AgriPath: A Systematic Exploration of Architectural Trade-offs for Crop Disease Classification

Hamza Mooraj; George Pantazopoulos; Alessandro Suglia

cs.CVMar 8, 2026

AgriPath: A Systematic Exploration of Architectural Trade-offs for Crop Disease Classification

Authors: Hamza Mooraj, George Pantazopoulos, Alessandro Suglia

Organizations: School of Mathematical and Computer Sciences Heriot-Watt University · Institute for Language, Cognition and Computation, School of Informatics University of Edinburgh

Abstract

Reliable crop disease detection requires models that perform consistently across diverse acquisition conditions, yet existing evaluations often focus on single architectural families or lab-generated datasets. This work presents a systematic empirical comparison of three model paradigms for fine-grained crop disease classification: Convolutional Neural Networks (CNNs), contrastive Vision-Language Models (VLMs), and generative VLMs. To enable controlled analysis of domain effects, we introduce AgriPath-LF16, a benchmark of 111k images spanning 16 crops and 41 diseases with explicit separation between laboratory and field imagery, alongside a balanced 30k subset for standardised training and evaluation. We train and evaluate all models under unified protocols across full, lab-only, and field-only training regimes using macro-F1 and Parse Success Rate (PSR) to account for generative reliability (i.e., output parsability measured via PSR). The results reveal distinct performance profiles: CNNs achieve the highest accuracy on in-domain imagery but exhibit pronounced degradation under domain shift; contrastive VLMs provide a robust and parameter-efficient alternative with competitive cross-domain performance; generative VLMs demonstrate the strongest resilience to distributional variation, albeit with additional failure modes stemming from free-text generation. These findings highlight that architectural choice should be guided by deployment context rather than aggregate performance alone.

Figures & tables

Panel of Figure 1 — Figure 1: Example of source difference in AgriPath-LF16. Left : Lab-sourced image of Tomato with Bacterial Spot. Right : Field-sourced image of the same disease, illustrating background clutter and lighting variation.

Figure 2: Class and source distribution of AgriPath-LF16-30k across 65 crop-disease pairs. Blue refers to lab-based samples and orange refers to field-based samples.

	PSR (%)	Macro-F1	Macro-Precision	Macro-Recall
Random	–	1.9	1.9	1.9
Majority	–	0.07	0.03	1.5
ResNet-50	–	91.0	92.2	90.7
ConvNeXt-Tiny	–	92.9	93.2	93.0
SigLIP-ZS	–	0.2	0.2	0.7
CLIP/L/14-ZS	–	14.3	19.8	19.0

Table 1: Overall performance on the combined test set. Macro-F1, precision, recall, and Parse Success Rate (PSR) are reported in percentage scale (0–100). PSR is applicable only to generative VLMs. Bold values denote the strongest F1-score within each experimental group.

Trained on full dataset
	PSR (%)	F1(Combined)	F1(Lab)	F1(Field)
ResNet-50	–	91.0	97.2	69.9
ConvNeXt-Tiny	–	92.9	97.1	74.6
SigLIP	–	90.0	93.1	72.9
CLIP/L/14	–	91.1	95.6	74.5
SmolVLM-500M	100.0	87.9	92.9	68.7

Table 2: Generalisation performance across domain-specific training regimes. Macro-F1 scores (percentage scale) are reported for combined, Lab-only, and Field-only test sets. PSR is shown for generative models. Bold values indicate the strongest F1 within each test domain.

Figure 3: A corn crop with common rust in the field with probabilities for a CNN and CLIP prediction.

Figure 4: A potato crop with an ambiguous blight disease in the field with an uncertain CLIP prediction.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset Name	Source	Crop	# Diseases	# Samples	Citation/Link
New Bangladeshi Crop Disease	Lab	Rice	4	4,078	Moin (2023)
PlantVillage	Lab	Apple	4	4,645	G. & J. (2019)
		Blueberry	1	1,502
		Cherry	2	2,502
		Corn	4	4,345
		Grape	4	4,639

Appendix

Table 3: Datasets used to compile AgriPath-LF16 along with links to the download source

	Total Params	Trainable Params	Ratio (%)
Convolutional Neural Networks
ResNet-50	23.6M	15.1M	63.98
ConvNeXt-Tiny	27.9M	14.3M	51.25
Contrastive Vision-Language Models
SigLIP	203M	50k	0.024
CLIP-L/14	427M	50k	0.012

Appendix

Table 4: Model size and trainable parameter ratios. Total and trainable parameters for each architecture and adaptation regime.

Batch Size	Learning Rate	F1(Combined)	F1(Lab)	F1(Field)
16	1e-4	91.0	97.2	69.9
16	2e-4	89.3	95.5	66.1
16	5e-4	88.7	94.8	69.3
32	1e-4	90.7	94.0	70.0
32	2e-4	90.0	93.7	69.3
32	5e-4	88.8	91.4	63.8

Appendix

Table 5: The F1-Scores of the ResNet-50 experiments trained on the full dataset (lab and field)

Batch Size	Learning Rate	F1(Combined)	F1(Lab)	F1(Field)
16	1e-4	56.4	96.7	4.3
16	2e-4	55.4	94.4	4.6
16	5e-4	55.7	96.8	3.6
32	1e-4	56.0	97.0	4.0
32	2e-4	55.6	95.4	4.8
32	5e-4	56.3	96.3	3.5

Appendix

Table 6: The F1-Scores of the ResNet-50 experiments trained on the lab-only subset of the full dataset

Batch Size	Learning Rate	F1 (Combined)	F1(Lab)	F1(Field)
16	1e-4	37.1	13.4	73.2
16	2e-4	35.4	11.0	69.0
16	5e-4	36.7	11.8	69.5
32	1e-4	40.1	15.3	76.7
32	2e-4	33.8	8.4	68.5
32	5e-4	34.4	11.4	68.3

Appendix

Table 7: The F1-Scores of the ResNet-50 experiments trained on the field-only subset of the full dataset

Batch Size	Learning Rate	F1(Combined)	F1(Lab)	F1(Field)
16	1e-4	92.9	97.1	74.6
16	2e-4	92.1	96.5	71.9
16	5e-4	92.6	97.0	73.6
32	1e-4	93.5	97.3	76.7
32	2e-4	92.4	97.3	70.5
32	5e-4	91.6	95.2	69.2

Appendix

Table 8: The F1-Scores of the ConvNeXt-Tiny experiments trained on the full dataset (lab and field)

Batch Size	Learning Rate	F1(Combined)	F1(Lab)	F1(Field)
16	1e-4	56.9	97.5	6.3
16	2e-4	56.7	97.1	6.3
16	5e-4	56.2	96.9	5.8
32	1e-4	57.9	98.1	6.5
32	2e-4	56.5	96.3	7.6
32	5e-4	56.0	96.2	6.0

Appendix

Table 9: The F1-Scores of the ConvNeXt-Tiny experiments trained on the lab-only subset of the full dataset

Batch Size	Learning Rate	F1(Combined)	F1(Lab)	F1(Field)
16	1e-4	40.1	12.8	78.8
16	2e-4	39.2	13.8	74.4
16	5e-4	37.6	12.4	73.4
32	1e-4	40.1	12.0	79.7
32	2e-4	41.7	14.3	80.0
32	5e-4	38.8	13.8	74.9

Appendix

Table 10: The F1-Scores of the ConvNeXt-Tiny experiments trained on the field-only subset of the full dataset

	PSR (%)	F1	Precision	Recall
SmolVLM-500M-ZS-Pure	0.2	0	0	0
SmolVLM-500M-ZS-Context	23.8	0.2	0.6	0.4
SmolVLM-500M-ZS-MCQ	0	0	0	0
Qwen2.5-VL-3B-ZS-Pure	90.7	4.2	5.8	6.3
Qwen2.5-VL-3B-ZS-Context	19.1	2.5	18.9	2.0
Qwen2.5-VL-3B-ZS-MCQ	21.3	24.6	66.2	17.1

Appendix

Table 11: All ZS experiments for all Generative VLMs, reporting PSR, F1, Precision, and Recall in a range of 0-100. Experiments mentioned in the paper are in bold

Parameters	Distribution	Range
Learning Rate	Uniform	$5\times 10^{-5}$ to $2\times 10^{-4}$
LoRA Rank $r$	Categorical	32, 64, 128 (Qwen3) \| 64, 128 (Qwen7)
Weight Decay	Uniform	$0$ to $0.1$

Appendix

Table 12: The initial parameters used for sweeps. Qwen3 and Qwen7 refer to the 3B and 7B variants of the Qwen2.5-VL-xB-LoRA experiments respectively.

