Source-linked AI summary
AgroVisNet: A lightweight Convolutional Network and the BD-PlantDX Expert-Validated Benchmark for Radish, Potato and Pointed Gourd Disease Classification
Md. Abdullah Mandal, Saad Ahmed, Md. Khalid Syfullah
TL;DR
Plant disease diagnosis needs region-specific, expert-validated data and models small enough for low-cost devices. The paper introduces BD-PlantDX and AgroVisNet, then reports strong benchmark performance, efficient computation, transfer to independent datasets, and lesion-focused explanations, while identifying limits under uncontrolled field conditions.
Problem
Public benchmarks underrepresent region-specific crops, expert validation is uncommon in local datasets, and competitive architectures can exceed low-cost hardware budgets.
Method
The paper builds the expert-validated BD-PlantDX benchmark and proposes AgroVisNet, a compact convolutional network trained from scratch with attention, grouped residual, multi-scale depthwise and dual-pooling components.
Results
99.52% accuracy and weighted F1 on BD-PlantDX exceeded six ImageNet-pretrained lightweight backbones, while transfer accuracy reached 98.71% and 99.05% on two independent datasets.
Takeaways & Limitations
The reported results support a compact from-scratch architecture as a competitive option for region-specific plant disease classification and cross-dataset evaluation.
Takeaways & Limitations
BD-PlantDX was acquired under controlled illumination and uniform backgrounds, so performance under field illumination, clutter and partial occlusion remains unmeasured.
Abstract
from arXiv · showhide
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. Exported for deployment, the model quantises to a 0.46 MB full-integer network at a 0.22 percentage-point accuracy cost and 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, and 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.
1. Introduction
The paper addresses region-specific plant disease diagnosis by introducing an expert-validated Bangladesh benchmark and a compact network trained from scratch. It evaluates accuracy, efficiency, generalization, reproducibility, component contributions, and whether predictions use lesion-bearing regions.
- Contributions: 12,432 expert-validated images cover 12 healthy and diseased classes of radish, potato and pointed gourd from two Bangladesh districts.The benchmark includes root and foliar presentations, with class assignments verified by practicing agronomists.
- Contributions: AgroVisNet combines attention-equipped grouped residual blocks, multi-scale depthwise blocks and dual pooling within 290,572 trainable parameters.The network is trained from scratch and composes established components into a compact architecture.
- Research questions: The study compares AgroVisNet with six ImageNet-pretrained lightweight backbones under an identical pipeline, split and evaluation protocol.The comparison targets both predictive performance and computational cost.
- Research questions: The evaluation tests component contributions, transfer without redesign, stability across random seeds, and localization of model attention on disease-bearing regions.These tests correspond to the paper’s architectural, generalization, reproducibility and interpretability questions.
2. Background
The background frames plant disease recognition as fine-grained image classification and motivates efficient feature extraction. It introduces residual, grouped, and depthwise operations as mechanisms for retaining representational capacity while reducing computational cost.
- Pipeline context: Figure 1 places these concepts within an end-to-end image diagnosis pipeline, while formal mechanism descriptions are deferred to the model section.The conceptual pipeline links image acquisition, preprocessing, representation learning, interpretability and diagnosis.
- Plant disease classification: Plant disease classification assigns a leaf or root image to the highest-scoring healthy or diseased condition in a fine-grained label set.Classes may differ mainly in lesion texture, margins or yellowing while sharing broader morphology and color.
- Residual learning: Residual learning adds a direct input-to-output path, while bottlenecks reduce the channel space used by expensive spatial convolutions.The direct path supports gradient transmission, and the bottleneck lowers spatial-convolution cost.
- Efficient convolutions: Grouped convolution partitions channels into independent groups, and depthwise separable convolution extends this idea before recombining channels with a pointwise projection.These operations reduce the dominant channel-product cost of conventional convolution.
1 FIELD ACQUISITION
The field-acquisition stage captures leaf or root images under real-world variation in lighting, viewpoint and background.
