Doctoral Thesis
Advances in deep learning techniques with uncertainty quantification for image-based detection of agricultural pests
Doctor of Philosophy (PhD), Murdoch University
2024
Abstract
Insects and pests are major threats to agricultural production and quality. Both visual inspection for pests and conventional computational models used to detect tiny objects such as insects are slow, expensive and time-consuming. The advent of modern Artificial Intelligence (AI)-based approaches present new ways to detect better and address these threats. Increasing detection, accuracy, and computational speed, as well as quantifying the accuracy of uncertainty in AI prediction, has become a significant challenge for AI techniques. This thesis presents novel deep-learning methodologies for image-based pest detection with uncertainty quantification.
The thesis begins by surveying the current status of detecting tiny objects and presenting four major contributions.
The first formulates pest detection as tiny object detection in images. It integrates an adaptive feature fusion strategy into an existing object detection framework to cater for variable-sized small objects, and this significantly improves the accuracy of pest detection.
The second contribution focuses on the communal behaviour of aphid pests, correlating their tendency to form colonies with increasing severity of infestation. A new approach, involving colony detection using deep learning, is proposed. The AphColDat dataset, the first for aphid colonies, was repurposed from an existing pest dataset. The methodology developed provides an innovative annotation strategy of bounding box merging and sets the stage for evaluating the efficacy of deep learning models in colony detection.
The third contribution emphasises the importance of uncertainty and its quantification for enhancing the precision and reliability of a deep-learning model for pest detection. It introduces a novel Bayesian multi-task learning model that detects pests, estimates their sizes, and simultaneously provides uncertainty metrics.
The final contribution presents an innovative approach that integrates gradient-based uncertainty and an attention mechanism into a feature pyramid network architecture. This strategy is specially designed to tackle the challenges in detecting small objects against complex backgrounds, enhancing the model’s focus on areas of images containing small objects and significantly improving detection accuracy across diverse datasets.
Overall, the proposed techniques developed in this thesis can better detect agricultural pests and increase sustainable agriculture and global food security. This endeavour pushes the frontiers of AI in agriculture and offers scalable solutions that could be adapted to other fields requiring precise object detection under varying conditions.
Details
- Title
- Advances in deep learning techniques with uncertainty quantification for image-based detection of agricultural pests
- Authors/Creators
- Abderraouf Amrani
- Contributors
- Ferdous Sohel (Supervisor) - Murdoch University, Centre for Crop and Food InnovationMichael G. K. Jones (Supervisor) - Murdoch University, Centre for Crop and Food InnovationDean Diepeveen (Supervisor)David Murray (Supervisor)
- Awarding Institution
- Murdoch University; Doctor of Philosophy (PhD)
- Identifiers
- 991005693761307891
- Murdoch Affiliation
- Centre for Crop and Food Innovation; School of Information Technology
- Resource Type
- Doctoral Thesis
UN Sustainable Development Goals (SDGs)
This output has contributed to the advancement of the following goals:
Metrics
105 File views/ downloads
155 Record Views