Parameters	Distribution	Range
Learning Rate	Uniform	$5\times 10^{-5}$ to $1.5\times 10^{-4}$
Weight Decay	Uniform	$0$ to $0.1$
LoRA Rank $r$	Fixed	128

Appendix

Table 13: The refined parameters used for sweeps after the initial sweeps showed some trends of convergence.

Sweep Name	Regime	Search Space
Qwen3	Full LoRA	Initial
Qwen3-LAB	Lab LoRA	Initial
Qwen3-FV	Frozen Vision	Refined
Qwen3-FIELD	Field LoRA	Refined
Qwen7	Full LoRA	Initial
Qwen7-LAB	Lab LoRA	Initial

Appendix

Table 14: A summary of the search spaces used for each sweep. Qwen3 relates to Qwen2.5-VL-3B, Qwen7 relates to Qwen2.5-VL-7B, and Smol relates to SmolVLM-500M. Initial relates to section E.3.1 , and Refined relates to section E.3.2

Figure 5: Top confusion pairs are discussed in Section 4.6.1 . Actual labels are along the y-axis, and Predicted labels are along the x-axis.

Figure 6: Top confusion pairs are discussed in Section 4.6.2 . Actual labels are along the y-axis, and Predicted labels are along the x-axis.

Plant Disease Classification

Plant disease classification uses image and sensor data to identify crop diseases and support timely, targeted agricultural intervention.

See all 58 papers

Vision-language models know more about agriculture than they show and rubric-grounded verifications close the gap
Sep 8, 2026Earl Ranario, Jared Smith, Lars Lundqvist +2Agricultural Image Analysis Plant Disease Classification
University of California, Davis
SAGE: Scalable Agentic Grounded Evaluation for Crop Disease Diagnosis
May 10, 2026Muhammad Arbab Arshad, Tirtho Roy, Yanben Shen +7VLM Evaluation Agricultural Image Analysis
Iowa State University, USA · New York University, USA
Multi-Crop Leaf Disease Recognition: A Unified Benchmark and Cross-Region Study
Oct 3, 2026Rosemary Nalwanga, Sebastian Bunda, Godliver Owomugisha +2Plant Disease Classification Cross-Domain Generalization
Data Management and Biometrics University of Twente Enschede, Netherlands · Department of Computer Engineering Busitema University Tororo, Uganda

Sep 8, 2026J/K move · Enter open · S save

Vision-language models know more about agriculture than they show and rubric-grounded verifications close the gap

Earl Ranario, Jared Smith, Lars Lundqvist, Urmil Jatin Chandarana+1

University of California, Davis

Agricultural Image Analysis Plant Disease Classification VLM Reasoning Classification Rubric-Based Evaluation VLM Calibration

Open paper AskPDF arXiv

Vision-language models (VLMs) show promise for agricultural classification, but zero-shot performance on disease, pest, damage, quality, and species identification remains poor, and it is unclear whether this reflects weak visual features or a failure to connect them to domain knowledge. We build a benchmark of 116 datasets, 834 classes, and 8,324 images spanning these tasks to isolate where the gap arises. Linear probing shows VLM vision encoders already encode agricultural features nearly as separable as a self-supervised DINOv3 baseline, ruling out weak visual representations as the primary bottleneck. Conditioning each model on an oracle reference description (an upper bound on its parametric knowledge) closes most of the gap left by an unaided lower bound, showing VLMs already know more about agriculture than they show. To close this gap without an oracle description at inference time, we structure test-time reasoning around a fixed, per-task diagnostic rubric: the model generates

K

candidate responses and a Probabilistic Pivot Tournament (PPT) verifier, scored pairwise against the rubric, selects the best one. This nearly doubles judged F1 over the lower bound and matches or exceeds the upper bound on several tasks, notably pushing Gemma 4 E4B-it's disease F1 to 0.71, above its own upper bound of 0.60. However, the verifier's letter-scale confidence score has the opposite of its intended effect: filtering to its most confident predictions does not improve accuracy and correlates negatively with correctness across every model and pool size tested, so the score cannot serve as a measure of predictive uncertainty, and most of the observed gain likely comes from rubric-grounded generation rather than pairwise verification.

Abstract

Plant Disease Classification

Explore similar work

Vision-language models know more about agriculture than they show and rubric-grounded verifications close the gap

SAGE: Scalable Agentic Grounded Evaluation for Crop Disease Diagnosis

Multi-Crop Leaf Disease Recognition: A Unified Benchmark and Cross-Region Study