- 2 FIELD ACQUISITION: Leaf or root images are captured in real field conditions with variation in lighting, viewpoint and background.
2 PREPROCESSING
The pipeline standardizes incoming imagery before representation learning and uses attention, regularization and interpretability tools to focus diagnosis on informative evidence.
- 2 PREPROCESSING: Images are resized, normalized and optionally augmented so disease evidence reaches the network in a consistent form.
- Representation learning: Channel and spatial attention rescale feature maps to amplify informative filters and locations while suppressing less useful responses.Channel attention summarizes feature maps into channel multipliers, whereas spatial attention produces location-specific multipliers.
- Representation learning: Batch normalization stabilizes optimization, dropout reduces unit co-adaptation, and swish provides a smooth non-monotonic activation used by AgroVisNet.Spatial dropout removes entire feature maps, matching convolutional tensors.
- Interpretability: Grad-CAM weights late-layer activations by their influence on the predicted class and overlays the resulting localization map on the input image.In this pipeline, the method helps distinguish lesion-focused evidence from background or acquisition artifacts.
3. Literature Review
Prior work spans transfer learning, crop-specific models, efficient architectures, regional datasets, and interpretability, but existing methods rarely satisfy all three requirements of local validation, low-cost deployment, and cross-dataset robustness.
- Regional datasets improve local relevance, but prior releases cover single crop families or lack documented per-image expert validation.The cited regional collections include Bangladeshi vegetable and radish imagery, while BD-PlantDX combines broader coverage with validation.
- Interpretability studies use Grad-CAM, LIME, or SHAP, but report selected visual explanations without quantitative agreement against expert annotations.These works treat interpretability primarily as a qualitative trust aid.
- Existing methods satisfy at most two of three requirements: expert-validated local data, low-cost parameter budgets, and accuracy that survives dataset changes.BD-PlantDX and AgroVisNet are presented as addressing the combined gap.
- Transfer-learning studies rely on public benchmarks, while purpose-built crop models, efficient backbones, and regional releases each address only subsets of the target setting.The review distinguishes their respective limitations across benchmark fit, efficiency, and regional coverage.
4. The BD-PlantDX Dataset
BD-PlantDX is an expert-validated, multi-crop field-image benchmark from northern Bangladesh, covering diverse disease etiologies and using controlled, stratified specimen-level partitioning.
- 4. The BD-PlantDX Dataset: Field surveys in Bogura and Nilphamari used agronomist-guided plot selection and consumer smartphones, aligning acquisition with realistic grower-held inference hardware.Agronomists identified diseased plots before specimen harvesting and imaging.
- 4. The BD-PlantDX Dataset: Images underwent screening and practicing-agronomist review of pathology, affected organ, and severity, with disagreements removed rather than resolved by majority.The signed validation record accompanies the release.
- 4. The BD-PlantDX Dataset: The dataset covers six radish, four potato, and two pointed-gourd classes spanning healthy states, fungal and viral infections, insect damage, and nutrient disorder.Radish includes root and leaf healthy presentations; potato and pointed gourd classes are primarily foliar.
- 4.5. Partitioning: The fixed split uses 8,694 training, 1,868 validation, and 1,870 test images under stratified 70/15/15 partitioning, with specimen-level separation and one-time test evaluation after model selection.Specimen-level partitioning prevents multiple views of one physical leaf from crossing splits.
- 4.5. Partitioning: The maximum class imbalance ratio is 1.16 across the training, validation, and test partitions.
- 4. The BD-PlantDX Dataset: BD-PlantDX contains 12,432 field images from 12 classes across radish, potato, and pointed gourd, combining local multi-crop coverage, root and foliar presentations, four etiologies, and per-image expert validation.The collection is positioned against PlantVillage and earlier Bangladeshi releases as broader in crop-system and validation scope.
5. Methodology
AgroVisNet is formulated as a sub-0.3M-parameter classifier and combines grouped residual attention, multi-scale depthwise processing, and dual pooling for compact plant-disease recognition.
