Lightweight Deep Learning Architectures for Avocado Disease and Pest Recognition: Two Case Studies with Laboratory Fruit and Field Leaf Images

Lightweight deep learning architectures are attractive for diagnosing plant diseases on low-cost field devices, but recent designs have rarely been evaluated on crop-specific tasks. We compared five recent lightweight architectures (MambaOut-Femto, FastViT-S12, MobileOne-S0, MobileNetV4-Conv-Small and EfficientViT-M0) with two conventional baselines (EfficientNet-B0 and ResNet-50) in two independent avocado (Persea americana) case studies. In Scenario 1, 3983 laboratory fruit images (healthy, scab and anthracnose), corresponding to 569 original photographs with seven flipped or rotated versions each, were grouped by photograph and evaluated with photograph-level stratified 5-fold cross-validation. In Scenario 2, 1897 field images of avocado leaves affected by three pests were evaluated with the same protocol. Differences in the Matthews correlation coefficient (MCC) were assessed with frequentist and Bayesian tests. On fruit images, MambaOut-Femto obtained the best performance (sensitivity 85.39%, specificity 92.78%, precision 86.00%, F1-score 85.44%, MCC 78.71% ± 2.05%, AUC 96.17%), followed by MobileOne-S0 (MCC 76.22%) and ResNet-50 (MCC 72.64%), with no significant differences among them, whereas the other four architectures obtained MCC values of 64.35–67.65%. On leaf images, all architectures exceeded an MCC of 93%; FastViT-S12 (sensitivity 97.70%, specificity 98.84%, precision 97.76%, MCC 96.55%) and MambaOut-Femto (97.60%, 98.78%, 97.69% and 96.42%, respectively) ranked highest, and the rankings of the two case studies were only partially consistent (Spearman’s ρ=0.64). MambaOut-Femto was the only architecture among the two best in both case studies, and MobileOne-S0, with 1.06 M parameters and 0.54 GFLOPs after reparameterization, is a lower-cost alternative for the fruit task.

Authors

Institutions

Publication Details

Journal
Agriculture
Published
2026-11-16
DOI
https://doi.org/10.3390/agriculture16222524
Primary Topic
Smart Agriculture and AI
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Lightweight Deep Learning Architectures for Avocado Disease and Pest Recognition: Two Case Studies with Laboratory Fruit and Field Leaf Images

Vidalina Andía-Ayme, Vladimir Acori Flores, José Yauri, Gilmar Peña-Rojas et al.
2 links
Agriculture
Smart Agriculture and AI
article

Lightweight Deep Learning Architectures for Avocado Disease and Pest Recognition: Two Case Studies with Laboratory Fruit and Field Leaf Images

Vidalina Andía-Ayme, Vladimir Acori Flores, José Yauri, Gilmar Peña-Rojas, Diana Luján-Pérez, Emerson Navarro-Guillén, Yerson Oseda, Roy Gomez-Prado, Juan Carlos Quispe-Mendoza, Román Canchari-Gutiérrez
article en

Abstract

Lightweight deep learning architectures are attractive for diagnosing plant diseases on low-cost field devices, but recent designs have rarely been evaluated on crop-specific tasks. We compared five recent lightweight architectures (MambaOut-Femto, FastViT-S12, MobileOne-S0, MobileNetV4-Conv-Small and EfficientViT-M0) with two conventional baselines (EfficientNet-B0 and ResNet-50) in two independent avocado (Persea americana) case studies. In Scenario 1, 3983 laboratory fruit images (healthy, scab and anthracnose), corresponding to 569 original photographs with seven flipped or rotated versions each, were grouped by photograph and evaluated with photograph-level stratified 5-fold cross-validation. In Scenario 2, 1897 field images of avocado leaves affected by three pests were evaluated with the same protocol. Differences in the Matthews correlation coefficient (MCC) were assessed with frequentist and Bayesian tests. On fruit images, MambaOut-Femto obtained the best performance (sensitivity 85.39%, specificity 92.78%, precision 86.00%, F1-score 85.44%, MCC 78.71% ± 2.05%, AUC 96.17%), followed by MobileOne-S0 (MCC 76.22%) and ResNet-50 (MCC 72.64%), with no significant differences among them, whereas the other four architectures obtained MCC values of 64.35–67.65%. On leaf images, all architectures exceeded an MCC of 93%; FastViT-S12 (sensitivity 97.70%, specificity 98.84%, precision 97.76%, MCC 96.55%) and MambaOut-Femto (97.60%, 98.78%, 97.69% and 96.42%, respectively) ranked highest, and the rankings of the two case studies were only partially consistent (Spearman’s ρ=0.64). MambaOut-Femto was the only architecture among the two best in both case studies, and MobileOne-S0, with 1.06 M parameters and 0.54 GFLOPs after reparameterization, is a lower-cost alternative for the fruit task.

AgricultureVol. 16(22)
San Cristóbal of Huamanga University (PE)
Zero hunger
Smart Agriculture and AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon