A sample academic presentation, a full slide deck with speaker notes, prepared to Australian postgraduate standard. Preview the slides below, or download the editable PowerPoint.
Download slides (PowerPoint) Download as PDF
Slide-by-slide outline
Slide 1: Machine Learning for Early Detection of Crop Disease
- Discipline: applied machine learning and computer vision
- Task: image-based classification of leaf disease at an early stage
- Context: Australian grain and horticulture biosecurity
- Approach: a convolutional model benchmarked against a classical baseline
Speaker notes: This presentation reports a study applying deep learning to image-based detection of crop disease before symptoms become severe. The work targets the Australian context, where early detection supports biosecurity and yield protection. It compares a convolutional model against a traditional method on the same data.
Slide 2: Agenda
- Problem and motivation
- Aim and research questions
- Data and modelling method
- Results, comparison and deployment
- Recommendations and future work
Speaker notes: The deck opens with why early detection matters economically and for biosecurity. It then defines the research questions and the modelling pipeline. Results, practical deployment and next steps follow.
Slide 3: Background
- Crop disease reduces Australian yields and export quality
- Manual scouting is slow, costly and inconsistent
- Earlier detection limits spread and unnecessary chemical use
- Bodies such as GRDC and Plant Health Australia prioritise surveillance
Speaker notes: Field detection still relies heavily on manual inspection, which is labour-intensive and often detects problems late. Earlier detection reduces both yield loss and needless spraying. Automated image analysis offers a scalable alternative that aligns with national surveillance priorities.
Slide 4: Aim and Research Questions
- Aim: assess whether deep learning enables reliable early detection
- Can a convolutional network classify disease from leaf images accurately?
- Does transfer learning outperform a classical baseline?
- Is the model efficient enough for in-field devices?
Speaker notes: The aim tests both accuracy and practicality, not accuracy alone. The questions compare a deep model against a traditional method and consider whether it can run on modest hardware. Efficiency matters because deployment is on farm, not in a data centre.
Slide 5: Method
- Dataset of 24,000 labelled leaf images, six disease classes plus healthy
- Split: 70 per cent training, 15 per cent validation, 15 per cent test
- Model: EfficientNet-B0 with transfer learning; a support vector machine baseline
- Tools: Python and PyTorch, with augmentation for lighting and rotation
Speaker notes: A labelled image dataset was partitioned into training, validation and test sets. Transfer learning with EfficientNet-B0 was chosen for accuracy at low computational cost. A support vector machine on handcrafted features served as the baseline, and augmentation improved robustness to field conditions.
Slide 6: Results, Model Performance
- Test accuracy of 96.2 per cent across seven classes
- Macro F1 score of 0.95, with precision 0.96 and recall 0.94
- Most errors fell between two visually similar fungal classes
- Inference time of 45 milliseconds per image on a mobile processor
Speaker notes: The deep model reached high accuracy with balanced precision and recall. The main confusion occurred between two fungal diseases with similar early symptoms, which matches agronomic expectations. Inference speed was well within the range needed for field use.
Slide 7: Results, Model Comparison
- EfficientNet-B0: 96.2 per cent accuracy with 5.3 million parameters
- Support vector machine baseline: 81.4 per cent accuracy on the same test set
- Transfer learning cut training time substantially
- The deep model generalised better to unseen lighting conditions
Speaker notes: The deep model clearly outperformed the classical baseline, especially under variable lighting. Transfer learning reduced both training time and data requirements. The parameter count remains small enough for constrained devices.
Slide 8: Results, Explainability and Deployment
- Grad-CAM heatmaps localised lesions, supporting user trust
- The model compressed to 12 megabytes for on-device use
- Offline operation suits low-connectivity paddocks
- Confidence thresholds flag uncertain cases for human review
Speaker notes: Grad-CAM visualisations showed the model attending to actual lesion areas rather than background, which supports trust. Compression allowed offline operation, important given rural connectivity. A confidence threshold routes uncertain images to an agronomist rather than forcing an automated call.
Slide 9: Discussion
- High accuracy is promising for early, low-cost detection
- Curated images may overstate real field performance
- Class imbalance and rare diseases remain challenging
- A human-in-the-loop design mitigates automation risk
Speaker notes: The results are encouraging but were obtained on curated images, so field performance is likely lower. Rare diseases with few examples are hard to learn and need targeted data collection. A human-in-the-loop workflow keeps an expert accountable for consequential decisions.
Slide 10: Recommendations
- Validate on multi-season, in-field imagery before any rollout
- Expand data for rare and early-stage disease presentations
- Integrate detection into existing farm-management applications
- Retain agronomist review for low-confidence predictions
Speaker notes: The priority is field validation across seasons and regions to confirm robustness. Targeted data collection would address the rare-disease gap. Integration with tools growers already use, plus expert review, would support safe adoption.
Slide 11: Conclusion
- Deep learning detected crop disease with high accuracy
- It clearly outperformed a classical baseline
- The model is efficient enough for on-device field use
- Field validation and human oversight are the next steps
Speaker notes: The study shows efficient deep models can detect crop disease accurately and quickly. The gains over traditional methods are substantial. Careful field validation and human oversight are needed before operational deployment.
References
Australian Bureau of Agricultural and Resource Economics and Sciences 2022, Australian Crop Report, ABARES, Canberra.
Barbedo, JGA 2019, ‘Plant disease identification from images: challenges and prospects’, Computers and Electronics in Agriculture, vol. 165, article 104958.
Commonwealth Scientific and Industrial Research Organisation 2021, Digital Agriculture: Opportunities for Australian Farming, CSIRO Publishing, Canberra.
Grains Research and Development Corporation 2022, Crop Disease Management Update, GRDC, Canberra.
Mohanty, SP, Hughes, DP & Salathe, M 2016, ‘Using deep learning for image-based plant disease detection’, Frontiers in Plant Science, vol. 7, article 1419.
Plant Health Australia 2021, National Plant Biosecurity Status Report, PHA, Canberra.
Selvaraju, RR, Cogswell, M & Das, A 2020, ‘Grad-CAM: visual explanations from deep networks’, International Journal of Computer Vision, vol. 128, pp. 336 to 359.
Tan, M & Le, Q 2019, ‘EfficientNet: rethinking model scaling for convolutional neural networks’, Proceedings of the International Conference on Machine Learning, pp. 6105 to 6114.
Speaker notes: These sources support the modelling choices, the explainability method and the biosecurity framing used across the deck.