- Problem formulation: AgroVisNet enforces a budget below 0.3M trainable parameters while targeting performance comparable to much larger backbones.
- Architecture overview: The network uses a preprocessing front end, strided stem, four feature stages, and a dual-pooling classification head; stage widths and group counts increase from (32, 2) to (128, 16).
- Architecture overview: Each feature stage pairs two Enhanced Residual Blocks with one Multi-Scale Depthwise Block, combining grouped bottlenecks, sequential channel-spatial attention, and parallel 3 × 3 and 5 × 5 depthwise responses.
- Enhanced Residual Block: The grouped residual bottleneck reduces spatial-convolution weights by 4g, reaching a factor of 64 in the final stage where g = 16.
- Classification head: Dual pooling combines global average and maximum responses so diffuse and focal disease evidence can both contribute to classification.
- Evaluation protocol: The evaluation reports accuracy, weighted precision, weighted recall, and weighted F1 on a held-out test split, with repeated five-seed training for AgroVisNet and pretrained baselines.
6. Experimental Results
AgroVisNet achieves high accuracy on BD-PlantDX while using substantially fewer parameters and compute than the evaluated alternatives. Its compact deployment formats retain accuracy, reduce latency, and show stable performance across random seeds.
- Classification performance: 99.52% accuracy is supported across the 1,870-image test set, with five classes error-free and the lowest class F1 at 97.75%.Radish Flea Beetle Damage has 95.60% precision but 100% recall, while macro F1 also remains 99.52%.
- Classification performance: Nine of 1,870 test images are misclassified, with seven assigned to Radish Flea Beetle Damage because sparse punctures resemble incidental leaf damage at 224 × 224 resolution.The authors characterize this residual error as a resolution limitation and identify higher-resolution acquisition as the natural remedy.
- Computational efficiency: 132.70M MACs and 1.13 MB make AgroVisNet the cheapest and smallest model, with no competitor simultaneously more accurate and cheaper to run.Its MAC cost is 1.3 to 8.5 times lower and its parameter budget is 8.7 to 16.8 times lower than the pretrained backbones.
- Deployment and computational efficiency: Full-integer INT8 quantization produces a 0.46 MB model within 0.22 percentage points of the original accuracy, while the deployment benchmark measures 2.33 ms per image at four CPU threads.The INT8 format is fastest across the tested thread counts on the evaluated Android handset.
- Stability across random seeds: Five training seeds yield 99.57% mean test accuracy with a 0.10-point standard deviation and a 99.47% to 99.68% range.The lower bound exceeds every pretrained baseline, although best epochs vary from 50 to 110 and training time from 56.6 to 98.7 minutes.
6.6. Statistical Significance of the Margin over the Baselines
Across five seeds, AgroVisNet outperformed every pretrained backbone, with paired confidence intervals excluding zero and convergent significance evidence supporting a reliable accuracy margin.
- AgroVisNet exceeded every pretrained backbone on all five seeds, with mean paired advantages of 0.33 to 0.61 percentage points and 95% confidence intervals excluding zero.The paired differences ranged from 0.33 percentage points against RepViT-M1.0 to 0.61 against MobileNetV4-Conv-Small.
- The Wilcoxon signed-rank test reached no significance because five same-sign pairs impose a minimum exact two-sided p-value of 0.0625, above 0.05 before correction.Holm correction increased the raw value to 0.375 for every baseline.
- The evidence combines a five-of-five sweep, six confidence intervals excluding zero and significant paired t-tests, while a larger seed budget is needed for the rank test to certify the difference.
- The ablation protocol fixed the seed, split and schedule, but its full-model accuracy was 99.04%, compared with 99.52% in the main run.Ablation comparisons therefore remain internal to that protocol.
