PlantVillageVQA: Visual Q&A for Plant Health

• PlantVillageVQA is a plant-science visual question answering dataset that pairs over 55,000 leaf images with 193,609 QA pairs to enable interactive plant disease analysis.
• It organizes questions into three cognitive complexity levels across nine categories, fostering tasks from basic plant identification to high-order causal reasoning.
• Benchmark results reveal that models like CLIP and LXMERT excel, demonstrating the dataset’s potential to advance multimodal plant pathology research.

PlantVillageVQA is a plant-science visual question answering dataset built from the PlantVillage leaf-image corpus to benchmark and train vision-LLMs for plant disease diagnosis, agricultural decision support, and multimodal reasoning. It comprises 193,609 question-answer pairs grounded over 55,448 images spanning 14 crop species and 38 disease conditions; questions are organized into 3 levels of cognitive complexity and 9 categories. The dataset was designed to move plant-health modeling beyond closed-set image classification toward interactive plant disease analysis, and its creators describe it as, to their knowledge, the first multimodal plant-science VQA dataset created under extensive guidance by domain experts specifically for model training and evaluation (Sakib et al., 23 Aug 2025).

1. Scope, composition, and representational form

PlantVillageVQA is grounded entirely in PlantVillage images and labels, but its supervision target is not a class label alone. Each sample binds an image to a natural-language question and answer intended to probe crop recognition, health assessment, symptom grounding, disease identification, explanation, causation, and counterfactual reasoning. The released dataset includes 2 annotation files (.csv and .json), and each annotation record includes image_id, question_type, question, answer, image_path, and split (Sakib et al., 23 Aug 2025).

Item	Value
Images	55,448
QA pairs (final)	193,609
Initial QA pairs before refinement	278,255
Crop species	14
Disease conditions / classes	38
Question categories	9
Cognitive-complexity levels	3

The dataset also records substantial linguistic expansion relative to its initial template output. Question vocabulary grows from 178 to 818 words, answer vocabulary from 148 to 201, and the mean unique words per QA pair is 14.31. These properties reflect a deliberate shift from rigid metadata verbalization toward more varied image-grounded language. The released image layout is flattened into paths such as images/image_000001.jpg, while provenance remains encoded in the annotations through metadata-derived question generation.

Distribution is provided through Figshare, Hugging Face, Kaggle, and GitHub, and the paper does not explicitly state a software or data license in the provided text (Sakib et al., 23 Aug 2025).

2. Dependence on PlantVillage and the consequences of that grounding

The source corpus is the PlantVillage image repository. In the PlantVillageVQA paper, PlantVillage is described as an open-access repository containing 55,448 images of plant leaves across 38 categories. The foundational PlantVillage repository paper reported 54,309 images spanning 14 crop species and 38 classes, collected as detached leaves photographed against grey or black paper backgrounds, outdoors, typically with 4–7 images per leaf, then cropped and oriented so the tip pointed upward (Hughes et al., 2015). This suggests reliance on a later mirrored or re-packaged release rather than strict reuse of the original 2015 count.

Grounding in PlantVillage is operational rather than merely thematic. Folder names such as Tomato_Late_blight and Apple_Healthy provide the crop label and the disease-or-health label, and these metadata are inserted directly into question and answer templates. That design gives PlantVillageVQA auditable image-level provenance and makes large-scale QA synthesis feasible without unconstrained hallucination-prone generation. It also means that the ontology of PlantVillageVQA inherits the PlantVillage crop–disease label space almost exactly.

The same inheritance imposes strong constraints. PlantVillage images were acquired under controlled conditions: detached leaves, homogeneous backgrounds, standardized orientation, and high-quality close-up imaging (Hughes et al., 2015). PlantVillageVQA therefore supports fine-grained leaf-level visual reasoning, but it does not by itself supply whole-plant context, field clutter, multi-organ evidence, or naturally occurring environmental variation. A plausible implication is that benchmark performance on PlantVillageVQA measures competence on image-grounded pathology language under controlled conditions more directly than it measures readiness for in-field agronomic deployment.

3. Generation pipeline, balancing, and scientific validation

PlantVillageVQA was constructed through a multi-stage workflow centered on auditable metadata grounding and repeated quality control (Sakib et al., 23 Aug 2025). The initial stage read each PlantVillage image’s parent directory, extracted crop and disease-or-healthy labels, and populated predefined templates for nine question categories. This produced 278,255 initial QA pairs. The authors explicitly avoided free-form LLM-based generation at this stage to reduce hallucinations, unsupported claims, and unverifiable text, and they formulated and refined the questions with feedback from three graduate students in Botany at the University of Dhaka.

