AgroVisNet: A Lightweight Convolutional Network and the BD-PlantDX Expert-Validated Benchmark for Radish, Potato and Pointed Gourd Disease Classification

Forum topic · 小凯 · 2026-09-11

Summary

This arXiv paper (2609.10469) introduces AgroVisNet, a compact convolutional neural network trained from scratch for plant disease diagnosis on low-cost, farmer-held devices, and BD-PlantDX, a new expert-validated benchmark of 12,432 field images covering 12 classes of healthy and diseased radish, potato, and pointed gourd collected in the Bogura and Nilphamari districts of Bangladesh. AgroVisNet combines grouped bottleneck residual blocks with sequential channel and spatial attention, multi-scale depthwise blocks, and a dual-pooling classification head, using only 290,572 trainable parameters. It achieves 99.52% test accuracy and weighted F1 on BD-PlantDX, outperforming six ImageNet-pretrained lightweight backbones with 8.7 to 16.8 times fewer parameters. Quantized to a 0.46 MB full-integer model, it loses only 0.22 percentage points of accuracy and classifies an image in 8.40 ms on a single CPU. The work addresses benchmark bias toward non-native crops, lack of expert validation, and hardware constraints in deployed agricultural AI.

Overview

Research area: Computer Vision (plant disease classification)
Authors: Md. Abdullah Mandal, Saad Ahmed, Md. Khalid Syfullah
Published: 2026-09-09
arXiv: 2609.10469

Abstract

Automated plant disease diagnosis is increasingly deployed on farmer-held devices in regions where agronomic expertise is scarce and network connectivity is unreliable. Three obstacles limit its practical value: public benchmarks are dominated by a small set of non-native crops, region-specific datasets are rarely validated by domain experts, and the architectures that reach competitive accuracy carry parameter budgets that are unsuited to low-cost hardware. We propose AgroVisNet, a compact convolutional network trained from scratch, together with BD-PlantDX, an expert-validated benchmark of 12,432 field images spanning 12 classes of radish, potato and pointed gourd in healthy and diseased states, collected across the Bogura and Nilphamari districts of Bangladesh.

AgroVisNet couples grouped bottleneck residual blocks carrying sequential channel and spatial attention with multi-scale depthwise blocks and a dual-pooling classification head, reaching 290,572 trainable parameters. On BD-PlantDX the model attains 99.52% test accuracy and 99.52% weighted F1, exceeding all six ImageNet-pretrained lightweight backbones evaluated under an identical protocol while using 8.7 to 16.8 times fewer parameters and 1.3 to 8.5 times fewer multiply-accumulate operations.

Deployment and Robustness

Quantized to a 0.46 MB full-integer network with only a 0.22 percentage-point accuracy cost; classifies an image in 8.40 ms on a single CPU.
Across five random seeds, accuracy remains at 99.57 ± 0.10%.
A ten-variant ablation isolates the contribution of each component.
The same architecture transfers without redesign to two independently collected datasets at 98.71% and 99.05% accuracy.
Grad-CAM evidence indicates that predictions rest on lesion-bearing leaf regions rather than on background cues.

---

*Automatically collected on 2026-09-11.*