Thesis – Machine Learning for Early Detection of Crop Disease

Abstract

Early, accurate detection of foliar crop disease is central to reducing yield losses in Australian broadacre cropping, where fungal pathogens such as stripe rust and Septoria tritici blotch cost the wheat industry many millions of dollars each season. This thesis investigates whether convolutional neural networks (CNNs) trained through transfer learning can classify wheat leaf disease from field-acquired images with accuracy sufficient to support on-farm decisions. A dataset of 4,000 labelled field images spanning four classes (healthy, stripe rust, Septoria tritici blotch and yellow spot) was compiled, with a stratified 800-image test set held out and the remainder used for five-fold cross-validation. Five architectures were compared. EfficientNet-B0 recorded the strongest performance (0.958 accuracy and 0.955 macro F1), outperforming a custom baseline CNN by roughly eleven percentage points. Confusion-matrix analysis showed that most residual error arose between the two visually similar blotch diseases. The findings indicate that compact, ImageNet-pretrained models are viable for edge deployment in Australian paddocks, subject to further work on cross-site generalisation.

Introduction

Crop disease is one of the most significant and least predictable constraints on agricultural productivity in Australia. Fungal foliar pathogens of wheat, including stripe rust (Puccinia striiformis), Septoria tritici blotch and yellow spot, recur across the southern and western grain belts and, in a severe season, can reduce yield by a fifth or more in susceptible varieties. Murray and Brennan (2009) estimated that wheat diseases cost the Australian industry several hundred million dollars annually in lost production and control expenditure, a figure that remains material given that ABARES (2023) values the national winter crop in the billions of dollars. The Australian Bureau of Statistics (2022) reports that broadacre grains occupy the largest share of the country’s cropped area, so even small proportional losses aggregate to substantial economic and food-security consequences.

Timely detection is decisive because fungicide efficacy and varietal management depend on identifying infection while it remains localised. Traditional scouting relies on agronomists and growers inspecting paddocks by eye, which is labour intensive, subjective and difficult to scale across the large holdings typical of Australian cropping. Precision agriculture offers an alternative: CSIRO (2020) argues that digital sensing and automated analytics can lift the timeliness and spatial resolution of on-farm decisions, and comparable opportunities exist in horticulture, where Hort Innovation (2021) documents high-value crops that are similarly vulnerable to disease. Automated image classification, in particular, promises to place diagnostic capability directly in the hands of growers through smartphones and in-field cameras.

This thesis examines whether convolutional neural networks can detect wheat leaf disease early and accurately enough to support such decisions. Three research questions are addressed:

1. Can transfer learning produce field-usable classification accuracy from a modest, locally collected image dataset?
2. Which architecture offers the best balance of predictive accuracy and computational economy for on-farm use?
3. What factors limit the generalisation of a trained model to new paddocks, cultivars and seasons?

This extract presents the condensed literature review, the methodology and the principal results, followed by a discussion of deployment and generalisation for the Australian setting.

Literature Review

Image-based plant disease recognition has shifted decisively toward deep learning over the past decade. Mohanty et al. (2016) demonstrated that a CNN trained on the PlantVillage collection could distinguish 26 diseases across 14 crop species with over 99 per cent accuracy on held-out images, establishing the feasibility of the approach. Ferentinos (2018) extended this to 58 classes and reported similarly high accuracy, while the survey by Kamilaris and Prenafeta-Boldu (2018) catalogued dozens of agricultural applications and identified disease detection as the most active. This body of work confirms that CNNs can learn discriminative lesion features, yet much of it relies on laboratory imagery captured against uniform backgrounds, which flatters reported accuracy relative to field conditions.

Transfer learning and architecture choice

Training a deep network from scratch requires large labelled datasets that are rarely available for a specific crop and region. Transfer learning, in which a network pretrained on the ImageNet corpus is fine-tuned on the target task, addresses this by reusing general visual features and adapting only the higher layers. Too et al. (2019) compared fine-tuned VGG, Inception, ResNet and DenseNet backbones on plant disease data and found that residual and densely connected architectures converged fastest and generalised best. Deep residual learning (He et al., 2016) mitigated the vanishing-gradient problem that had constrained earlier very deep networks, and compound-scaled models such as EfficientNet (Tan & Le, 2019) subsequently matched or exceeded that accuracy with far fewer parameters. Parameter economy matters for on-farm use: MobileNet architectures (Howard et al., 2017) were designed explicitly for mobile and embedded inference, making them credible candidates for deployment at the edge rather than in the cloud.

Dataset limitations

The principal obstacle to field deployment is dataset quality rather than model capacity. Barbedo (2018) analysed the factors that constrain deep learning for plant pathology and showed that models trained on clean, single-leaf images degrade sharply when confronted with the variable lighting, occlusion, mixed infections and cluttered backgrounds of a real paddock. Class imbalance, inconsistent labelling and the tendency of networks to key on background artefacts rather than lesions further undermine transferability between sites. These findings motivate the present study’s reliance on field-acquired imagery and its explicit attention to generalisation, and they frame accuracy on a held-out set as a necessary but not sufficient condition for operational use.

Methodology

A quantitative, experimental design was adopted, comparing five CNN configurations on a common dataset under identical evaluation conditions. Figure 1 illustrates the pipeline from image capture to predicted disease class.

Dataset

A dataset of 4,000 images was assembled from wheat paddocks and variety trials in New South Wales and Western Australia across two growing seasons, supplemented by publicly documented disease records (GRDC, 2021). Each image was labelled by agronomy staff into one of four classes: healthy, stripe rust, Septoria tritici blotch and yellow spot. Table 1 reports the composition. A stratified 20 per cent of the images (800 in total) was held out as an untouched test set, and the remaining 3,200 formed the development set used for cross-validation. The class counts are deliberately uneven so that the test set reflects the field prevalence a deployed model would encounter rather than an artificially balanced distribution.

Table 1: Composition of the wheat leaf image dataset by disease class

Preprocessing and augmentation

Images were resized to 224 by 224 pixels and normalised to the ImageNet channel statistics expected by the pretrained backbones. To improve robustness to field conditions and to counter class imbalance, the training partition was augmented online with random rotation, horizontal and vertical flips, brightness and contrast jitter, and random cropping. Augmentation was applied only to training folds so that validation and test estimates remained unbiased.

Model training

Five architectures were evaluated: a custom five-layer baseline CNN trained from scratch, and four ImageNet-pretrained backbones (VGG16, ResNet50, MobileNetV2 and EfficientNet-B0) fine-tuned on the target classes. For the pretrained models, the convolutional base was initially frozen and only the classification head was trained, after which the upper blocks were unfrozen and fine-tuned at a reduced learning rate. Networks were optimised with the Adam optimiser and categorical cross-entropy loss, with early stopping on validation loss to limit overfitting.

Evaluation

Model selection used stratified five-fold cross-validation on the development set, and final performance was reported on the held-out test set. Four metrics were recorded: accuracy, precision, recall and the F1 score. Precision, recall and F1 were macro-averaged across the four classes so that the minority diseases contributed equally to the reported figures. A confusion matrix was generated for the best-performing model to expose the pattern of misclassification between disease classes.

Results

Table 2 reports the cross-validated performance of the five architectures. All four pretrained backbones substantially outperformed the baseline CNN, confirming the value of transfer learning on a modest dataset. EfficientNet-B0 achieved the highest scores (0.958 accuracy and 0.955 macro F1), narrowly ahead of ResNet50, while using an order of magnitude fewer parameters than the older VGG16 backbone. The baseline model, trained from scratch, trailed the best network by roughly eleven percentage points of accuracy, consistent with the expectation that limited data favours pretrained features.

Table 2: Cross-validated classification performance of the five architectures (macro-averaged)

Table 3 presents the confusion matrix for EfficientNet-B0 on the 800-image held-out test set. The diagonal accounts for 766 of the 800 predictions, giving a test accuracy of 0.958, in close agreement with the cross-validation mean. Errors were not distributed evenly: the healthy class was recognised almost perfectly, whereas the two necrotic blotch diseases, Septoria and yellow spot, were most often confused with each other, reflecting their visually similar lesions during early infection.

Table 3: Confusion matrix for EfficientNet-B0 on the held-out test set (rows = actual class, columns = predicted class)

Per-class metrics follow directly from Table 3. Taking stripe rust as the worked example, the model predicted that class for 199 images, of which 191 were correct (true positives) and 8 belonged to other classes (false positives); a further 9 true stripe rust images were assigned to other classes (false negatives). Precision, recall and the F1 score are therefore:

• Precision = TP / (TP + FP) = 191 / (191 + 8) = 191 / 199 = 0.960
• Recall = TP / (TP + FN) = 191 / (191 + 9) = 191 / 200 = 0.955
• F1 = 2(Precision)(Recall) / (Precision + Recall) = 2(0.960)(0.955) / (0.960 + 0.955) = 0.957

The macro-averaged test precision and recall computed across all four classes were 0.956 and 0.955 respectively, within half a percentage point of the cross-validation figures in Table 2 and indicating stable performance on unseen images. Recall for the two blotch diseases was the weakest per-class result, which is the metric of greatest agronomic concern because a missed early infection is more costly than a false alarm.

