Applicability of Deep Learning Network on Gray Mold Disease Detection on Strawberry Leaves

( Sijan Karki ),( Jayanta Kumar Basak ),( Bhola Paudel ),( Na Eun Kim ),( Nibas Chandra Deb ),( Hyeon Tae Kim ) 한국농업기계학회 2022 한국농업기계학회 학술발표논문집 Vol.27 No.2

Early and accurate disease detection in the plant is crucial to mitigate its effect and maximize the yield. Gray mold is considered the most devastating strawberry disease, leading to complete plant death. Various machine learning and deep learning-based models were developed in the past. However, most studies used a controlled environment to capture the images and trained a model whose performance decreased when the models were tested images captured in the field. Therefore, there has been a need for a model that can detect and quantify plant disease accurately, especially in the natural environment. Therefore, this study developed an image segmentation model based on deep learning to distinguish the gray mold disease in strawberry plants. Three groups of strawberry plants (ten plants in each group) were inoculated with different concentrations of necrotrophic fungus pathogen (Botrytis cinerea) and observed the resulting disease. The deep learning model (Unet) was trained with images captured in a natural environment non-destructively. Model performance was assessed using evaluation metrics like intersection over union (IoU), pixel accuracy, and dice accuracy. Furthermore, two machine learning-based models (K-means and XGBoost) were also trained with the same images, and the performance of these models was compared. The deep learning-based model had an average IoU accuracy of 82.12%, dice accuracy of 89.71%, and pixel accuracy of 98.24%, surpassing both machine learning models in multiple aspects. The XGBoost model had an average IoU accuracy of 80.89%, dice accuracy of 85.40%, and pixel accuracy of 98.16%, which performed consistently well in identifying the disease following the deep learning-based model. In conclusion, the developed model could be a valuable tool for strawberry farmers with a simple computational setup in gray mold disease detection.

Applicability of Deep Learning Network on Gray Mold Disease Detection on Strawberry Leaves

( Sijan Karki ),( Jayanta Kumar Basak ),( Bhola Paudel ),( Na Eun Kim ),( Nibas Chandra Deb ),( Hyeon Tae Kim ) 한국농업기계학회 2022 한국농업기계학회 학술발표논문집 Vol.27 No.2

Early and accurate disease detection in the plant is crucial to mitigate its effect and maximize the yield. Gray mold is considered the most devastating strawberry disease, leading to complete plant death. Various machine learning and deep learning-based models were developed in the past. However, most studies used a controlled environment to capture the images and trained a model whose performance decreased when the models were tested images captured in the field. Therefore, there has been a need for a model that can detect and quantify plant disease accurately, especially in the natural environment. Therefore, this study developed an image segmentation model based on deep learning to distinguish the gray mold disease in strawberry plants. Three groups of strawberry plants (ten plants in each group) were inoculated with different concentrations of necrotrophic fungus pathogen (Botrytis cinerea) and observed the resulting disease. The deep learning model (Unet) was trained with images captured in a natural environment non-destructively. Model performance was assessed using evaluation metrics like intersection over union (IoU), pixel accuracy, and dice accuracy. Furthermore, two machine learning-based models (K-means and XGBoost) were also trained with the same images, and the performance of these models was compared. The deep learning-based model had an average IoU accuracy of 82.12%, dice accuracy of 89.71%, and pixel accuracy of 98.24%, surpassing both machine learning models in multiple aspects. The XGBoost model had an average IoU accuracy of 80.89%, dice accuracy of 85.40%, and pixel accuracy of 98.16%, which performed consistently well in identifying the disease following the deep learning-based model. In conclusion, the developed model could be a valuable tool for strawberry farmers with a simple computational setup in gray mold disease detection.

Colour space selection for enhanced machine learning model performance in classifying strawberry ripeness

( Sijan Karki ),( Jayanta Kumar Basak ),( Bhola Paudel ),( Nibas Chandra Dev ),( Hyeon Tae Kim ) 한국농업기계학회 2023 한국농업기계학회 학술발표논문집 Vol.28 No.1

Color space is a fundamental concept in image processing and machine learning, as it plays a crucial role in determining how an image is represented and processed by the computer. The choice of color space can have a significant impact on the performance of machine learning algorithms, as different color spaces have different properties and emphasize different features. Therefore, this study aimed to evaluate the efficacy of machine learning models in classifying strawberry ripeness stages using color spaces: RGB, HLS, HSV, CIELab*, and YCbCr. The study results indicate that the four ripeness stages: unripe, semi-ripe, ripe and over-ripe exhibited significant differences in biochemical and color features. While the Unripe stage was the most correctly classified stage, the Semi-ripe stage was the most challenging. The Feed Forward Artificial Neural Network using the CIELab* colour space was the most successful in classifying ripeness stages with an average accuracy of 96.7%. This combination with other features, which indicate fruit ripeness, may be utilized in the automatic detection of strawberry ripeness.

Machine learning approaches in modeling of digestible energy demand in mature phase of swine

( Nibas Chandra Deb ),( Jayanta Kumar Basak ),( Sijan Karki ),( Daeyeong Kang ),( Niraj Tamrakar ),( Eun-wan Seo ),( Seong Woo Jeon ),( Hyeon Tae Kim ) 한국농업기계학회 2023 한국농업기계학회 학술발표논문집 Vol.28 No.2

Digestible energy (DE) is an essential component of swine production, required to maintain proper growth, and reproducibility of swine. Therefore, the objective of this study was to evaluate the DE demand in mature phase of pigs by proving different types of feed in different pig barns during 2021 and 2022. A load cell and a livestock environment management system (LEMS) were used to measure the body weight of swine’s, room temperature, and humidity level inside the barn, respectively. A total of four ML models (MLR, SVR, RFR, and MLP) were used to forecast the DE demand, with the input parameters of room temperature, humidity, age, and body mass of swine's. The result of the study showed that age and body mass were positively correlated with DE (r>0.90). It was revealed that RFR model provided the best result (R2>0.94) compared to other models. However, MLR model provided the worst result (R2<0.94) compared to other models. Additionally, sensitivity analysis indicated that body mass had the greatest impact on forecasting DE compared to other input parameters. Finally, the study colcluded that RFR model can predict DE accurately, which can assist farmer's to proper management of DE in mature phase of swine production.

Classification of Pig Behaviour Experiencing Impaired Air Quality Due to Elevated GHG Concentration

( Niraj Tamrakar ),( Jayanta Kumar Basak ),( Nibas Chandra Deb ),( Sijan Karki ),( Myeongyong Kang ),( Daeyeong Kang ),( Seong Woo Jeon ),( Hyeon Tae Kim ) 한국농업기계학회 2023 한국농업기계학회 학술발표논문집 Vol.28 No.2

Pig welfare and health are the primary concerns in livestock management. Pig behaviors serve as important early indicators of pig stress. This research utilizes AI-based computer vision with a top-view RGB camera to monitor the behavior of pigs. The pigs were detected in an experimental pig barn using a CNN-based deep learning model, and pig motions were tracked in real-time using the Deep SORT algorithm. This approach helps segment different pig postures and assigns an activity score based on activity tracking. Various CNN-based deep learning models (Yolov7, Faster R-CNN, SSD ResNet101) were implemented to detect pigs in different postures. When comparing different pig identification and posture detection models, Yolov7 was found to be the fastest and most accurate, with a mean average precision (MAP) of 97.84%. Similarly, the accuracy of multiple object tracking was 93.2%, and the precision was 81.4% for the tracking algorithm. The study observed behavioral changes in both groups and individuals due to the natural elevation of GHG concentration in the experimental pig barn. Higher GHG concentrations were found to have a negative correlation with pig postures like standing, walking, and sternal lying activities, whereas lateral lying had a positive correlation.

Estimation of CO₂ emissions in a swine barn based on age, body weight gain and different activities of swine

( Nibas Chandra Deb ),( Jayanta Kumar Basak ),( Bhola Paudel ),( Sijan Karki ),( Daeyeong Kang ),( Junghoo Kook ),( Myeongyong Kang ),( Seongwoo Jeon ),( Hyeon Tae Kim ) 한국농업기계학회 2023 한국농업기계학회 학술발표논문집 Vol.28 No.1

In the modern world, global warming is a serious problem that is predominantly caused by greenhouse gases (GHGs). However, due to the demand of pork, carbon dioxide (CO₂) emissions are increasing dramatically from swine burns, which have a significant impact on increasing GHGs in the atmosphere. Therefore, the objectives of this study were to measure the CO₂ emissions based on age, body weight gain and different activities of swine. The experiment was conducted in an experimental swine barn from September to December, 2022. A load cell and a livestock environment management system (LEMS) were used to measure the body weight of swine’s and CO₂ emissions level inside the barn, respectively. A 2d camera was used to record the swine’s different activities on a daily basis. The findings of the study showed that the CO₂ emissions were strongly correlated with body weight (r = 0.83) and age of swine (r = 0.86). In this study, we also found that the CO₂ emissions were highest at sleeping time (1-2 PM) and lowest at feeding time (5-6 PM). Moreover, the CO₂ emissions during sleeping (1-2 PM) and feeding (5-6 PM) time were significantly different from other activities (P < 0.05). In conclusion, this study recommends additional research need to be conducted in the different seasons to estimate the CO₂ emissions in concern to swine’s age, body weight and different activities by providing different additives of diets.