The initial corpus was then linguistically re-engineered. Repetition was severe: “Provide a full description of the plant and its condition.” appeared 27,809 times; “If this plant were healthy, what visual features would be different?” 25,152 times; “Please provide a diagnosis for the condition shown” 15,004 times; “What pathogen is responsible for the symptoms displayed here?” 13,118 times; and “Does this leaf appear to be free of disease?” 10,842 times. The top five repeated questions together accounted for 151,952 instances. To reduce this rigidity, the authors manually built paraphrase pools, typically with 10–15 high-quality paraphrases, and diversified answer bundles so that disease responses were not limited to a single generic statement.

A separate structural-balancing stage targeted categories with highly skewed binary outputs. Visual Attribute Grounding, Detailed Verification, and General Health Assessment were 100% binary and 78.6% skewed toward negative answers. The authors therefore applied targeted stratified undersampling: all “Yes” answers were preserved, “No” answers were randomly reduced, and the goal was a 40/60 binary ratio within the problematic subsets. After reintegration, the overall binary answer ratio became 59.4%, and the final dataset size became 193,609 QA pairs.

Counterfactual questions required the most substantial repair. In phase-one expert review, the main issue was that counterfactual answers often failed to answer the hypothetical. The hierarchical correction pipeline was applied to all 27,242 counterfactual QA pairs. It mined 16,486 Visual Attribute Grounding questions, extracted canonical symptom phrases, built a canonical_phrase map, and used a “fix when verifiable; delete when not” rule. A pair was considered verifiable only if it had valid crop_disease provenance, a disease_keyword mapping, and an expert symptom phrase. Verifiable answers were regenerated using the template “A healthy leaf would not show … [canonical symptom] …”. The result was 9,981 (36%) corrected/regenerated pairs and 17,261 (63.7%) deleted as non-verifiable. Generic fallback answers dropped from 14,991 to 1,496, a 90% reduction in generic answers as a share of all counterfactuals.

The final quality-control layer combined automated outlier mining with a second expert review. Automated analysis assessed relative simplicity, vagueness by TF-IDF, and semantic similarity using sentence-transformer all-MiniLM-L6-v2 with cosine similarity, flagging 25,418 QA pairs. In phase two, experts reviewed a random sample of 2,837 flagged items, discarded 91, and reported 97.57% correctness. The paper presents this figure as its main quantitative claim about annotation quality, while explicitly noting that it is not an inter-annotator agreement statistic (Sakib et al., 23 Aug 2025).

4. Cognitive hierarchy and question taxonomy

A distinguishing feature of PlantVillageVQA is its explicit organization into three cognitive-complexity levels and nine question categories (Sakib et al., 23 Aug 2025). The taxonomy is intended to separate direct visual recognition from symptom-level verification and from higher-order explanation or hypothetical reasoning.

At Level 1: Foundational Perception and Identification, the categories are Existence / Sanity Check, Plant Species Identification, and General Health Assessment. These questions test whether the image contains the expected plant subject, whether the crop can be identified, and whether the plant is healthy or unhealthy. They are the closest analogues to classical image classification, but they are expressed through natural-language QA.

At Level 2: Detailed Analysis and Verification, the categories are Visual Attribute Grounding and Detailed Verification. Visual Attribute Grounding asks whether canonical symptoms are visible, as in “Does the leaf exhibit dark, concentric 'bullseye' rings?” Detailed Verification asks joint crop+disease binary questions such as “Is this [Crop] leaf infected with [Disease_Name]?” This level makes symptom wording explicit and converts disease recognition into a compositional verification task.

At Level 3: Higher-Order Reasoning and Inference, the categories are Specific Disease Identification, Comprehensive Description, Causal Reasoning, and Counterfactual Reasoning. These questions require open-ended diagnosis, holistic image description, inference about cause, and hypothetical contrastive answers such as how a healthy leaf would differ from the image shown. The taxonomy is therefore not limited to label retrieval; it attempts to elicit explanation, abstraction, and conditional reasoning.

The symptom vocabulary used in Visual Attribute Grounding is reported as being based on established phytopathology literature. This design choice matters because it constrains question semantics to domain-recognizable evidence rather than purely generic computer-vision phrasing. It also partly explains why counterfactual questions were the most fragile category: they required symptom-grounded negation rather than simple label restatement.

