A Data-Centric Review of Plant Disease Datasets: Taxonomy, Critical Analysis, Environmental Variability, and Implications for Precision Agriculture

Aamir Hilal; Shabir Ahmad Sofi; Neeraj Goel

cs.CVOct 5, 2026

A Data-Centric Review of Plant Disease Datasets: Taxonomy, Critical Analysis, Environmental Variability, and Implications for Precision Agriculture

Authors: Aamir Hilal, Shabir Ahmad Sofi, Neeraj Goel

Organizations: Department of Information Technology, National Institute of Technology, Hazratbal Srinagar 190006, Jammu & Kashmir, India · Department of Computer Science and Engineering, Indian Institute of Technology, Ropar Rupnagar 140001, Punjab, India

Abstract

Despite rapid advances in artificial intelligence, reliable real-world plant disease detection remains a persistent challenge. Visual and deep learning approaches have shown promising results, but their deployment under field conditions remains limited. A key bottleneck is the reliance on laboratory-generated datasets that lack environmental diversity, realistic backgrounds, and balanced class distributions, resulting in poor generalization. In contrast, datasets collected directly from agricultural environments capture natural variability and better reflect challenges faced by farmers across regions. This review presents a critical analysis of visual and deep learning approaches for plant disease detection, with emphasis on plant disease datasets. It establishes a taxonomy based on acquisition setting, accessibility, plant diversity, disease composition, class structure, and imbalance severity, and examines their implications for model generalization and real-world deployment. A comparative analysis of laboratory and real-field datasets identifies critical gaps that hinder disease detection. The review further analyzes how multi-level dataset imbalance, including intra-class, inter-crop, and cross-dataset imbalance, and limited environmental variability affect model performance and robustness, an area insufficiently examined in existing surveys. Beyond image-based approaches, it highlights the importance of integrating environmental parameters such as temperature, humidity, and leaf wetness with image data to improve prediction under dynamic field conditions. Finally, the review identifies key challenges, research gaps, and future directions concerning dataset construction, environmental variability, structural imbalance, standardization, and multimodal disease monitoring. It provides a foundation for developing next-generation multimodal frameworks for precision agriculture.

Figures & tables

Figure 1: Taxonomy of deep learning models for plant disease detection structured along two dimensions: learning paradigm and architectural strategy.

Ref	Year	Model Coverage	Architecture Analysis	Dataset Inclusion	Dataset Analysis	Data Diversity	Class Imbalance	Acquisition Setting	Dataset Availability	Env Consideration	Dataset Limitations
[ 50 ]	2025	✓	Moderate	✓	Limited	✗	✗	✗	✗	✗	✗
[ 53 ]	2025	✓	Moderate	✓	Limited	✗	✗	✗	✗	✗	✗
[ 51 ]	2025	✓	Comprehensive	✓	Limited	✗	✗	✗	✗	✗	✗
[ 52 ]	2024	✓	Comprehensive	✓	Limited	✗	✗	✗	✗	✗	✗
[ 9 ]	2024	✓	Moderate	✓	Limited	✗	✗	✗	✗	✗	✗
[ 54 ]	2024	✓	Moderate	✓	Limited	✓	✗	✗	✓	✗	✗

Table 1: Comparison of existing survey papers on plant disease detection highlighting dataset-centric research gaps.

Database	Retrieved Articles
ACM Digital Library	50
Scopus	1654
SpringerLink	1328
IEEE Xplore	1742
Google Scholar	1934
Other sources	76

Table 2: Number of articles retrieved from each database.

Considered articles	Unconsidered articles
Studies on plant disease detection using deep learning techniques.	Studies on general agriculture without focus on disease detection or deep learning applications
Articles that describe or analyze plant disease datasets relevant to deep learning models.	Articles focused on traditional plant pathology methods without using datasets or DL models
Research that includes single-plant or
multi-plant datasets for identifying diseases in crops or plants	Studies that do not provide or analyze datasets or only mention deep learning models in theoretical terms.
Articles discussing image datasets specifically collected for plant disease classification	Research on crop yield prediction or growth monitoring without a focus on disease detection.
Articles discussing the comparison of DL and datasets in plant disease detection across different plant species.	Papers using non-disease-related datasets (e.g., plant growth data, soil composition data) without disease-specific context.