- Removing swish caused the largest ablation loss at −0.27 percentage points, while removing the multi-scale depthwise block or grouped convolutions raised parameters and lowered accuracy.The multi-scale depthwise removal increased parameters by 56.8%, and grouped-convolution removal increased them by 41.0%.
6.10. Interpretability
AgroVisNet’s explanations concentrate on lesion-bearing leaf tissue rather than controlled backgrounds, and faithfulness and randomization tests support model-dependent localization. The analysis also shows that errors arise mainly from visual confusion between classes, while field-condition robustness remains unmeasured.
- Qualitative attribution: Attribution peaks on symptomatic leaf tissue and falls away over the uniform background across eight representative classes.The maps concentrate on chlorotic, necrotic, or mildew-covered regions rather than acquisition artifacts.
- Where the attribution refines: Across network stages, Grad-CAM evolves from broad contour and venation responses into a consolidated symptomatic region.The first stage is weakly class-selective, whereas later stages separate affected lamina from background.
- Failure analysis: All five analyzed errors localize to the correct leaf, indicating discrimination failure rather than attention failure; seven of nine errors involve Radish Flea Beetle Damage.Four of those errors come from Radish Leaf Healthy, with others from Potato Mosaic and Gourd Healthy.
- Faithfulness: Grad-CAM is more faithful than Grad-CAM++ on masking metrics, with confidence drop of 41.5% versus 61.6% and insertion AUC of 0.932 versus 0.925.Both methods outperform random ordering on insertion, but deletion AUC does not separate either method from the random control in leaf-dominated imagery.
- Sanity check: After cascading weight randomization, Grad-CAM maps become near-uniform instead of reverting to input contours, supporting dependence on learned parameters.This sanity check addresses whether localization is merely an edge-detection artifact.
- 99.52% accuracy remains above all six pretrained baselines, while five-seed accuracy is 99.57±0.10%.The interpretability section situates the attribution findings within the model’s strong and stable benchmark performance.
- Interpretability conclusions remain bounded because imagery used controlled illumination and uniform backgrounds, while field clutter and occlusion were not measured.Synthetic corruptions showed tolerance to compression and blur but sensitivity to additive noise and brightness shift.
7. Conclusion
The paper presents AgroVisNet and BD-PlantDX as a compact model–benchmark combination for plant disease classification. It reports strong benchmark performance and identifies field-condition evaluation and embedded deployment characterization as remaining work.
- AgroVisNet has 290,572 parameters, while BD-PlantDX contains 12,432 expert-validated field images across 12 classes from three Bangladeshi crops.The network combines grouped bottleneck residual attention blocks, multi-scale depthwise blocks, and dual pooling.
- Future work must test natural backgrounds and variable illumination, measure sustained-load latency and memory on target hardware, and add severity grading.The current synthetic corruption sweep is described as an initial step rather than a substitute for field-condition evaluation.
Declaration of Generative AI and AI-Assisted Technologies in the Manuscript Preparation Process
The manuscript reports that LLMs supported language editing and code debugging, while authors verified and assumed responsibility for the resulting work. The paper states that AI assistance did not contribute to the scientific claims or technical outputs.
- LLMs were used only for language improvement and code debugging, with all AI-assisted material examined, verified, and corrected by the authors.
- Figures, tables, quantitative outcomes, technical descriptions, and reported numerical values were generated solely by the authors without AI contribution.
- AI-generated coding recommendations required manual design, correction, and iterative testing before author-verified execution produced the reported results.
- The authors state that no sensitive or proprietary data were disclosed to AI tools and that LLM use had no bearing on the scientific contributions.
CRediT authorship contribution statement
The authors’ contributions span conception, data, analysis, investigation, methodology, software, validation, supervision, and writing. Md. Abdullah Mandal led the broadest set of implementation and study responsibilities.
- Md. Abdullah Mandal contributed across conceptualization, data curation, methodology, software, visualization, validation, and original drafting.
- Saad Ahmed and Md. Khalid Syfullah contributed to analysis, resources, supervision, and manuscript review, with additional project administration by Syfullah.