Answers are grounded in this paper and cite their sources.
Plant disease detection systems are accurate on laboratory datasets, but struggle to maintain high accuracy in real field settings because of the variability in the environment, background noise, changes in lighting, and the diversity of plant species in terms of morphology. The similar disease pattern in the leaves are also challenging in the CNN-based models due to Global Average Pooling (GAP), which discards the spatial information that is specific to the lesion for the classification. The solution to this is inter-subject training which leads to population based tensor of the features. The tensor-based methods are based on frozen deep learning features and fail to take advantage of population-level disease distribution. The gap between the lab and the field is a continuous one and hinders the application in precision agriculture. A population-aware multilinear learning framework, PALM-Net, is proposed that consists of three components which groups features into a 4th-order population tensor, and supports spectral decorrelation and class-sensitive Discriminant Projection (DP) simultaneously. The experimental evaluation is carried out on the PlantVillage (54,305 images, 38 classes) dataset, Cassava Leaf Disease (26,220 images, 5 classes) dataset, and Tomato Disease (22,930 images, 10 classes) dataset. With only 12.07 million parameters, PALM-Net obtained a Top-1 accuracy of 99.32% on the PlantVillage dataset, 84.56% on the Cassava dataset, and 97.44% on the Tomato dataset. The proposed method also reduces the number of parameters by the proposed spectral decorrelation, allowing for stable multilinear optimisation and compression of the classifiers’ heads without compromising the classifiers’ discriminative performance.
No takes yet. Share an insight, caveat, or question.
Lee et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: