Answers are grounded in this paper and cite their sources.
Most image-based plant disease studies train deep networks on large laboratory datasets, yet many practical projects have only a few hundred field photographs and no GPU. This paper evaluates eight standard machine learning classifiers on a small field image dataset of 1,532 leaf photographs labelled Healthy, Powdery (powdery mildew) and Rust, using 129 hand-crafted colour, texture and shape features. Classifiers were compared by nested, repeated stratified cross-validation on the 1,382 training and validation images, and the best model was tested once on the untouched official 150-image test set. Gradient boosting achieved the highest cross-validated accuracy (92.04% ± 1.41%) and significantly outperformed random forest (89.87%; corrected resampled t-test, p = 0.021). On the official test set it reached 89.33% accuracy (95% CI 83.4–93.3%), a macro F1-score of 89.37% and Cohen's κ = 0.84, within four points of a published DenseNet121 result on the same data. Colour and texture features were the most informative; adding shape features reduced k-nearest-neighbour accuracy from 87.37% to 66.79%. Random forest was the most data-efficient model, reaching 68.4% test accuracy with only five training images per class. The commonest error was a diseased leaf predicted as healthy (10 of 100 diseased test leaves). A bias control using only eight border pixels per image reached 58.1% in cross-validation but 38.7% on the test set (chance: 33.3%), revealing that capture conditions shared within the training split inflate cross-validated estimates. Classical machine learning is therefore a strong, low-cost baseline for field disease recognition, provided performance is reported on an independent test set.
No takes yet. Share an insight, caveat, or question.
Battu et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: