BUTTON MUSHROOM FARMING USING IMAGE PROCESSING

KAVEESHA PERERA, RISHMA PACKEERAN, YASASSI SURIYABANDARA, HUMAIRA RIZWAN, ANURADHA

KARUNASENA, LOKESHA WEERASINGHE

Department of Information Technology, Sri Lanka Institute of Information Technology, Malabe, Sri Lanka.

The field of mushroom cultivation is on the rise within Sri Lanka’s agricultural sector. This trend is gaining traction among
farmers, entrepreneurs, and consumers due to its cost-effectiveness, substantial yields, and growing demand both locally and
globally. Despite its potential, mushroom farming remains a relatively fresh concept in Sri Lanka, leaving many cultivators
unfamiliar with the necessary techniques. Button mushrooms, a widely consumed variety globally, are a prime example. These
mushrooms hold significant value as a crop, suitable for year-round cultivation under controlled conditions. Their short
production cycle enables farmers to yield multiple crops annually, ensuring a consistent income stream. This aspect elevates their
importance as a livelihood source for countless farmers and business owners across the world. While cultivating button
mushrooms presents numerous benefits, specific obstacles still impede farmers from achieving optimal harvests. Conventional
manual techniques often lead to deficiencies and imperfections in mushroom farming. In response, this paper proposes an ICT
solution to address these inefficiencies in the cultivation process. To tackle issues such as insect infestations and diseases,
advanced models like MobileNetV2 and DenseNet121 have demonstrated remarkable accuracies of 88% and 90%, respectively,
in identification. Furthermore, the implementation of the MobileNetV2 architecture attains an 85% accuracy in categorizing
mushroom quality attributes. Predicting harvest timing, a critical factor, employs the YOLOv8l instance segmentation model,
offering a substantial 0.657 [email protected] value. These cutting-edge techniques not only provide growers with crucial insights but
also empower them to take timely actions to enhance their cultivation practices.

Keywords— Button Mushrooms, Image Processing, diseases, Mushroom quality, insects, harvest time

Mushroom cultivation holds a prominent position in the significant risks arise from insect infestations and the spread of
agricultural landscape of Sri Lanka, with particular emphasis on the diseases within the crop.
cultivation of button mushrooms (Agaricus bisporus). This species
flourishes in the tropical and subtropical regions of the country, Identifying insect infestation in button mushrooms is crucial for
showcasing a rich diversity of fungal species (Mba. et al., 2020). several reasons. Firstly, it is essential for quality control since infested
The cultivation of button mushrooms, however, poses inherent mushrooms can be discolored, deformed, and emit unpleasant odors,
challenges, necessitating vigilant monitoring of critical parameters making them unsuitable for consumption or sale. Secondly, some
such as temperature, humidity, ventilation, and other environmental insects that infest button mushrooms can carry harmful bacteria such as
factors. Ensuring optimal growth conditions, employing effective Escherichia coli and Salmonella, which can cause food poisoning and
composting techniques, implementing meticulous harvesting pose health risks to consumers. Thirdly, insect infestation can have a
practices, and exhibiting expertise in post-harvest handling significant economic impact on mushroom farmers, resulting in crop
collectively contribute to enhancing the overall quality and flavor failures, reduced yields, and financial losses. Therefore, managing and
profile of button mushrooms. identifying insect infestation in button mushrooms is crucial to ensure
In Sri Lanka, button mushroom growing has grown in popularity their quality, safety, and economic viability in the market.
recently. They are also referred to as white mushrooms or white When considering insects, there are several insects that can harm
button mushrooms. They can be found in a wide range of foods and mushrooms, including flies and beetles. Beetles, including the lycid
cooking processes, such as pizza, pasta, pies, and omelets. White beetle, can also be a problem for mushrooms. Flies (Rinker D.L.,
mushrooms are high in nutrients, such as protein, Vitamin C, and D, 2017) such as the Bradysia, Cecid Fly, Phorid Fly, and Sciarid Fly can
yet low in calories (Mba. et al., 2020). White mushrooms are highly lay their eggs on compost, and their larvae can feed on the nutrients in
nutritious and also have a number of therapeutic applications. Some the compost, causing further damage. Farmers are often unable to detect
of the health benefits include lowering the risk of cancer, these infestations in their early stages. If this spreads, they will have to
controlling blood sugar, and boosting immunity (Desai R.V., replace the entire bag of compost mixture. As a result of the insect
2023). infestation in their compost mixtures, farmers may suffer significant
Button mushroom cultivation encompasses multiple stages losses in their mushroom cultivation.
(Vikaspedia., 2022), including substrate preparation, spawning, Button mushroom is among the greatest consumable cultivated
casing, and fruiting. The initial step entails the meticulous mushrooms in Sri Lanka due to their medicinal and nutritional
preparation of the substrate, typically composed of composted horse properties. One specific kind of fungal plant that is deficient in
manure, straw, and other organic matter. This substrate undergoes chlorophyll is the button mushroom (Ojha S. et al, 2 0 1 7 ). As
sterilization to eradicate any potentially harmful bacteria or fungi mushrooms are grown, a variety of dangerous fungi are discovered in
and is subsequently conditioned to reach the appropriate the casing soil and compost. The fruit body is hindered during
temperature. Once the substrate has cooled, it undergoes inoculation different phases of crop growth, resulting in specific disease
with mushroom spawn, a specialized mycelium cultivated on a symptoms (Ojha S. et al, 2 0 1 7 ). Currently, farmers in Sri Lanka
nutrient-rich medium. Following the spawning process, the casing detect mushroom diseases with their naked eye. However, this
layer is applied atop the compost layer, serving to facilitate the method may not be accurate or effective, especially if they are not
growth of mushroom fruiting bodies. The final stage entails familiar with disease types and their symptoms. As most of the
cultivation within a controlled environment to optimize growth farmers have very limited knowledge on the diseasetypes and their
conditions. Given the fungal nature of button mushrooms, the most symptoms and they are also not aware of control measures to prevent
the nearby crops from the diseases. This is because they consult based on diseases and symptoms once the data has been collected,
experts at that time, and it would be costly and time-consuming as the and they will be maintained in the knowledge base and accessed using
experts take time in classifying the diseases and identifying the the forward chaining approach. The system will then fetch the
symptoms. And most farms would not be able to afford to consult relevant rules and provide the proposed remedy when the farm owner
experts regularly. However, implementing this approach would be inputs the symptom. The user should input the answer to questions
time-consuming, require specialized knowledge of mushroom based on the symptoms and conditions of mushrooms for the system to
diseases, and involve an extensive workload, thereby potentially detect the associated diseases and suggest control measures. This
facilitating the rapid spread of diseases. Therefore, it is essential to system is being developed using methods including rule-based
identify the types of diseases to prevent their occurrence and algorithms and forward-chaining inference engines. The user may
minimize their impact on the crop. identify mushroom diseases using the online system with the assistance
The quality and taste of button mushrooms depend on appropriate of the findings processed by this system, which also offers helpful
composting, proper growing, careful harvesting, and professional suggestions.
post-harvest practices. Natural Harvest Mushroom (NH) mushrooms Using a support vector machine model, A. Masoudian and K.
are well known for their superior quality and flavor in the Sri Lankan A. Mcisaac (Masoudian et al, 2013) describes a system that identifies
market. The market values white button mushrooms (Agaricus bis- the microbial decay of mushrooms. According to the study,
porus) for their white appearance and glossy surface (Masoud et al, mushrooms will be classified as fresh or harmful to guarantee the
2012).Browning of the cap or discoloration shriveled or distorted quality of crops. This uses image processing and a support vector
caps, and open gills are some of the indicators of poor-quality white machine model for classification. The classification is done by
button mushrooms. This will lead to a reduction in the shelf life of capturing images, and techniques such as image preprocessing, image
the button mushroom and lower financial returns for the button enhancement, and SIFT feature extraction methods are used. This
mushroom farmers. The quality of the mushroom can be identified study has used images of both fresh and diseased mushrooms as the
using different factors. Size, Color, Shape, and gills are some of the dataset. To assess the effectiveness of the SVMs, the dataset makes use
main characteristics that can be checked when identifying the quality of several kernel functions.
of the mushroom. The surface of fresh button mushrooms should be Da-Wen Sun and Xiaohui Lin (Sun et al, 2019) thoroughly explore
smooth and spotless. If the mushroom is discolored, has abnormal the complex domain of mushroom browning in a comprehensive
shapes, or has black gills, it is of low quality. Overall, the quality of manner. The study not only explores a multitude of methodologies for
button mushrooms may significantly affect their flavor and texture, appraising the coloration of mushroom surfaces but also delves into the
so it’s critical to choose mushrooms that are vibrantly colored, fresh, pivotal determinants behind the browning process. These encompass
and flavorful. factors such as microbial infections, postharvest conditions, and the
For a batch of button mushrooms, there are three flushes. The first catalytic role of mushroom tyrosinase. Moreover, the research
flush has a higher yield than the other two because it absorbs the rigorously examines and contrasts diverse strategies aimed at curbing
majority of the nutrients from the soil and casing. Three weeks after browning occurrences, encompassing Modified Atmosphere Packaging
casing, the first crop appears. Every flush will be spaced out by 7 to (MAP), inhibitors of polyphenol oxidase (PPO), edible coatings,
14 days. The harvest is carried out during the button stage, with caps irradiation, freezing, and innovative cutting-edge techniques. Notably,
measuring 2.5 to 4 cm, according to the Vikaspedia (Vikaspedia., the empirical investigations were conducted on post-harvest white
2022) in India. In Sri Lanka, the harvesting is done by the most button mushrooms, providing valuable insights into the intricate
experienced workers. Based on the size, shape, color, and texture of dynamics of browning phenomena and its potential management
the cap, they are selecting mushrooms with their naked eyes. The strategies.
mushrooms may be small and immature if the workers are M. Nadim et al. (Nadim et al., 2019) have conducted a thorough
inexperienced in mushroom picking, and they may even be already investigation into the various image processing techniques utilized to
decomposing if they are harvested too late. Additionally, training carefully evaluate the quality of mushrooms. The research underscores
new employees is difficult. Therefore, it is essential to harvest at the the application of advanced image processing techniques to gauge the
proper time to increase productivity, quality, and profitability in the visual attributes of mushrooms, thereby elucidating their quality. This
mushroom farming industry. endeavor encompasses a comprehensive analysis of parameters like
color, area, density, and capacity, which were meticulously extracted
MATERIALS AND METHODS through data mining techniques. Leveraging the power of artificial
neural networks and fuzzy logic systems, the study employs a three-
A. Sanchez-Ortiz et al. (Sanchez-Ortiz et al., 2017) proposes a tier classification framework—comprising flawed, moderately
novel method for the larva image classification of Aedes mosquitos satisfactory, and pristine conditions—to meticulously analyze 250
based on convolutionalneural networks (CNN). The proposed scheme photographic instances. Remarkably, the image processing framework
provides a high identification rate of the Aedes larvae, giving a attains an impressive 95.6% accuracy in its adeptness at precisely
false negative error rate is 0% and a true negative rate is 70%. This discerning objects and their quality attributes.
makes the system suitable for quickly detecting them in the field, The growth phase of apricot mushrooms was captured using the
which in turn canlead to effective mosquito control measures. YOLOv4 model by B.R. Huang et al. (Huang et al., 2021), and they
The research paper, conducted by R. Rijal, P. Maity, and A. created an intelligent system that relies on image recognition that has
Kumar (Rijal et al., 2021) discusses the topic of pest control within beencreated and put to the test. The growth status of the mushroom has
mushroom farming. According to their findings, the main challenges been categorized using labeling classification as good, none, little,
in mushroom pest management are a lack of scientific research, the or red. The distance between the lens and the mushrooms was also
use of less innovative technologies, a lack of effective formulations, measured using an image distance identification approach.
and a lack of thorough analysis of mushroom ecosystems. Integrated A harvesting robot with a flexible joint was designed in a study
Pest Management (IPM), which includes biological, cultural, from China, and static analysis was suggested to measure the end-
physical, mechanical, and chemical methods, is the most effective and contact effector’s force during the plucking of button mushrooms.
efficient pest management tool. IPM helps ecosystem mushroom With the harvest robot, they have an 88.2% success rate and a 2.9%
growers manage mushroom pests by providing natural and injury rate according to (Yang et al., 2022).
environmentally friendly tools, allowing them to produce healthy and For oyster mushroom farms in Sri Lanka, an IOT-based monitoring
nutritious mushrooms. This study provides background information system has been created to identify harvest time by Y. Surige et al.
on pestmanagement in mushrooms. (Surige et al., 2021). They have captured oyster mushrooms in their
An intelligent system for identifying diseases of oyster five life stages and utilized CNN models to predict when they would
mushrooms is developed by M. Y. Munirah, M. Rozlini, and Y. S. be harvested. And they had over-fitting accuracy of 92% and 99%,
Mariam (Munirah e t a l , 2 0 1 3 ). using a set of rules regarding the respectively.
symptoms of the diseases. According to the study, the main sources A 99.24% mean average precision deep learning method called
for the system were obtained from one of the oyster mushroom farms Mushroom-YOLO is suggested for identifying mushroom growth. A
through an interviewing process. A set of rules will be developed prototype system called iMushroom is exhibited in (Wang et al., 2022),
showing how well it can regulate the growth quality of shiitake that needs to be checked for diseases will be captured through the
mushrooms withoutthe need for human involvement. mobile application.
b) Data Preparation - Prior to model training, the collected images
The primary objective of this research is to address the challenges underwent preprocessing steps to enhance the quality and ensure
encountered by button mushroom cultivators resulting from their consistency. These techniques encompassed resizing the images,
current manual farming practices. As a result, a study was conducted normalizing the pixel values, and performing one-hot encoding. As the
with the aim of developing a smart mushroom farming solution produced images exhibited variations in size and color formats, they
specifically for mushroom cultivators in Sri Lanka, effectively underwent a transformation process. This involved resizing all images
resolving the unaddressed issues they face. to a uniform size of 224 × 224 and converting their color format to
RGB. The normalization process involved scaling the pixel values to a
A. Insect Infestation Identification range between 0 and 1 by multiplying them by 255. These steps aimed
to reduce the computational complexity and improve the convergence
There are three primary categories (Paul R., 2019) of larvae
speed of the algorithms. To further augment the dataset and mitigate
that can have an impact on the growth of button mushrooms, as
potential overfitting, data augmentation techniques such as rotation,
illustrated in the accompanying Fig. 1.
flipping, and random cropping were applied to the images. This
helped introduce additional variations and increase the generalization
capability of the model.
c) Training the model - The model was implemented to identify
5 disease types that harm button mushrooms such as bacterial blotch,
green mold, black mold, cobweb and mite attack. Mainly five
disease types are divided based on their color and other features.
Fig. 1. Larvae Types Transfer learning-based CNN architectures were used to train the
models in the Google Colab environment. The architectures were
a) Data Collection - The experiment utilized a data set sourced DenseNet121, InceptionV3, MobileNetV2, ResNet50, VGG16,
from multiple channels, including online websites such as Kaggle, Xception and InceptionResNetV2. To ensure real-time predictions and
and on-site visits to a button mushroom farm situated in Ambewela, precise outcomes, the DenseNet121 architecture was selected due
Sri Lanka. A collection of 510 images was amassed during the data- to its accuracy and performance compared to other architectures.
gathering process and subsequently included in the data set. The DenseNet design encourages feature reuse and promotes efficient
training images constituted around 80% to 90% of the total data set, use of parameters through dense connections (Wang et al, 2021). The
