Phenotypic image dataset of naturally grown shiitake mushrooms across multiple varieties and growth stages

Abstract

As a valuable edible fungus with both culinary and medicinal value, shiitake mushrooms hold significant economic and social importance in promoting agricultural restructuring, increasing farmers’ income, and advancing the intelligent development of the edible fungus industry. However, in natural growing environments, the classification, grading, and inspection of shiitake mushrooms still rely primarily on manual labor. This approach is inefficient, highly subjective, and lacks standardized data support. Currently, publicly available datasets for edible fungi are limited in quantity, making it challenging to meet the research demands for intelligent detection and automated harvesting models. To address this issue, this study constructed a multi-variety shiitake mushroom image dataset under natural growth conditions, selecting three mainstream market varieties: 9608, Chunsheng No. 1, and Qihe No. 9. The dataset contains 1,782 high-resolution original images, which were expanded to 6,500 images through systematic data augmentation, containing a total of 43,752 precisely annotated mushroom instances. The dataset covers three distinct growth stages (maturity, growing period, and juvenile stage) as well as a dedicated category for deformed mushrooms. The images feature realistic characteristics, including complex lighting, occlusion, dense distribution, and varying viewing angles. This dataset can be widely applied to object detection tasks, including intelligent grading of shiitake mushrooms, visual recognition for harvesting robots, yield estimation, production management, and decision-making for market release. It provides a high-quality data foundation for intelligent production in the edible fungi industry. The dataset is freely available for academic research and technical exchange.

Background & Summary

Edible fungi, as a vital agricultural biological resource, possess exceptional nutritional, medicinal, and economic value, aligning with the health-conscious dietary principle of “one vegetable, one meat, one fungus”¹. In recent years, driven by rapid advancements in agricultural biotechnology, environmental control, and automated cultivation techniques, the edible fungi industry has globally adopted trends toward large-scale, intensive, industrialized, and intelligent development². Among these, shiitake mushrooms, as the world’s second-largest edible mushroom variety, have become China’s most produced and consumed edible mushroom, serving as a key representative driving the modernization and upgrading of the edible mushroom industry³.

However, in traditional cultivation and grading practices, shiitake mushroom production remains highly dependent on manual expertise. Particularly in identifying different varieties and growth stages, grading, and harvesting, issues persist, such as high labor intensity, low efficiency, and inconsistent standards⁴. In their natural growing environment, shiitake mushrooms are significantly influenced by factors such as light exposure, humidity, temperature, and ventilation. Their morphology, coloration, and cap shape exhibit complex variations, while subtle differences in form and appearance exist between different varieties. These characteristics make manual inspection highly susceptible to subjective factors, rendering it challenging to meet the precision and consistency requirements of large-scale production⁵. Against this backdrop, research on intelligent shiitake mushroom detection based on computer vision and deep learning technologies has gradually emerged⁶. By enabling automatic recognition, classification, and grading of shiitake mushroom images, this approach not only significantly enhances automation levels in production processes but also facilitates real-time monitoring and dynamic assessment of mushroom growth conditions, thereby providing technical support for intelligent cultivation management^7,8. Recent breakthroughs in agricultural computer vision have expanded beyond traditional convolutional networks toward specialized geometric and attention-driven architectures. For instance, comprehensive reviews on recent trends emphasize that despite significant progress, multi-class classification and dynamic on-field variations remain persistent challenges in fruit and vegetable sorting⁹. To preserve spatial hierarchies and fine-grained feature boundaries, advanced capsule frameworks have emerged; PotCapsNet introduces an explainable pyramid dilated capsule architecture tailored for early disease visualization¹⁰, while FruCapsNet integrates a shuffled convolutional block attention mechanism to deeply encode multi-fruit quality traits and part-whole morphological variations¹¹. Furthermore, to handle the chronic obstacle of class imbalance in real-field collections, unsupervised generative frameworks like FruitGAN have demonstrated exceptional capability in synthesizing ultra-realistic synthetic training images while maintaining subtle patch lesion integrity¹². While these capsule-based and generative methods provide unmatched precision in localized classification and data balancing, there remains a prominent operational gap for immediate multi-variety spatial localization and simultaneous multi-instance real-time inference under resource-constrained embedded edge scenarios. Consequently, exploring modern real-time object detection models (such as the lightweight YOLO series) validated against systematically structured multi-variety field datasets becomes imperative to bridge this deployment gap. However, publicly available image datasets for shiitake mushrooms remain extremely limited. A major gap in the existing literature is that available datasets predominantly focus on a single mushroom variety or rely on images captured in highly idealized laboratory environments with uniform lighting and clean backgrounds. In reality, such data fails to represent the complex, chaotic visual conditions of agricultural production. In particular, there is a lack of high-quality datasets covering multiple co-cultivated varieties, deformed mushrooms, and various growth stages simultaneously under natural greenhouse environments. This shortage constrains the generalization capabilities of deep learning models, leading to severe performance degradation when deploying algorithms onto harvesting robots or edge-inspection devices in actual fields.

To address these critical limitations and bridge the gap between academic research and practical application, this study constructed a comprehensive dataset specifically engineered for the fine-grained classification and grading of multiple shiitake mushroom varieties in natural growing environments. To contextualize our contribution, it is essential to compare our work against recent prominent open-source mushroom datasets. Recent advancements have introduced heavily structured benchmarks primarily optimized for 3D localization and robotic harvesting. For instance, the Realistic Synthetic Mushroom Scenes Dataset utilizes a generative diffusion pipeline to produce 15,000 highly realistic synthetic images with comprehensive 2D/3D annotations¹³. However, its scenes are largely synthesized around a singular 3D mushroom model structure, which limits its capacity to model subtle phenotypic variations among distinct biological varieties. Similarly, the SMS3D dataset provides an extensive scale of 40,000 synthetic scenes generated from 3D scans of button mushrooms¹⁴. While SMS3D offers exceptional customizable density and oriented 3D bounding boxes to mitigate data scarcity, it inherently relies on simulated environments that may exhibit a domain gap when encountering real-world dynamic greenhouse complexities. To bridge this synthetic-to-real gap, the M18K dataset introduces a multi-purpose real-world benchmark containing over 100,000 instances across 2,000 RGB-D images for object detection and growth monitoring¹⁵. Although M18K captures real-world cultivation environments, its scope is strictly constrained to two specific mainstream commercial button mushroom varieties.

In contrast to these existing frameworks, our constructed dataset establishes a distinct and complementary benchmark focused on fine-grained taxonomic and phenotypic grading under natural production conditions. Unlike previous works that treat mushroom detection as a coarse, binary, or single-class task¹⁶, our approach introduces a structured, hierarchical coding system. Furthermore, the data was gathered across 7 standardized production bases over a cumulative 3-year period, embedding authentic challenges such as dense overlapping, varying viewing angles, and complex illumination directly into the dataset. It encompasses three typical varieties and four distinct phenotypic categories, utilizing high-precision annotation and rigorous quality control procedures to ensure data accuracy and reusability. This dataset not only provides foundational data support for real-time detection, maturity assessment, and yield prediction of shiitake mushrooms, but also offers a standardized, systematic dataset for research.

Methods

The Shiitake Mushroom Dataset constructed in this study was collected from edible mushroom production bases that the research team has used for a long time. It encompasses seven standardized shiitake mushroom production bases, including Beiwudu Town in Wuyang County, Luohe City, Henan Province, China, as well as five large-scale shiitake mushroom cultivation enterprises, such as Yufu Mushroom Industry in Huojia County, Xinxiang City, Henan Province. Data collection was completed based on continuous production practices accumulated over the past three years. All images originate from raw photographic data captured in actual cultivation settings, demonstrating significant authenticity and representativeness. Before formally initiating data collection, the research team conducted thorough consultations with experts in edible mushroom production and technicians at relevant production bases. Through systematic investigation, they gained a comprehensive understanding of the morphological characteristics, visual variations, and developmental patterns across growth stages of different shiitake mushroom varieties. These insights provided a robust theoretical foundation and practical field experience to support the subsequent development of a classification and grading system. Based on the aforementioned preparatory work, this study established a comprehensive technical workflow spanning from raw shiitake mushroom image acquisition and data annotation to deep learning model validation. The overall framework is illustrated in Fig. 1.

Fig. 1.

Workflow for data collection, data annotation, and technical validation.

Data collection

During the data collection process, three mainstream shiitake mushroom varieties in the market, namely 9608, Chunsheng No. 1, and Qihe No. 9, were selected as experimental subjects because these three varieties have high similarity in appearance, such as shape, color, and flower shape, making them difficult for non-professionals to distinguish. This makes them highly targeted and challenging in the classification and grading of multiple shiitake mushroom varieties, as shown in Fig. 2. To ensure data representativeness and diversity, the photography process fully considered the impact of different time periods, multiple lighting conditions, and varied shooting angles on image quality and the visual characteristics of shiitake mushrooms.

Fig. 2.

Original images of three shiitake mushroom varieties: (a) Variety 9608; (b) Variety Chunsheng No. 1; (c) Variety Qihe No. 9.

Image acquisition equipment included the Sony A6400 and Canon EOS R6 cameras, along with mobile devices such as the iPhone 15 Plus and vivo NEX. The images in the dataset were collected under natural cultivation conditions with multiple image resolutions, primarily including 2,016 × 1,512 pixels, 2,312 × 1,736 pixels, and 2,856 × 2,142 pixels. These devices were mounted on a mobile support approximately 40–60 cm above the shiitake mushroom surface to ensure each image fully covers the target area of the mushroom. During data collection, strict control was maintained over the consistency of shooting parameters, while simultaneously recording environmental information, including temperature, humidity, and light intensity. These measurements will facilitate scene modeling and light normalization analysis in subsequent research. After multiple rounds of data collection, a total of 2,600 raw images were obtained. Following screening and quality control, 1,782 high-resolution samples with complete information and clearly annotated targets were ultimately retained for subsequent annotation and modeling tasks.

Referencing international standards and common market classification methods in China, and considering the practical application needs of shiitake mushrooms in production and research, this study categorizes shiitake mushrooms into four phenotypic categories: mature mushrooms, growing mushrooms, juvenile mushrooms, and morphologically deformed mushrooms¹⁷. Mature shiitake mushrooms feature fully expanded caps with distinct umbrella shapes and slightly inward-curling edges. These larger specimens exhibit cap expansion ranging from 60% to 90%. Mature shiitake mushrooms are in the middle stage of growth, with caps not yet fully opened. They are medium-sized, with caps opened to 60% or less, or not yet opened. Juvenile shiitake mushrooms are in the initial growth stage, characterized by smaller bodies that are mostly attached to the surface of the mushroom logs. Deformed mushrooms refer to samples exhibiting obvious phenotypic abnormalities, including severe discoloration, irregular morphology, malformed cap structures, or mechanical growth damage. During annotation, discoloration was identified based on visible deviations in chromatic consistency and surface appearance relative to healthy mushrooms of the same variety under similar acquisition conditions. Typical appearances of different shiitake mushroom categories in natural environments are shown in Fig. 3. No explicit barrel distortion correction or geometric rectification was applied to the wide-angle images. All images were retained in their original captured form to preserve the authenticity of natural cultivation environments.

Fig. 3.

Morphological characteristics of different categories of shiitake mushrooms in natural growth images: (a) Mature shiitake mushrooms; (b) Growing shiitake mushrooms; (c) Juvenile shiitake mushrooms; (d) Deformed shiitake mushrooms; (e) Densely overlapping; (f) Wide-angle view.