5. Benchmarking and observed model behavior

The dataset paper benchmarks PlantVillageVQA using CLIP, LXMERT, and FLAVA (Sakib et al., 23 Aug 2025). The annotations include train and test splits, although the paper does not provide exact train/test counts. Evaluation combines Accuracy, BLEU, METEOR, ROUGE-1, and ROUGE-L. Accuracy is used for binary yes/no and single-word answers; BLEU, METEOR, and ROUGE are used for more descriptive categories.

Model	Accuracy	BLEU	METEOR	ROUGE-1	ROUGE-L
FLAVA	0.3432	0.0909	0.2066	0.3826	0.3764
CLIP	0.6148	0.2452	0.4476	0.7323	0.7153
LXMERT	0.6034	0.2382	0.4393	0.7219	0.7061

The headline empirical result is that CLIP performs best overall, with LXMERT close behind and FLAVA substantially weaker. The paper interprets this as evidence that the dataset is both learnable and challenging. The stronger results on simpler tasks, together with only moderate scores on generative metrics, indicate that image-grounded plant QA cannot be reduced to trivial metadata memorization under the benchmark protocol.

The dataset paper reports that models perform relatively well on Plant Species Identification, Detailed Verification, and Existence / Sanity Check, while struggling more on Causal Reasoning, Counterfactual Reasoning, Detailed Verification, and Specific Disease Identification. The repeated appearance of Detailed Verification in both stronger and weaker groupings is a mild inconsistency in the paper’s own qualitative summary. The most charitable reading is that binary compositional verification is easier than open-ended reasoning on average, but disease-specific difficult cases remain.

The pattern of automatic metrics is also informative. CLIP and LXMERT obtain ROUGE-1 above 0.72, but BLEU remains around 0.24 and METEOR around 0.44. This suggests substantial keyword recovery without full semantic fidelity. A plausible implication is that models often retrieve the right crop or disease terms while still underperforming on richer explanatory or paraphrastically varied answers.

6. Bias inheritance, explainability concerns, and later methodological context

Because PlantVillageVQA is fully grounded in PlantVillage, it inherits PlantVillage’s acquisition biases as well as its strengths. A short diagnostic study showed that, on PlantVillage, a random forest trained on only eight pixels sampled from the image background achieved 49.0% test accuracy over 38 classes, far above the 2.6% random-guess baseline (Noyan, 2022). The same study argued that the dominant issue is capture bias rather than semantic background context. For PlantVillageVQA, this implies that some apparently correct answers could be supported by acquisition traces rather than by visible pathology.

Explainability analysis on PlantVillage classification reinforces the same concern. An ACE-based study using InceptionV3 on PlantVillage reported that automated concept discovery recovered biologically plausible symptom concepts, but also exposed background and shadow/light artifacts as salient features. It highlighted lower-performing classes such as Apple Scab with recall 0.74, F1 0.85, and Tomato Early Blight with recall 0.59, F1 0.72, and linked some failures to background-related concepts rather than disease symptoms (Amara et al., 2024). This suggests that PlantVillageVQA evaluation is strongest when paired with explicit grounding audits rather than answer-only scoring.

PlantVillageVQA also sits within a rapidly broadening agricultural VQA landscape. Later work has proposed training-free explainable agricultural QA pipelines based on neutral symptom-rich captions and judge-based answer selection, but without evaluating on PlantVillageVQA itself (Zhang et al., 31 Dec 2025). Other work has expanded plant VQA toward multi-step, intent-driven inquiry with severity labels, safety-sensitive evaluation, and question sequences rather than single-turn QA (Sakib et al., 22 Apr 2026). Still other research has supplied expert-verified rationale supervision and confidence labels for leaf-disease diagnosis, though not as a native multi-question VQA benchmark (Mahmood et al., 9 Apr 2026). This suggests that PlantVillageVQA occupies an important transitional position: it turns a major plant-disease image corpus into a large, expert-reviewed QA benchmark, but it remains closer to controlled leaf-image reasoning than to fully interactive botanical diagnosis.

Its main limitations follow directly from its construction. It depends on PlantVillage’s controlled distribution; much of its supervision is metadata-derived; several categories were initially structurally imbalanced; and a large fraction of counterfactuals had to be deleted to preserve verifiability (Sakib et al., 23 Aug 2025). Its main contribution is therefore best understood not as solving plant visual reasoning, but as defining a standardized, expert-guided benchmark on which that broader research program can be measured.