Work

AGRI-KD: A lightweight multi-task framework for rice disease diagnosis with localization and severity quantification

Year: 2026

Type: article

Abstract: Automated rice disease diagnosis has largely been framed as a classification problem, yet a class label alone is insufficient for field decisions: effective disease management depends on knowing where the infection is, how much leaf area it covers, and how severe it has become. Methods that supply this spatial and quantitative information, such as object detectors or segmentation networks, are typically too computationally heavyweight for the entry-level smartphones available to smallholder farmers, while lightweight classifiers omit it entirely. We present AGRI-KD, a single lightweight model that resolves this tension by producing five outputs simultaneously: disease classification, spatial localization, infected area percentage (IAP), severity grade, and lesion count. A three-teacher ensemble (ResNet-50, DenseNet-121, EfficientNet-B0) is distilled into a 6.97 MB MobileNetV3-Small student through multi-teacher knowledge distillation, learned attention supervision that trains localization maps directly rather than relying on post-hoc visualization, and dedicated agricultural prediction heads. On the field-collected RiceLeafDiseaseBD dataset (9,769 images, 21,460 verified bounding boxes, six classes), AGRI-KD preserves baseline classification accuracy (90.55% vs. 90.41%, 𝑝 = 0 . 7 2 4 ) while significantly improving localization Dice from 0.276 to 0.575 ( p < 0.001, Cohen’s 𝑑 = 1 . 4 9 9 ). Furthermore, it delivers robust agricultural quantification, achieving leaf-area IAP estimation (5.91% MAE, 𝑅 2 = 0 . 6 8 9 ), severity grading (75.1% accuracy, F1-macro 0.645), and lesion counting ( 𝑅 2 = 0 . 7 1 3 ). The model runs in 14.8 ms on a budget smartphone without GPU or network access. By integrating classification, trained localization, and agricultural regression in one deployment-validated model, AGRI-KD delivers comprehensive disease diagnosis within the constraints of entry-level mobile hardware.

Source: Smart Agricultural Technology

DOI: 10.1016/j.atech.2026.102550

Authors: Mohammad Mehedi Hasan Munna, Acramul Haque Kabir, Reefka Fabliha Tulona

Institutions: East West University, North South University, Independent University, Bangladesh

Language: English

Topic: Smart Agriculture and AI

Subfield: Plant Science

Field: Agricultural and Biological Sciences

Domain: Life Sciences

SDG: Zero hunger

Open Access status: gold

Estimated APC paid: $1,770