Data labeling

Prior to annotation, the collected shiitake mushroom image data underwent systematic screening and preprocessing. To eliminate potential selection bias, the exclusion of raw samples was governed strictly by objective operational criteria rather than subjective visual evaluation. First, images were designated as “unidentifiable” and subsequently excluded only under two specific scenarios: (1) physical labeling failures, where cultivation tags, crop bed markers, or QR codes on the mushroom logs were missing or unreadable due to greenhouse humidity, preventing absolute verification of the biological variety; and (2) record synchronization errors, where manual field photography logs could not be reliably cross-referenced with image file timestamps. No image was discarded based on visual quality, target complexity, or mushroom appearance. Through this rigorous traceability check, samples lacking verifiable origin were excluded to ensure data accuracy and traceability, while preserving the full range of natural environmental challenges across the remaining dataset. Second, crop and convert the format of selected images to standardize data dimensions and file formats. Finally, batch-remove the “EXIF” metadata from images to ensure consistency and reproducibility between training data and annotation files.

Data annotation was performed using the widely adopted object detection annotation tool LabelImg¹⁸. The specific workflow included: (1) Manually identifying each shiitake mushroom instance in the image to determine its variety and growth stage category; (2) Precisely annotating each mushroom by tightly enclosing its outline with a bounding box and assigning corresponding labels; (3) Saving the generated.txt format label files corresponding to each image. (4) Complete annotation for all images sequentially. (5) Finally, all annotations underwent three rounds of manual review conducted by trained annotators and domain specialists to ensure labeling accuracy, consistency, and completeness. To further clarify the dataset’s complexity regarding spatial overlapping, we quantified the proportion and severity of occluded mushroom instances. Among the annotated instances, approximately 32.4% exhibit varying degrees of occlusion. Specifically, these occluded targets are categorized into three severity levels based on the obscured area of the mushroom cap: light occlusion (<20% obscured area) accounts for 18.2%, moderate occlusion (20%–50% obscured area) accounts for 10.7%, and heavy occlusion (>50% obscured area) accounts for 3.5% of the total dataset. To ensure annotation completeness for overlapping targets, the bounding boxes for partially occluded instances were drawn to encompass the estimated full physical boundaries of the mushroom caps rather than being restricted to their visible fragments. This approach ensures that the object detection models receive consistent scale and geometric cues during training.

Label naming adheres to a design principle that combines systematicity and conciseness, establishing a structured coding system illustrated in Fig. 4, with its multi-class annotation performance visually demonstrated in Fig. 5. Specifically, Fig. 4 deconstructs each label into three logical components: the capital letter “L” (Lentinus), a three-digit variety identifier, and a two-digit growth stage marker; meanwhile, Fig. 5 provides a practical LabelImg interface example, showcasing a dense distribution of localized bounding boxes paired alongside color-coded tags for co-existing stages. Taking the 9608 variety as an example, mature, growing, juvenile, and deformed shiitake mushrooms are respectively labeled as L00101 through L00104. Here, “L” denotes shiitake mushroom, “001” represents the variety code, and the last two digits indicate the growth stage category. The other two varieties follow the same naming convention. Prior to batch annotation, small-sample pilot annotation, model pilot training, and inference validation were conducted to establish clear annotation standards and quality requirements.

Fig. 4.

Structured tag naming conventions.

Fig. 5.

Example of a label for shiitake mushroom variety 9608.

Data augmentation

Data augmentation plays a crucial role in training object detection models and is a key step in the process. In this study, data augmentation methods were employed to expand the dataset from 1,782 original images to 6,500 augmented images, thereby improving sample diversity and alleviating limitations related to illumination variability, viewpoint changes, environmental complexity, and class imbalance in natural cultivation environments. The augmentation strategy increased sample diversity and improved robustness under varying illumination conditions, dense occlusion, and complex natural cultivation environments, as illustrated in Fig. 6. The data augmentation pipeline included random rotation, vertical flipping, brightness adjustment, saturation adjustment, and random noise injection. Random rotation was employed to simulate variations in image orientation and object arrangement during image acquisition. Vertical flipping enhanced the model’s robustness to spatial orientation changes. Random adjustments to image brightness and saturation were applied to improve the model’s adaptability to lighting changes, compensating for the complex and variable lighting conditions within the shiitake mushroom cultivation greenhouses during original image data collection. To address potential sensor interference or environmental noise during image acquisition, random noise was added to enhance the model’s robustness against disturbances. The overall data augmentation workflow is summarized in Algorithm 1.

$$I\prime =\alpha I+\beta$$ 1

where $I$ and $I\prime$ represent the original and transformed images, respectively, $\alpha$ denotes the contrast adjustment coefficient, and $\beta$ represents the brightness offset parameter.

Fig. 6.

Various types of data augmentation effects: (a) Original image; (b) Random rotation; (c) Vertical flipping; (d) Brightness adjustment; (e) Random noise; (f) Saturation adjustment.

Algorithm 1

Pseudocode for cleaning data from abnormal values.

Through a systematic data augmentation strategy, the training dataset was significantly expanded, yielding a total of 6,500 augmented images. Meanwhile, the number of annotated shiitake mushroom instances increased to 43,752, encompassing diverse phenotypic characteristics across different varieties and phenotypic categories. The distribution of instances across varieties and phenotypic categories is shown in Fig. 7.

Fig. 7.

Instance distribution map.

Object detection model

This study focuses on intelligent vision technology to address the classification and grading detection of diverse shiitake mushroom varieties, aiming to enhance the intelligence and automation of shiitake cultivation processes. The task requires the detection model to adapt to variations in posture, occlusion, and lighting conditions under natural growth environments while maintaining cross-platform deployment capabilities. Desktop deployment focuses on leveraging the computational power and storage advantages of computing devices, prioritizing detection accuracy and stability¹⁹. Mobile deployment, however, must balance model lightweighting with real-time responsiveness to accommodate the resource constraints inherent in embedded and mobile devices²⁰.

Among current mainstream object detection models, approaches such as Faster R-CNN²¹ and SSD²² exhibit high detection accuracy but suffer from slow inference speeds and high resource consumption, making them unsuitable for real-time detection requirements at the edge²³. In contrast, the YOLO²⁴ series of models strikes a good balance between detection speed and accuracy, demonstrating powerful performance in network architecture optimization, feature extraction efficiency, and multi-scale perception capabilities. Considering factors such as model size, detection performance, and multi-scale adaptability, this study selected mainstream lightweight detection models, including YOLOv8n²⁵, YOLOv9t²⁶, YOLOv10n²⁷, and YOLO11n²⁸ as core frameworks for training and testing. This approach systematically validates the applicability and effectiveness of the constructed shiitake mushroom image dataset for object detection tasks. These models offer advantages such as compact size, fast inference speed, and low resource consumption, while delivering both high accuracy and real-time performance. They are suitable for deployment across multiple scenarios on both desktop and mobile platforms.

Experimental setup

Model training employs a stochastic gradient descent optimizer. The initial learning rate was adjusted through multiple trials to 0.01, with a weight decay coefficient of 0.005. Three warm-up phases were established during the early training stage, with the momentum parameter gradually adjusted from 0.8 to 0.937 to achieve stable convergence. The training runs a total of 400 epochs with a batch size of 8. Regarding data partitioning, the 5,375 annotated images are divided into a training set and a validation set in an 8:2 ratio for model training and performance evaluation. All quantitative evaluation metrics (such as Precision, Recall, F1-score, and mAP) reported in this study were strictly calculated using this fully annotated validation set. Additionally, a separate group of unlabeled raw images was reserved exclusively as an independent inference set to visually demonstrate the model’s practical detection performance and generalization capacity on entirely unseen data during the inference phase. For consistency across benchmark evaluations, all models were trained under the same experimental configuration without using pre-trained weights. All experiments were conducted on a workstation equipped with an NVIDIA RTX 4090 GPU, Intel Core i7 processor, and 24 GB RAM using the PyTorch deep learning framework under CUDA acceleration.

Evaluation indicators

To accurately evaluate model performance, the metrics employed in this study include accuracy, recall, F1 score, average precision (AP), and mean average precision (mAP). To ensure the accuracy of real-time detection for multi-variety shiitake mushroom classification and grading in practical applications, the IoU threshold for mAP evaluation is set to 0.5, enabling a comprehensive assessment of model performance. The calculation formulas are as follows:

$$F_1=\frac{2\times \mathrm{Precision}\times \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}$$ 2

$$mAP=\frac{1}{N}\sum _{i=1}^{N}A{P}_i$$ 3

$$AP={\int }_0^{1}P\left(R\right)dR$$ 4

Data Records

The complete phenotypic image dataset and validation source code described in this work have been permanently deposited in the Zenodo repository under https://doi.org/10.5281/zenodo.20373531²⁹. The shiitake mushroom image dataset used in this study was precisely annotated using LabelImg software, generating annotation files compliant with the YOLO format (.txt). Each annotation file records target instance information relative to the image dimensions, including parameters such as class ID (cls), center coordinates (x, y), and target width and height (width, height). These parameters ensure annotation accuracy remains unaffected during image scaling or resolution changes. This annotation format has become the de facto standard in the field of object detection, offering excellent algorithm compatibility and scalability. It can be used for model performance evaluation and comparative studies of different detection algorithms³⁰. The processed dataset is organized into two separate directories based on file type: the images folder stores all raw images, while the labels folder contains the corresponding annotation files. Each image is uniquely paired with its corresponding label file through its filename, ensuring structural and content consistency between data and annotations. The overall directory structure of the dataset is illustrated in Fig. 8.

Fig. 8.

Dataset directory structure.

Technical Validation

To evaluate the effectiveness of the proposed data augmentation strategy, comparative experiments were conducted on the YOLO11n model using datasets before and after augmentation under identical training conditions. As shown in Table 1, data augmentation significantly improved detection performance, with Precision increasing from 0.717 to 0.947, Recall from 0.682 to 0.929, mAP50 from 0.717 to 0.961, and mAP50–95 from 0.588 to 0.844.

Table 1.

Performance comparison of YOLO11n trained with and without data augmentation.

Training Strategy	Precision	Recall	mAP50	mAP50-95
Without augmentation	0.717	0.682	0.717	0.588
With augmentation	0.947	0.929	0.961	0.844

Meanwhile, the convergence curves in Fig. 9. demonstrate that the augmented dataset enabled faster and more stable training convergence. The model trained with augmentation achieved lower loss values and consistently higher mAP50 throughout the training process, indicating improved robustness and generalization ability under complex cultivation environments. Based on these results, the enhanced dataset was subsequently employed to train and validate the mainstream object detection models used in this study.

Fig. 9.

Training and validation curves of YOLO11n using augmented and non-augmented datasets: (a) training and validation loss curves; (b) mAP50 convergence curves.

We employed the enhanced image dataset to train and validate current mainstream deep learning models for object detection. The training and validation sets were allocated at an 8:2 ratio, while the unlabeled inference set was utilized afterward solely for qualitative verification, ensuring the reliability of the models’ visual effectiveness and generalization capabilities when deployed in real-world scenarios. We repeatedly split the training and validation datasets at a fixed ratio. Each model underwent three training iterations, with the reported results representing the average performance metrics from multiple validation evaluations to capture variability and the model’s actual capacity. Our objective is to conduct adaptive testing to enhance the reliability of models’ effectiveness and generalization capabilities in real-world scenarios.

To verify the effectiveness of the constructed mushroom image dataset in the object detection task, this study selected four representative lightweight detection models, YOLOv8n, YOLOv9t, YOLOv10n, and YOLO11n, as benchmark models and evaluated their performance under the same training strategy. Table 2 presents the overall key metrics for each model on this dataset, including Precision, Recall, F1-score, mAP50, and mAP50–95, providing objective quantitative evidence regarding the dataset’s usability and task difficulty.

Table 2.

Performance comparison of different YOLO nano models.

Model	Params (M)	Precision (overall)	Recall (overall)	F1-score (overall)	mAP50 (overall)	mAP50-95 (overall)
YOLOv8n	3.01	0.924	0.853	0.887	0.936	0.791
YOLOv9t	1.97	0.877	0.817	0.850	0.904	0.755
YOLOv10n	2.70	0.945	0.878	0.910	0.945	0.808
YOLO11n	2.58	0.947	0.929	0.938	0.961	0.844

The comprehensive evaluation results demonstrate that all models achieved high detection performance on this dataset, indicating that the annotation quality across different species, growth stages, and morphological categories within the dataset is high, and that the models can effectively learn visual differences. Among these, YOLO11n achieved the best overall performance with 2.58 million parameters, attaining overall Precision, Recall, and F1-score of 0.947, 0.929, and 0.938, respectively. mAP50 and mAP50–95 further improved to 0.961 and 0.844, demonstrating a superior balance between lightweight architecture and detection accuracy. In comparison, the YOLOv8n and YOLOv9t models exhibit slightly lower overall performance, particularly showing a noticeable gap in the more stringent fine-grained evaluation metrics mAP50–95. These results indicate that this dataset poses greater challenges to models in recognizing multi-scale, multi-morphology mushroom instances.

Figure 10 illustrates the evolution of key performance metrics across four benchmark models during training. Precision, Recall, mAP50, and mAP50–95 all demonstrate gradual improvement with increasing training iterations before eventually stabilizing, indicating the models effectively learned category features within the dataset. Overall, YOLO11n maintains higher curve levels across all four metrics, demonstrating superior convergence speed and final performance compared to other models. YOLOv8n, YOLOv9t, and YOLOv10n exhibit relatively similar convergence trajectories, reflecting the dataset’s strong adaptability to different models. The consistent trend across all four curves indicates stable annotation quality and well-designed tasks within the dataset, capable of supporting training and validation for models with diverse architectures.

Fig. 10.

Model performance metrics.

To further validate the effectiveness and generalization ability of the proposed dataset, we conducted additional benchmark experiments using representative detectors from different architectures, including Faster R-CNN³¹, RT-DETR³², YOLO11s, and YOLO11n. The comparison results are shown in Table 3. Among all evaluated models, YOLO11s achieved the best overall detection performance, with a precision of 0.951, a recall of 0.925, an F1 score of 0.938, and an mAP50 of 0.965. Meanwhile, YOLO11n also achieved highly competitive results, with a model size of only 2.58 million parameters, an F1 score of 0.936, and an mAP50 of 0.961. In contrast, Faster R-CNN and RT-DETR require over 30 million parameters but have lower overall performance.

Table 3.

Performance comparison of representative object detection models on the proposed shiitake mushroom dataset.

Model	Params (M)	Precision (overall)	Recall (overall)	F1-score (overall)	mAP50 (overall)	mAP50-95 (overall)
Faster R-CNN	35.41	0.931	0.883	0.906	0.932	0.823
RT-DETR	32.01	0.938	0.900	0.918	0.943	0.831
YOLO11s	9.46	0.951	0.925	0.938	0.965	0.857
YOLO11n	2.58	0.944	0.927	0.936	0.961	0.844

As shown in Fig. 11, YOLO11n achieves a good balance between detection accuracy and lightweight design, while YOLO11s achieves the highest detection accuracy among all comparison models. These results demonstrate that the proposed dataset can effectively support high-performance and resource-efficient target detection models, making it suitable for practical deployment in complex shiitake mushroom cultivation environments.