with the remaining 10% to 20% assigned for validation purposes. key benefit of utilizing DenseNet121 lies in its faster execution
b) Data Preparation - To enhance the dataset’s accuracy, compared to a standard convolutional network, making it particularly
several preprocessing techniques were employed. These techniques well-suited for mobile applications. Input images were resized into a
encompass resizing the images, normalizing the pixel values, and uniform resolution. Subsequently, these images were passed through
performing one-hot encoding. The normalization process involved several convolutional layers, pooling layers, and finally a fully
scaling the pixel values to a range between 0 and 1 by multiplying connected layer to generate the classification results.
them by 255. Moreover, data augmentation techniques were
implemented to tackle overfitting and introduce greater diversity to
the data set. These techniques encompassed operations like rotation, C. Mushroom Quality Identification
shear, shift, zoom, flip, and fill, resulting in the generation of a) Data Collection- As per the insights provided by domain experts,
additional sample images. By utilizing these methods, the dataset’s a multitude of factors exert their influence on the quality attributes
precision was improved, facilitating more effective analysis and of white button mushrooms. This meticulous process unveiled four
interpretation. primary factors that emerged as the focal points for classification. A
meticulously curated collection of 1031 images, showcasing affected
c) Training the model- The model was implemented to identify
areas, was amassed from a white button mushroom farm nestled in
three larval types that harm button mushrooms, such as Sciarid fly
Sri Lanka’s expanse. This repository encompasses 251 instances of
larvae, Phorid fly larvae, and Cecid fly larvae. Mainly, these larvae
’Discolored Mushrooms,’ 212 instances of ’Black Gills,’ 210 instances
types are divided based on their color and size. Six architectures of
of ’Texture’ anomalies, 108 instances of ’Abnormal Capsize,’ and 250
transfer learning were used to train the models in the Google
instances of ’Good Mushrooms.’ The training dataset was thoughtfully
Colab environment. The architectures were InceptionV3,
composed, constituting an approximate range of 80% to 90% of the
MobileNetV2, ResNet50, VGG16, Xception, and
total dataset, with the remaining 10% to 20% meticulously earmarked
InceptionResNetV2. To ensure real-time predictions and precise
for validation purposes. It is of paramount importance to note that the
outcomes, the MobileNetV2 architecture was selected due to its
image requisition for assessing white button mushroom quality will be
accuracy and performance compared to other architectures. The key
facilitated through a dedicated mobile application, thus enhancing the
benefit of utilizing MobileNetV2 lies in its faster execution
efficiency and convenience of the quality evaluation process.
compared to a standard convolutional network, making it particularly
well-suited for mobile applications. Input images were resized to a b) Data Preparation- To refine the dataset’s accuracy, we adopted
uniform resolution of 224x224 using 3 channels. Subsequently, these several careful preprocessing techniques. These included adjusting
images were passed through several convolutional layers, pooling image sizes, standardizing pixel values, and using one-hot encoding.
layers, and finally a fully connected layer to generate the To normalize the pixel values, we scaled them between 0 and 1 by
classification results. multiplying by 255. Additionally, we countered the risk of overfitting
and added diversity by applying data augmentation methods. These
methods involved various transformations like rotating, shearing,
B. Disease Detection shifting, zooming, flipping, and filling pixels. This effort resulted
in the creation of extra images, broadening the dataset’s range. These
a) Data Collection - Initially, a comprehensive survey of the farm adjustments significantly improved the dataset’s reliability.
was conducted in collaboration with domain experts to identify the Consequently, our subsequent analyses and interpretations were
key disease types affecting the crops. Through this process, five underpinned by a more robust foundation, affirming the overall
disease types were selected as targets for classification. A total effectiveness of our research.
of 800 images of infected regions were collected from a button
mushroom farm located in sri lanka. The dataset carries 216 images c) Training the model- The main goal of the study was to create
as ‘Black Mould’, 214 images as ‘Green Mould’, 220 images as mite a model that could assess the quality of button mushrooms by
attack’, 102 as ’Bacterial blotch, and 108 as cobweb. The training considering four key factors: discoloration, the presence of black gills,
images constituted around 80% to 90% of the total data set, with the texture, and any irregularities in cap size or shape. To achieve this, the
remaining 10% to 20% assigned for validation purposes. The image research utilized five distinct transfer learning architectures:
InceptionV3, MobileNetV2, ResNet50, Xception, and
InceptionResNetV2. These architectures were implemented and
trained within the Google Colab platform. Among these options, the
MobileNetV2 architecture stood out due to its remarkable accuracy
and superior performance when compared to the other architectures.
As a result, it was chosen as the most suitable architecture for
ensuring real-time predictions and highly accurate outcomes.