Table 3: Inclusion and Exclusion Criteria for Study Selection.

Figure 2: Overview of the systematic literature search and study selection process.

Ref	Dataset	Year	Acquisition	Availability	No. of Crops	No. of Classes	No. of Images
Single-Plant Datasets
[ 61 ]	ALDD Apple Dataset	2019	Lab	Private	1	5	26,377
[ 62 ]	Soybean Dataset	2019	Lab	Private	1	7	1,470
[ 63 ]	Citrus Dataset	2022	Lab	Private	1	3	2,684
[ 64 ]	Rice Dataset	2022	Lab	Private	1	3	442
[ 65 ]	Apple Plant Dataset	2022	Lab	Private	1	10	5,201

Table 4: Structural overview of plant disease datasets utilized in deep learning–based detection research.

Figure 3: Taxonomy of plant disease datasets based on plant scope.

Figure 4: Taxonomy of plant disease datasets based on acquisition environment.

Figure 5: Taxonomy of plant disease datasets based on dataset accessibility.

Ref	Dataset Name	Access Link
Publicly Available Single-Plant Datasets
[ 69 ]	Rice Dataset	https://archive.ics.uci.edu/ml/datasets/Rice+Leaf+Diseases
[ 70 ]	Soybean Dataset	https://datadryad.org/stash/dataset/doi:10.5061/dryad.41ns1rnj3
[ 45 ]	Beans Dataset	https://github.com/AI-Lab-Makerere/ibean/tree/master
[ 20 ]	Citrus Plant Dataset	https://www.kaggle.com/datasets/drlisbeek/citrus-leaves-prepared
Publicly Available Multi-Plant Datasets

Table 5: Publicly available plant disease datasets with access information.

Dataset	Class of diseases	No. of images	Total Images	Skewness measure
Apple dataset ALDD [ 61 ]	Alternaria Leaf Spot	5343	26,377	0.155
	Apple-Rust	5694
	Mosaic-Virus	4875
	Gray-Spot	4810
	Brown-Spot	5655
Soybean Dataset [ 62 ]	Bacterial	200	1,470	0.294

Table 6: Disease Distribution in Major Single-Plant Datasets

Plant Name	Healthy	Diseased	Total Images (Crop)	% of Dataset	Dominant Class	Skewness Measure
Arjun	220	232	452	11.08%	Diseased	0.052
Mango	170	265	435	10.66%	Diseased	0.358
Bael	0	118	118	2.89%	Diseased	1.00
Guava	277	142	419	10.27%	Healthy	0.487
Chinar	103	120	223	5.46%	Diseased	0.142
Lemon	159	77	236	5.78%	Healthy	0.516

Table 7: Crop distribution statistics for the Multi-Plant Dataset [ 80 ] .

Plant Name	Class of Diseases	No. of Images	Total Images (Crop)	% of Dataset	Dominant Disease Class	Skewness Measure
Apple	Apple Cider Rust	276	3,172	5.84%	Healthy	0.833
	Apple Black Rot	621
	Apple Scab	630
	Apple Healthy Leaf	1645
Peach	Peach Bacterial Spot	2297	2,657	4.89%	Bacterial Spot	0.843
	Peach Healthy Leaf	360

Table 8: Crop distribution statistics for the PlantVillage dataset [ 23 ] .

Plant Name	Class of diseases	No of images	Total Images (Crop)	% of Dataset	Dominant Disease Class	Skewness measure
Apple	Apple Scab Leaf	92	272	10.59%	Apple Scab	0.33
	Apple Rust Leaf	89
	Apple Healthy Leaf	91
Blueberry	Blueberry Healthy Leaf	117	117	4.55%	Healthy Leaf	1.00
	Blueberry Diseased Leaf	0
Cherry	Cherry Healthy Leaf	57	57	2.21%	Healthy Leaf	1.00

Table 9: Crop-wise structural distribution of disease categories in the PlantDoc dataset [ 77 ] .