Fig. 11.

Model performance comparison (bubble size represents Params (M)).

Discussion

The proposed dataset introduces substantial visual complexity compared with existing mushroom datasets, particularly in terms of illumination variability, overlapping growth patterns, dense occlusions, and morphological diversity across varieties and phenotypic categories. These characteristics better reflect real greenhouse cultivation environments and increase the difficulty of object detection tasks. The experimental results indicate that the constructed dataset can effectively distinguish the performance differences among different detection architectures. Lightweight YOLO-based models exhibited strong adaptability and stable convergence behavior, while comparisons with Faster R-CNN and RT-DETR demonstrated the dataset’s capability to support both CNN-based and transformer-based detectors. In particular, the trade-off analysis between model size and detection accuracy highlights the feasibility of deploying lightweight models in resource-constrained agricultural scenarios.

In addition, the introduced augmentation pipeline enhanced the diversity of training samples and reduced the influence of environmental disturbances during model training. The improved convergence stability and detection metrics suggest that the dataset is suitable for robust model evaluation under practical cultivation conditions. However, the current study mainly focuses on two-dimensional object detection and category recognition. More detailed annotations, such as instance segmentation masks, temporal growth information, and three-dimensional structural features, are not yet included. Future work will therefore extend the dataset toward fine-grained phenotypic analysis and multi-modal agricultural perception tasks.

Conclusion

In this study, we constructed a large-scale multi-variety shiitake mushroom image dataset collected from real greenhouse cultivation environments across multiple production bases and growth cycles. The dataset contains diverse visual characteristics, including different varieties, growth states, illumination conditions, occlusions, and complex background interference, providing realistic support for intelligent mushroom detection and grading tasks.

A standardized annotation framework and systematic data augmentation pipeline were established to improve dataset consistency, diversity, and robustness. Benchmark experiments on multiple mainstream object detection models demonstrated that the proposed dataset can effectively support both high-accuracy and lightweight detection networks. Among the evaluated models, YOLO11n achieved an optimal balance between detection accuracy and model complexity, while YOLO11s obtained the highest overall detection performance. Overall, the proposed dataset provides a reliable benchmark resource for mushroom detection, classification, and intelligent cultivation research under natural environments. It also offers practical value for future applications in automated harvesting, intelligent monitoring, and edge-device deployment in modern mushroom production.

Author contributions

X.W. and Y.Z. conceived the study. X.W., Y.Z., H.C., and Y.H. collected and curated the dataset. Y.Z. and Y.H. performed data annotation and organization. Y.T., X.Wan., and B.C. contributed to data validation and quality control. C.Z. and G.Z. supervised the project and provided critical revisions to the manuscript. X.W. and Y.Z. wrote the main manuscript text. All authors reviewed and approved the final manuscript.

Funding

This research was supported by the Major Science and Technology Special Project of Henan Province (No. 241100110200), the Science and Technology Project of Henan Province (252102111172, 242102110331, 252102111070, 242102111079 and 262102111107), the Training Plan for Young Backbone Teachers in Henan Province (2023GGJS117) and Henan Province Postgraduate Education Reform and Quality Improvement Project (Case Study) YJS2025AL69.

Data availability

The phenotypic image dataset of naturally grown shiitake mushrooms collected during the current study is permanently archived and publicly available in the Zenodo repository under https://doi.org/10.5281/zenodo.20373531.

Code availability

The implementation used in this study is based on the open-source Ultralytics YOLO project (https://github.com/ultralytics/ultralytics). Several training-related hyperparameters, including the number of epochs, input image resolution, and confidence threshold, were adjusted to adapt the model to the constructed dataset. The scripts and configuration files used for model training and inference are available through the Zenodo repository (https://doi.org/10.5281/zenodo.20373531), enabling reproduction of the reported experiments and evaluation procedures.

Competing interests

The authors declare no competing interests.