D. Harvesting Time Prediction

a) Data Collection- Images of button mushrooms at four
different growth phases were gathered from the Ambewela button
mushroom farm and several online sources. During the data
collection, a collec- tion of 724 photos was gathered and included to
the data set. Around 80% to 90% of the overall data set was made up Fig. 2. Accuracy graph for Insect Infestation Type Identification
of training images, with the remaining 10% to 20% being used for
validation and testing.
b) Data Preparation - The images have to be annotated in accor-
dance with the stages of harvest time after being collected. Mainly
these mushrooms are annotated based on their colour and size.
Prepossessing methods were used to improve the data set’s precision.
utilized the auto orient approach to remove EX-IF rotations and
uniformly organize pixels.Since the final images arrive in a range
of sizes, we reduced the images for smaller file sizes and quicker
training. Due to the strong contrast in the desired colors, we used
gray scale approaches. Data augmentation techniques like rotation,
crop, brightness, flip, and exposure were employed to enhance the
number of sample images in the data sets and prevent the models
from being over fit.
c) Training the model - The model was used to determine the
harvesting time of button mushroom as 4 stages, including 24 hours,
15 hours, 5 hours, and 0 hours to harvest. The models were trained Fig. 3. Loss graph for Insect Infestation Type Identification
in the Google Co-lab environment using four architectures. The two
primary architectures used to train instance segmentation models are
YOLOv5 and YOLOv8. As single-stage detectors that can predict
bounding boxes and class probabilities for all objects in an image B. Disease Detection
in a single pass, YOLOv5 and YOLOv8 are effective for instance An extensive study was undertaken to achieve precise disease
segmentation. They are thus faster than two-stage detectors like
classification, employing a transfer learning strategy encompass- ing
Faster R-CNN. Furthermore, it has been demonstrated that YOLOv5 seven distinct architectural models: InceptionV3, MobileNetV2,
and YOLOv8 are extremely accurate for instance segmentation.The DenseNet121, ResNet50, VGG16, Xception, and InceptionResNetV2.
architectures that were employed were YOLOv5m6u, YOLOv8l, Each architecture underwent training and subsequent evaluation,
YOLOv5lu, and YOLOv8x. All of these models underwent 500 yielding testing accuracies of 80%, 85%, 96%, 51%, 55%, 81% and
epochs of training. Based on its performance, the YOLOv8l was 75%, respectively. Among these, DenseNet121 emerged as the
selected as the best model to forecast the button mushroom harvest optimal choice, demonstrating superior accuracy and overall
period. performance compared to the remaining trained architectures. The
graphical representation of the model’s training accuracy is illustrated
RESULTS in Fig. 4, while Fig. 5 portrays the associated training loss throughout
the training process. This study stands out from previous research by
A. Insect Infestation Identification achieving a notable increase in accuracy through the implementation
of transfer learning methodologies. Unlike earlier studies, this research
A comprehensive study was conducted to identify insect infesta-
distinctly applies these techniques to precisely detect the diseases of
tions with high accuracy, implementing a transfer learning approach
button mushrooms, resulting in enhanced outcomes.
with six distinct architectures such as InceptionV3, MobileNetV2,
ResNet50, VGG16, Xception, and InceptionResNetV2. Each archi-
tecture was trained and evaluated, with testing accuracies of 81%,
88%, 58%, 71%, 87%, and 79%, respectively. The MobileNetV2 ar-
chitecture exhibited good performance, achieving a training
accuracy of 98.96%. Consequently, MobileNetV2 was selected as
the optimal model due to its accuracy and overall performance
compared to the other trained architectures. The training accuracy
of the model can be observed in Fig. 2, while Fig. 3 represents the
corresponding loss encountered during the training process. When
compared to previous studies in this field, the presented approach
stands out for its significantly improved accuracy through the use of
transfer learning techniques.

Fig. 4. Accuracy graph for the Disease Type Identification

 D. Harvesting Time Prediction
Five main strategies were utilized to train the instance segmentation
model to predict when button mushrooms will be harvested:
YOLOv8l, YOLOv8x, YOLOv5lu, and YOLOv5m6u. Each trained
model’s performance was compared using the metrics [email protected],
[email protected]:0.95, precision, and recall in order to determine its accu-
racy. The model built on YOLOv8l using derived hyper-parameters
performed the best.To be more precise, it achieved [email protected] of 0.657
and [email protected]:0.95 of 0.581. The [email protected] for YOLOv8x,
YOLOv5lu, and YOLOv5m6u was 0.621, 0.641, and 0.604, respec-
tively. Fig. 8 shows the model’s [email protected]. The model’s box loss
and the class loss represents from the Fig. 9 and Fig. 10 respectively.
Most other growth detection studies employed yolov4 and yolov5
models, but we used yolov8, which is a more powerful and versatile
object detection model, in our study. Fig. 11 shows an example of the
Fig. 5. Loss graph for the Disease Type Identification harvest time detection where only mushrooms from 2 stages exist.
The mushroom will be ready for harvest in 24 hours, according to the
orange bounding box. The mushroom will be ready for harvest in 15
C. Mushroom Quality Identification hours, according to thelight red bounding box.
Five distinct neural network architectures, namely MobileNetV2,
InceptionResNetV2, InceptionV3, ResNet50, and Xception, under-
went an exhaustive analysis to assess their effectiveness in cate-
gorizing mushroom quality. Following the training and evaluation
process, the testing accuracies achieved by each architecture were as
follows: 85% for MobileNetV2, 82% for InceptionResNetV2, 83%
for InceptionV3, 42% for ResNet50, and 82% for Xception. Notably,
the MobileNetV2 architecture exhibited exceptional performance,
achieving a remarkable training accuracy of 85.58%. This outcome
underscores its proficiency in capturing intricate patterns within the
mushroom-quality dataset during the training phase. The graphical
representation of the model’s training accuracy is illustrated in Fig.
6, while Fig. 7 depicts the corresponding loss experienced throughout
the training iterations. A key differentiator of this study from prior
research lies in the significant enhancement of accuracy achieved
through the application of transfer learning techniques. Unlike
previous investigations, this study uniquely employs these approaches
to accurately identify the quality of button mushrooms, yielding
improved results.

