Logo image
Deep Neural Network for Data Sparsity Scenarios in Plant Disease Detection
Doctoral Thesis   Open access

Deep Neural Network for Data Sparsity Scenarios in Plant Disease Detection

Masoud Rezaei
Murdoch University
Doctor of Philosophy (PhD), Murdoch University
2025
DOI:
https://doi.org/10.60867/00000046
pdf
Whole Thesis13.73 MBDownloadView
Open Access

Abstract

Barley--Diseases and pests Barley--Yields Plant diseases—Diagnosis
Plant disease causes severe yield loss and affects production quality. Traditional methods are ineffective in managing plant diseases on large-scale farms. Deep learning (DL) has advanced automated disease detection methods, but they face data-related challenges. In a realistic approach, this study addressed some challenges in plant disease detection, including limited data availability, complexity of field-collected samples, pixel-level annotation, and severity estimation. The contributions are as follows: First, this study reviewed recent developments in AI-based disease detection techniques and highlighted the challenges of developing an efficient disease recognition method. Second, using a consumer-grade RGB camera, we collected a barley disease dataset from barley test-bed trials across multiple paddocks. The dataset captured samples of three barley diseases, such as net form net blotch, spot form net blotch, and scald. Given the small size of the dataset, MobileNet outperformed other well-known deep learning models, such as ResNet, Xception, and Inception, achieving high accuracy in the defined barley disease classification task. Moreover, we introduced a novel FSL method (PMF+FA) to address the limited data challenge. It incorporated a feature attention (FA) module into the pipeline based on pre-training, meta-training, and fine-tuning (PMF) steps. Vision and Swin transformers demonstrated high performance with only five training samples per class, achieving over 90% accuracy given complex field samples. The thesis proposed a practical approach for automatic pixel-level annotation to label disease lesions. Finally, this thesis presented a novel framework for disease severity estimation by extracting morphological, spectral, and texture features. ABCD (asymmetry, border, colour, diameter) features are commonly used for human skin cancer classification; Considering the similarities between tasks, this study introduced these features to plant disease severity classification.

Details

Metrics

14 File views/ downloads
18 Record Views
Logo image