NobleBlocks

Research indexJournal article

A Genetic Algorithm-Driven Framework for Convolutional Kernel Selection in Lightweight Plant Disease Classifiers

HSSAYNI

International Journal of Advanced Computer Science and Applications1 Jan 2026Open access

NobleIDark:/48914/w1/20261105/40A28820DOI10.14569/ijacsa.2026.0170784

Free to read at the publisher

You can read the abstract here, or the full text at the publisher.

Read at the publisher

Abstract

Foliar pathogens destroy roughly a third to two-fifths of harvestable crops each year, yet the deep convolutional classifiers that have proven most effective at automated leaf-image diagnosis carry parameter counts in the tens of millions prohibitive for the microcontrollers and low-power accelerators aboard agricultural drones and field-deployed IoT nodes. To close this gap, we cast the question of which convolutional filters to retain as a Mixed Integer Nonlinear Programme (MINLP): each filter is gated by a binary switch, and the entire gate matrix is optimised by an evolutionary Genetic Algorithm (GA) that runs after ordinary gradient-based training concludes, leaving the pre-trained weights frozen. Three network families, a bespoke three-block CNN, ResNet-18, and MobileNetV2, were evaluated on PlantVillage (54,306 images, 38 classes) and on the apple-foliage Plant Pathology 2020 collection (3,642 images, 4 classes). Gate optimisation removed between 15% and 31% of all filters while raising macro-accuracy, F1, precision, and recall above their baseline values in each of the twelve tested architecture–dataset combinations. Against five reference methods (ℓ1, T ℓ1, structured pruning, knowledge distillation, and no compression), our scheme was strictly superior in every cell of the comparison table (p<0.05, Wilcoxon test). Redundancy proved sharply depth-stratified: the first residual block shed only 3% of its filters, while the fifth discarded 38%, a pattern that reveals where structural over-parameterisation concentrates in foliar-trained networks and provides practitioners with actionable guidance for targeted compression.

Discussion

More

Suggest a correction or report