Fig. 8. mAP50(B) graph for the Harvest time prediction of the mushroom

Fig. 6. Accuracy graph for the identification of the mushroom

quality
Fig. 9. Box Loss graph for the Harvest time prediction of the mushroom

Fig. 7. Loss graph for the identification of the mushroom quality

 incorporating dropouts have emerged as pivotal strategies in averting
Fig. 10. Class Loss graph for the Harvest time prediction of the mushroom overfitting during the model's training phase.
Methodology for harvest forecasting Farmers may find it useful to
monitor the growth of mushrooms to determine when to harvest them.
Since mushrooms will be harvested at the appropriate time, a rise in
product quality will be expected as a result. In addition, various
prospective applications may be supported by the properly recognized
harvest time of the mushrooms. For instance, it might be a helpful
component of a robotic system that identifies, selects, and harvest.
Additionally, there are certain drawbacks to the YOLOv8-based
approach of mushroom time prediction. We used several angles of
mushroom photos in our method. As a result, it was quite challenging
to properly annotate them. But the results were good from the given
amount of collected data. However, we could utilize a bigger dataset
with better annotated mushrooms to get better findings if we needed to
improve them.

Fig. 11. Example of Harvest time detection in 2 different stages CONCLUSION

DISCUSSION In conclusion, the MobileNetV2 Algorithm was employed to

achieve the highest accuracy in identifying various insect types. These
The research paper investigates the use of advanced deep infestations pose a significant threat, causing considerable economic
learning methods to identify insect infestations in button mushrooms. damage to mushroom cultivation. Identifying the sort of insect
It evaluates six different architectures, including InceptionV3, infestation will assist in considerably limiting the harm that happens.
MobileNetV2, ResNet50, VGG16, Xception, and Disease type identification is presented based on transfer learning-
InceptionResNetV2, ultimately selecting MobileNetV2 as the most based CNN architectures. DenseNet121 was selected as the best
effective model. The study emphasizes meticulous data collection architecture for disease type identification with 96.77% testing
from various sources, resulting in a dataset comprising 510 images. accuracy. Mushroom Quality identification is the process of deter-
Several preprocessing techniques and methods to augment the data mining mushroom quality through the utilization of transfer learning-
are applied, enhancing both accuracy and diversity. The research based Convolutional Neural Network (CNN) models is outlined.
specifically targets the detection of three types of larvae in button Among the various architectures considered, MobileNetV2 emerged
mushrooms and validates MobileNetV2's superiority, especially in as the most optimal choice for mushroom quality identification,
real-time predictions. This study represents a significant advancement achieving a notable testing accuracy of 85.58%. In order to identify the
in agricultural pest management, providing valuable insights for harvesting time of button mushrooms, the YOLOv8l algorithm based
farmers and researchers and promoting the adoption of technology- on instance segmentation model acquired the greatest mAP@50 value
driven sustainable farming practices. of 0.657.
The outcomes of this research hold significant impacts on the For future scope, The insect infestation identification model can
field of mushroom disease classification using deep learning and be expanded to cover a wider range of insect larva types, and we can
image processing techniques. The achievement of an accuracy of utilize a robotic system to pick mushrooms to accurately determine the
96% with DenseNet121 architecture demonstrates the potential of harvest time, and pick them without manual inspection. Disease
transfer learning-based CNN architectures in accurately identifying identification models can be further elaborated to detect other types
diseases affecting button mushrooms. This can play an important role of button mushroom diseases and further development could be done
in aiding farmers by suggesting rapid and reliable disease to detect the different stages of button mushroom diseases based
assessments. The model successfully classified diseases with on the severity level of the disease. To enhance the evaluation of
comparatively fewer samples by using information gained from a mushroom quality, we can segment the assessment into multiplestages.
bigger source domain dataset. This approach showcases the This approach helps us gain a more thorough understanding of the
adaptability of deep learning techniques to domains with limited mushroom’s overall condition.
labeled data, which is a common constraint in agriculture. To address
underfitting, adjustments such as augmenting the number of neurons REFERENCES
within each layer, expanding the layers in the model, and modifying
the existing layer types have proven beneficial. Meanwhile, in the Desai R. V. (2023). Button Mushrooms: Health Benefits and Nutritional
training phase to mitigate overfitting, techniques such as early Value. Bajaj Finserv Health.
stopping, data augmentation, and implementing dropouts have been https://www.bajajfinservhealth.in/articles/button-mushrooms.
utilized. The training of a CNN architecture depends heavily on
parameters such as learning rate, optimizer selection, and batch size.
Proper tuning of these hyperparameters can lead to better D.-W. Sun and X. Lin, Research advances in browning of button
convergence and accuracy. mushroom (Agaricus bisporus, ser. Trends in Food Science
Technology. Elsevier, 2019, vol. 90.
The use of deep learning and image processing techniques in
this study has significant implications for the assessment of
mushroom quality. The MobileNetV2 architecture was used to Huang B. R., Chen J. Y., Lai X. Y., Chen G. W., and Huang K. W.
achieve an outstanding accuracy rate of 86%, which highlights the (2021). Application of deep learning for mushrooms cultivation.
capability of transfer learning-based convolutional neural networks International Conference on Technologies and Applications of
(CNNs) to effectively judge the grade of button mushrooms. This Artificial Intelligence (TAAI). IEEE, pp. 250–253.
invention has the potential to greatly help farmers assess mushroom
quality, an important step in their farming process. Notably, the Mba, A. L. B. (2020). White Mushrooms: Nutrition, Benefits, and Uses.
model's performance in quality classification with a little dataset Healthline. https://www.healthline.com/nutrition/white-mushroom-
illustrates how deep learning techniques may be used to problems nutrition.
when labelled data is few, which is a prevalent issue in agriculture.
To enhance the model's efficacy, effective measures involve
adjustments such as augmenting the neuron count within each layer Munirah M. Y., Rozlini M., and Mariam Y. S. (2013). An expert
and modifying layer types, which serve to mitigate underfitting. system development: Its application on diagnosing oyster
Simultaneously, tactics like early stopping, data augmentation, and mushroom diseases. 13th International Conference on Control,
 Automation and Systems (ICCAS 2013), pp. 329–332.

