf diseases, leading to cost and labour savings. The study’s findings and ideas can benefit Smart Agriculture applications.
History
Sub-type
- Master's Thesis