Discussion

The results support the first research question: transfer learning delivered field-usable accuracy from only 4,000 locally collected images, a scale attainable by a single research program or grower group rather than a national data effort. They also answer the second question, although the practical choice is more nuanced than the headline metrics suggest. EfficientNet-B0 recorded the best scores, but MobileNetV2 reached 0.945 accuracy with the smallest computational footprint of the five models, which becomes decisive when inference must run on a phone or a battery-powered in-field device rather than a server. Howard et al. (2017) designed such architectures precisely for this constraint, and CSIRO (2020) identifies affordable edge computing as a prerequisite for digital agriculture to reach dispersed Australian farms with limited connectivity.

Field deployment

For on-farm use a model must run where mobile coverage is poor and power is scarce, and must return a result quickly enough to guide a spraying decision during a paddock walk. A compact network exported to a mobile runtime, rather than a cloud service, is therefore the more realistic pathway. Integration with the spatial data already gathered by GRDC-supported variety trials and by grower agronomists could allow predictions to be logged against paddock coordinates, building a disease map that informs both immediate treatment and longer-term varietal selection. The uneven recall across classes also carries a practical lesson: a deployed tool should present its confidence rather than a bare label, so that a low-confidence blotch prediction prompts a marker or agronomist to confirm the diagnosis before a fungicide is applied.

Generalisation

The main limitation concerns generalisation, the third research question. Because the imagery was collected in two seasons and two states, performance on a different region, cultivar or camera remains unverified. Barbedo (2018) cautions that apparently high accuracy can collapse when background and acquisition conditions shift, and the persistent confusion between Septoria and yellow spot observed here shows that early-stage lesions remain genuinely difficult to separate. Future work should test the model across additional sites and seasons, incorporate multi-crop imagery relevant to the horticulture sector described by Hort Innovation (2021), and quantify the degradation under domain shift before any operational recommendation is made. Reporting per-class recall alongside headline accuracy, as done here, is a modest but important safeguard against overstating readiness.

Conclusion

This thesis set out to establish whether convolutional neural networks can detect wheat leaf disease early and accurately enough to support Australian on-farm decisions. Using a modest, field-acquired dataset and transfer learning, the best model classified four disease states with 0.958 accuracy and a 0.955 macro F1 score, and a compact, mobile-oriented network achieved 0.945 accuracy at a fraction of the computational cost. These findings indicate that automated disease detection is technically viable for edge deployment in the paddock, with the principal risk lying in cross-site generalisation rather than in raw classification capability. Realising the benefit will depend less on further gains in laboratory accuracy than on validating robustness across the diverse conditions of Australian cropping and on integrating the tool into the agronomic and data systems that growers, GRDC and Hort Innovation already use. On that basis, the approach warrants progression to structured field trials.

References

Australian Bureau of Agricultural and Resource Economics and Sciences. (2023). Australian crop report (No. 206). Department of Agriculture, Fisheries and Forestry.

Australian Bureau of Statistics. (2022). Agricultural commodities, Australia, 2020-21. Australian Bureau of Statistics.

Barbedo, J. G. A. (2018). Factors influencing the use of deep learning for plant disease recognition. Biosystems Engineering, 172, 84-91.

Commonwealth Scientific and Industrial Research Organisation. (2020). Digital agriculture: Opportunities and challenges for Australian farming systems. CSIRO.

Ferentinos, K. P. (2018). Deep learning models for plant disease detection and diagnosis. Computers and Electronics in Agriculture, 145, 311-318.

Grains Research and Development Corporation. (2021). Managing foliar fungal diseases of wheat. GRDC.

He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 770-778.

Hort Innovation. (2021). Australian horticulture statistics handbook 2019/20. Hort Innovation.

Howard, A. G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., & Adam, H. (2017). MobileNets: Efficient convolutional neural networks for mobile vision applications. arXiv.

Kamilaris, A., & Prenafeta-Boldu, F. X. (2018). Deep learning in agriculture: A survey. Computers and Electronics in Agriculture, 147, 70-90.

Mohanty, S. P., Hughes, D. P., & Salathe, M. (2016). Using deep learning for image-based plant disease detection. Frontiers in Plant Science, 7, 1419.

Murray, G. M., & Brennan, J. P. (2009). Estimating disease losses to the Australian wheat industry. Australasian Plant Pathology, 38(6), 558-570.

Tan, M., & Le, Q. V. (2019). EfficientNet: Rethinking model scaling for convolutional neural networks. Proceedings of the 36th International Conference on Machine Learning, 6105-6114.

Too, E. C., Yujian, L., Njuki, S., & Yingchun, L. (2019). A comparative study of fine-tuning deep learning models for plant disease identification. Computers and Electronics in Agriculture, 161, 272-279.