Nadim M., Ahmadifar H., Mashkinmojeh M., and Yamaghani M.

(2019). Application of image processing techniques for quality
control of mushroom. Caspian Journal of Health Research, vol.
4, pp. 72–75.

Ojha S. and Chowdhury D. (2017). An empirical study on mushroom

disease diagnosis: A data mining approach. International
Research Journal of Engineering and Technology (IRJET), vol.
04, pp. 529–534

Paul R. (2019), Pests of Mushroom and Its Control — India —

Agriculture. https://www.agricultureinindia.net/pest-
control/mushroom/pests-of-mushroom-and-its-control-india-
agriculture/15890 /.

Rijal R., Maity P., and Kumar A. (2021), pests of mushroom and their
ecological management strategies: A review. Research Trend.

Rinker D.L. (2017). Edible and Medicinal Mushrooms: Technology

and Applications. 1st ed. John Wiley Sons Ltd.

Sanchez-Ortiz A., Fierro-Radilla A., Arista-Jalife A., Cedillo-

Hernandez M., Nakano-Miyatake M., Robles-Camarillo D., and
Cuatepotzo-Jiménez. (2017). Mosquito larva classification
method based on convolutional neural networks. International
Conference on Electronics, Communications and Computers
(CONIELECOMP).

Surige Y., Perera W., Gunarathna P., Ariyarathna K., Gamage N., and
Nawinna D. (2021). Iot-based monitoring system for oyster
mushroom farming. 3rd International Conference on
Advancements in Computing (ICAC). IEEE, pp. 79–84.

Taghizadeh A. G. M. and O’Donnell C. P. (2012). White button mush-

rooms, quality evaluation and shelf life prediction. pp. 231–250

Vikaspedia. (2022). Button Mushroom Production.

https://vikaspedia.in/agriculture/farm-based-
enterprises/mushroom- production/button-mushroom-production.

Wang R., Li J., Chencho, An S., Hao H., Liu W., and Li L. (2021).
Densely connected convolutional networks for vibration based
structural damage identification. Engineering Structures, vol. 245,
p. 112871.

Wang Y., Yang L., Chen H., Hussain A., Ma C., and Al-gabri M.
(2022). Mushroom-yolo: A deep learning algorithm for
mushroom growth recognition based on improved yolov5 in
agriculture 4.0. 20th International Conference on Industrial
Informatics (INDIN). IEEE, pp. 239–244.

Yang S., Ji J., Cai H., and Chen H. (2022). Modeling and force
analysis of a harvesting robot for button mushrooms. IEEE
Access, vol. 10, pp. 78 519–78 526.