Plant Name	Class of Disease	No. of Images	Total Images (Crop)	% of Dataset	Dominant Disease Class	Skewness Measure
Apple	Apple Aphis spp	162	1416	31.84%	Venturia inaequalis	0.744
	Apple Venturia inaequalis	633
	Apple Monilia lava	255
	Apple Eriosoma lanigerum	366
Peach	Peach Monilia lava	314	746	16.77%	Parthenolecanium corni	0.264
	Peach Parthenolecanium corni	427

Table 10: Crop-wise structural distribution of disease categories in the Turkey Plant Dataset [ 59 ] .

Plant Name	Class of Disease	No. of Images	Total Images (Crop)	% of Dataset	Dominant Disease Class	Skewness Measure
Corn	Leaf Blight	1655	3,194	37.0%	Leaf Blight	0.996
	Brown Spots	220
	Corn Stripe	164
	Corn Yellowing	493
	Corn Streak	251
	Chlorotic Leaf Spot	26

Table 11: Crop distribution statistics for the FieldPlant Dataset [ 81 ] .

Plant Name	Class of Disease	No. of Images	Total Images (Crop)	% of Dataset	Dominant Disease Class	Skewness Measure
Apple	Apple Leaf Black Spot	81	1202	33.90%	Apple Leaf – Spot	0.760
	Apple Leaf – Mosaic Virus	224
	Apple Leaf – Spot	271
	Apple Stem – Canker	169
	Apple Fruit – Black Rot	65
	Apple Leaf – Healthy	215

Table 12: Crop-wise structural distribution of disease categories in the NZDL Fruit Plant Dataset [ 78 ] .

Plant Name	Class of Diseases	No. of Images	Total Images (Crop)	% of Dataset	Dominant Disease Class	Skewness Measure
Apple	Scab Apple G	241	1,769	6.61%	Healthy Apple	0.871
	Scab Apple S	174
	Healthy Apple	1354
Grape	Leaf-Blight Fungus Grape G	70	3,167	11.80%	Leaf-Blight Fungus S	0.903
	Black Rot Fungus Grape G	435
	Black Measles Fungus Grape G	587

Table 13: Crop distribution statistics for the AI Challenger Dataset [ 79 ] .

Plant	Class of disease	No. of images	Total Images (Crop)	% of Dataset	Dominant Disease Class	Skewness Measure
Rice	Bacterial Leaf Blight	40	120	60%	Balanced	0.00
	Brown Spot	40
	Leaf Blast	40
Wheat	Leaf Rust	40	80	40%	Balanced	0.00
	Powdery Mildew	40

Table 14: Crop-level distribution statistics for the UCI Rice and Wheat dataset [ 41 ] .

Plant	Class of disease	No. of images	Total Images (Crop)	% of Dataset	Dominant Disease Class	Skewness Measure
Rice	Bacterial Leaf Blight	438	2710	73.05%	Brown Spot	0.041
	Brown Spot	466
	Healthy	464
	Leaf Blast	454
	Leaf Scald	448
	Narrow Brown Spot	440

Table 15: Crop-level distribution statistics for PlantDet dataset (Rice and Betel Leaf dataset) [ 82 ] .

Dataset	Total Crops	Total Images	Avg. Crop-Level NIR	Inter-Crop Variance	Variance Level	Acquisition
PlantVillage	14	54,306	0.77	0.97	High	Lab
PlantDoc	13	2,569	0.68	0.92	High	Mixed
AI Challenger	10	26,747	0.77	0.92	High	Lab
NZDL Fruit Dataset	5	3,545	0.53	0.63	Moderate	Field
Turkey Plant Dataset	7	4,447	0.79	0.95	High	Field
FieldPlant Dataset	3	8,629	0.99	0.18	Moderate	Field

Table 16: Cross-dataset structural imbalance comparison across representative multi-plant plant disease datasets.

Figure 6: Structural contrast between laboratory and field datasets and their influence on imbalance severity (NIR), learning complexity, and reported performance inflation.

Figure 7: Conceptual pathway illustrating how structural dataset imbalance propagates into model optimization bias, representation skewness, evaluation distortion, and reduced generalization performance.

Imbalance Type	Impact on Model Learning	Potential Risk
Intra-Class Imbalance	Majority classes dominate gradient updates and decision boundaries	Low recall for minority diseases; biased classification
Inter-Crop Imbalance	Species with higher sample volume shape learned representations	Reduced cross-species generalization
Missing or Underrepresented Healthy Classes	Distorted separation between healthy and diseased states	Overconfident predictions in real-world deployment
Long-Tail Disease Distribution	Rare diseases insufficiently encoded in feature space	Poor detection of uncommon or emerging diseases

Table 17: Model-level implications of structural imbalance in plant disease datasets.

NIR Range	Structural Interpretation	Typical Model Behavior	Deployment Risk
0.0 – 0.3	Relatively balanced class distribution	Stable training; improved per-class recall; reduced need for heavy augmentation	Lower risk of minority-class suppression; better generalization potential
0.3 – 0.7	Moderate imbalance across disease categories	Bias toward majority classes; reliance on augmentation or class-weighting; possible overfitting	Reduced reliability for underrepresented diseases in field conditions
0.7 – 1.0	Severe skewness; dominance of few classes or absence of certain categories	Majority-class dominance; inflated overall accuracy; poor minority recall; architectural compensation (ensemble/hybrid models)	High risk of failure in real-field deployment, especially for rare but critical diseases

Table 18: Relationship between normalized imbalance ratio (NIR) and typical model learning behavior in plant disease detection datasets.

Ref	Dataset(s) Name	Learning Model	Data Augmentation Used	Model Accuracy (%)
[ 88 ]	PlantVillage Dataset	AlexNet	Yes	86.5%
[ 90 ]	PlantVillage Dataset	Comparison between VGG16, MobileNetV2, InceptionV3, Xception and DenseNet121 (Modified Transfer Learning)	Yes	DenseNet121 achieved the highest accuracy (99.9%)
[ 91 ]	Proprietary Dataset (Mango Dataset collected from different fileds in Bangladesh) [ 92 ]	Comparison between VGG19, InceptionV3, ResNet152V2, DenseNet121, InceptionResNetV2, MobileNetV2 and Xception	Yes	InceptionV3 achieved the highest accuracy (99.87%)
[ 93 ]	PlantVillage Dataset	CNN Models - VGG16 and Xception	Yes	VGG with 98.18%
[ 10 ]	PlantVillage Dataset	AlexNet	NA	95.75%
[ 74 ]	Proprietary Dataset (Images collected from Myanmar Grapevine Yard)	VGGNet 16	Yes	98.4 %

Table 19 : Single deep learning frameworks applied to plant disease datasets.

Ref	Dataset(s) Name	Learning Model	Data Augmentation Used	Model Accuracy (%)
[ 109 ]	PlantVillage Dataset	Ensemble of CNN (VGG-16, ResNet-50, DenseNet-121, Xception) and ViT (MobileViT, MViTv2, DeiT3, MaxViT)	Yes	Ensemble of DeiT3 and MaxViT achieved accuracy of 99.83%
[ 110 ]	Kaggle Wheat Disease Datasets	DenseNet201, MobileNetV2, Custom CNN (Ensemble - Bagging based Multi stage CNN)	No	99.16%
[ 111 ]	PlantVillage Dataset	VGG16, VGG19, InceptionV3, ResNet101V2 (Ensemble Voting)	Yes	92.3%
[ 107 ]	PlantVillage Dataset	United Ensemble Model - GoogLeNet and ResNet	Yes	98.57%
[ 112 ]	PlantVillage Dataset	EfcientNetB0 and DenseNet 121 in ensemble mode.	No	98.56%
[ 42 ]	Proprietary Dataset (The Turkey-Plant Dataset – images collected from the agricultural fields of the Bingol and Inonu Universities in Turkey)	AlexNet, GoogleNet, ResNet18, ResNet50, ResNet10 and DenseNet201.	No	97.56%

Table 20: Ensemble deep learning frameworks applied to plant disease datasets.

Ref	Dataset(s) Name	Learning Model	Data Augmentation Used	Model Accuracy (%)
[ 114 ]	PlantVillage Dataset (Apple, Corn)	VGG16 + Inception-V3 + DenseNet201 + ViT (Hybrid (CNN Ensemble + Vision Transformer))	Yes	99.24% (Apple), 98% (Corn)
[ 72 ]	Proprietary Dataset (Images collected from fields)	Four CNN models VGG16, SSD, Faster R-CNN, and Yolov4 were used directly, with Yolov4 employed in a hybrid approach.	Yes	96.47%
[ 83 ]	Proprietary Dataset (Images captured using multispectral camera)	CNN (ResNet) with Vision Transformers	Yes	88.88%
[ 115 ]	Proprietary Dataset (Coffee [ 103 ] )	SUNet (SegNet encoder + U-Net decoder + VGG16 backbone + Mask R-CNN + PSPNet)	Yes	98.45%
[ 116 ]	PlantVillage Dataset	AlexNet and GoogleNet	NA	97.62 %
[ 104 ]	PlantVillage Dataset	VGGNet 19 with EfficientNetB0, MobileNetV2	Yes	96.86 %

Table 21 : Hybrid deep learning frameworks applied to plant disease datasets.

Ref	Dataset Name	Learning Model	Data Augmentation Used	Model Accuracy (%)
[ 76 ]	Proprietary Dataset (Images collected from a vineyard scanned with lab-scale HSI system in 900–1700 nm spectral range)	Partial least squares discriminant analysis (PLS-DA)	No	97.17%
[ 124 ]	Combined Dataset (PlantVillage, VegNet, Bangladeshi Crop Disease, Eggplant Dataset [ 125 ] )	PlantCareNet (5 convolution layers, 3 pooling layers, followed by fully connected layers before the final Softmax classifier.)	Yes	97%
[ 122 ]	PlantVillage Dataset	Custom CNN with 3 convolution and 3 max-pooling layers followed by 2 fully connected layer	Yes	91.2%
[ 123 ]	PlantVillage Dataset	Custom CNN incorporates several convolutional layers, with a pooling, activation, and fully connected layer.	No	87.47% to 99.25%
[ 126 ]	PlantVillage Dataset	Custom lightweight 6-layer CNN	No	98.4%
[ 64 ]	Proprietary Dataset (Images collected from a farmland)	Custom CNN model designed with two shallow convolutional layers followed by 3 Fully Connected layers, along with batch normalization and max-pooling layers.	Yes	93.25%

Table 22: Custom deep learning frameworks applied to plant disease datasets.

Panel of Figure 8 — Figure 8: Heatmap-style summary of model usage across different dataset characteristics in plant disease detection research. Darker shades indicate higher frequency of usage.

Disease Type	Affected Plant Part/Appearance	Favorable Environmental Conditions
Apple Scab	Leaf and Fruit	Temperature on average: 20°C
Apple Powdery Mildew	Leaf and Fruit	Temperature ranging from 10°-25° C
Apple Marssonina Leaf Blotch	Leaf	High rainfall and temperatures between 20° and 22° C
Apple Alternaria Leaf Spot	Leaf	Temperature ranging from 25°-30° C
Apple Black Rot	Leaf and Fruit	20°-22° C Temperature range and Moisture
Apple Sooty Blotch	Fruit	Temperature range 18°-27° C

Table 23: Apple Diseases Due to Temperature [ 133 ] .

Disease Type	Affected Plant Part/Appearance	Favorable Environmental Conditions
Apple Powdery Mildew	Leaf and Fruit	Humidity greater than 70%
Apple Collar Rot	Root	Over-watering and Moisture
Apple Seedling Blight	Root	High Humidity
Apple Core Rot	Fruit	High humidity and warm temperature
Apple Brown Rot	Fruit	Wet humid conditions

Table 24: Apple Diseases Due to Humidity [ 134 ] .

Disease Type	Affected Plant Part/Appearance	Favorable Environmental Conditions
Apple Scab	Leaf	Light infection: 9 hours at 18°C
Apple Alternaria Leaf Spot	Leaf	More than 6 hours of LWD

Table 25: Apple Diseases Due to Leaf Wetness Duration [ 135 ] .

Challenge	Implication for Model Development and Deployment
Lack of synchronized environmental metadata	Prevents context-aware learning and limits multimodal benchmarking consistency.
Sensor heterogeneity across studies	Reduces reproducibility and complicates cross-dataset comparison.
Environmental variability across regions	Introduces domain shift and weakens model generalization in field deployment.
Absence of standardized multimodal benchmarks	Hinders fair architectural comparison and systematic performance evaluation.
Limited public availability of multimodal datasets	Restricts collaborative research and slows development of field-ready systems.

Table 26: Structural challenges limiting environmental integration in plant disease datasets.

Open Challenge	Detailed Description and Implications
Scarcity of Large-Scale Real-Field Datasets	Although numerous plant disease datasets have been introduced, a substantial proportion remains laboratory-generated, characterized by uniform backgrounds, controlled illumination, and limited environmental noise. Real-field datasets are often smaller in size, geographically constrained, crop-specific, and frequently proprietary. This structural dependency on controlled datasets inflates experimental performance while limiting ecological validity and real-world generalization. The development of large-scale, geographically diverse, publicly accessible field datasets remains a pressing research priority.
Multi-Level Class Imbalance and Data Skewness	Class imbalance persists at multiple levels, including intra-class imbalance within disease categories, inter-crop imbalance across plant species, and cross-dataset imbalance due to inconsistent dataset design. Severe skewness biases model learning toward dominant classes, suppresses minority disease detection, and produces misleading overall accuracy metrics. While augmentation and resampling strategies are commonly applied, they do not fundamentally resolve structural imbalance inherent in dataset construction. More principled imbalance-aware learning and evaluation strategies remain underexplored in agricultural AI research.
Lack of Standardized Benchmarking Protocols	The absence of unified benchmarking frameworks complicates fair comparison across studies. Variability in train–test splits, preprocessing pipelines, augmentation policies, and performance metrics limits reproducibility. Additionally, architectural diversity across single, ensemble, hybrid, and custom models further hinders standardized evaluation. Without common validation protocols and cross-dataset testing strategies, reported improvements may reflect experimental configuration rather than genuine methodological advancement.
Limited Cross-Dataset Generalization Studies	Most existing research evaluates models on the same dataset used for training. Cross-dataset validation—training on one dataset and testing on another is rarely performed. This limits understanding of domain shift effects, acquisition variability, crop morphological differences, and environmental transferability. Robust real-world deployment requires models that generalize across datasets, regions, and acquisition conditions, yet systematic cross-dataset evaluation remains largely unexplored.
Environmental Context Integration Constraints	Environmental factors such as temperature, humidity, leaf wetness, and soil conditions significantly influence disease progression. However, most detection frameworks rely exclusively on visual inputs. Multimodal datasets combining image data with synchronized environmental measurements remain scarce. Key challenges include temporal alignment between sensor data and image capture, sensor calibration variability, missing environmental readings, and increased fusion complexity. Establishing standardized multimodal benchmarks is critical for advancing environmentally aware disease prediction systems.
Annotation Quality and Label Noise	Accurate annotation of plant disease images requires domain expertise. Field images may contain mixed symptoms, early-stage infections with subtle visual cues, or co-existing diseases on the same leaf. Inconsistent labeling protocols introduce annotation noise that negatively impacts model learning, generalization, and reproducibility. Establishing standardized annotation guidelines and expert-verified labeling pipelines is essential for improving dataset reliability.

Table 27: Open challenges and emerging issues in real-world deep learning–based plant disease detection.

Explore similar work

May 10, 2026cs.MA

SAGE: Scalable Agentic Grounded Evaluation for Crop Disease Diagnosis

Plant disease diagnosis is critical for food security, yet training disease-recognition models that generalize across crops, pathogens, and field conditions remains challenging because labeled disease images are far less abundant and standardized than data for other biotic stresses such as insects or weeds. Frontier vision-language models offer new opportunities through improved visual reasoning, but they still struggle with fine-grained disease identification due to the lack of structured, crop-specific symptom knowledge. To address this gap, we curate the largest plant disease image--symptom dataset to date, covering 335 crops, 1{,}251 disease classes, and approximately 839K images, designed to support training-free, agentic disease prediction. A scalable automated pipeline generates source-grounded symptom descriptions in which each claim is linked to a verbatim web quote; domain experts validate sampled crops and reconcile disease-name variants across sources. As a baseline, we introduce an autonomous visual reasoning agent that identifies anatomical context, narrows candidate diseases using symptom knowledge, sequentially compares reference images, and produces a fully explainable reasoning trace. Incorporating symptom knowledge improves accuracy by 16.2 percentage points on average at the full reference budget, with consistent gains across all four evaluation crops. Because the framework only requires crop-specific reference images and symptom knowledge, it can be extended to new crops without retraining, while the agentic baseline can directly benefit from future improvements in foundation model capabilities. Dataset and code are available at:https://sage-dataset.github.io/.

Muhammad Arbab Arshad, Tirtho Roy, Yanben Shen +7

Iowa State University, USA · New York University, USA

Leaf Disease Classification Disease Diagnosis Recent Vision-Language Models Readmission Prediction

Ask PDF

May 5, 2026cs.CV

ReLeaf: Benchmarking Leaf Segmentation across Domains and Species

Rising global food demand and growing climate pressure increase the need for sustainable, precise agricultural practices. Automated, individualized plant treatment relies on fine-grained visual analysis, yet leaf-level segmentation remains underexplored despite its value for assessing crop health, growth dynamics, yield potential and localized stress symptoms. Progress is limited by a lack of dedicated datasets, especially regarding species coverage, and by the absence of systematic evaluations of modern instance-segmentation architectures for this task. We address these gaps by surveying current data and identifying four suitable, publicly available leaf-segmentation datasets. Using them, we compare one-stage, two-stage and Transformer-based detectors and identify a YOLO26 model configuration to provide the best trade-off for real-world precision-agriculture tasks. Extensive cross-domain generalization experiments reveal substantial performance drops across plant species and recording setups, especially for models trained solely on laboratory data. To strengthen data availability, we introduce a new benchmark dataset with leaf-level masks for 23 plant species, created via semi-automatic annotation of selected CropAndWeed images. A model trained on all four existing datasets achieves a mean mAP50-95 of 83.9% across their corresponding test sets and 40.2% on our new benchmark, demonstrating improved generalization and highlighting the need for diverse leaf-segmentation datasets in robust precision agriculture.

Robert Martinko, Daniel Steininger, Julia Simon +2

1AIT Austrian Institute of Technology, Center for Vision, Automation & Control · University of Applied Sciences Technikum Wien, Computer Science & Applied Mathematics

Leaf Disease Classification Precision Agriculture Vision Datasets Clover Species Treatment

Ask PDF

Oct 5, 2026cs.CV

Environmental sensor readings in two crop disease image datasets identify the session in which each image was taken

Integrating environmental sensor data with leaf imagery is widely reported to boost crop disease classification accuracy. In this work, we reveal that these reported gains are often artifacts of dataset construction: because a single sensor reading is shared across many images collected in a single session (one farm on one date), multimodal networks can predict disease simply by memorizing session identities. Analyzing two widely used Korean datasets, the Crop Disease Diagnosis (CDD) benchmark and an AI Hub pest/disease dataset, we demonstrate that nearly all images share sensor values, with 91.9% of CDD test images having exact sensor duplicates in the training set. Remarkably, an image-free classifier given only timestamps matches or exceeds sensor-driven predictions across all seven evaluated crops, and matches the published macro-F1 of a state-of-the-art CDD fusion model. These results indicate that performance gains on standard random splits cannot be disentangled from session leakage. We propose that multimodal crop studies must evaluate on session-held-out splits and report performance against sensor-free date-time baselines to ensure genuine generalization.

Sungwoo Kang

Department of Electrical and Computer Engineering, Korea University Seoul, Republic of Korea

Leaf Disease Classification Multi-Modal Sensing

Ask PDF