یادگیری عمیق

In this research, field experiments were carried out in two types of soil (sandy clay loam and clay) to model the traction performance of the tractor considering drawbar power, rolling resistance, and traction efficiency, using deep learning and convolutional neural network while having some parameters such as soil type and conditions, tool parameters, and operation parameters. The tests were conducted within each soil texture in the form of factorial tests based on the randomized complete block design (RCBD) in triplicates. The tests were done in various moisture levels (8-17% for dry soils and 18-40% for moist soils), tractor forward speed (1.2, 1.6, 1.8, and 2.2 km h-1), working depth (30 and 50 cm), the number of pass (2 and 6 times), and tire inflation pressure (20 and 25 psi). The cone index, dynamic load, draft force, and moisture content were measured in each tests. The networks designed to model the drawbar power, rolling resistance, and traction efficiency were of convolutional neural network type. Various algorithms such as Sgdm, Adam, and Rmsprop were utilized to train the network. The results showed that the neural network developed by Sgdm algorithm outperformed the others. Therefore, this algorithm was utilized for the modeling process. Statistical criteria such as R2 and MSE were also employed to evaluate the performance of the network. For the drawbar power, the 8-499-499-1 architecture showed the best performance with R2=0.9953 and MSE=0.0016. Concerning the rolling resistance, the best performance was observed in 8-301-305-1 architecture with R2=0.9903 and MSE=0.0039. The best performance for the traction efficiency was obtained by 8-371-371-1 architecture with R2=0.9888 and MSE=0.003. The results showed that these networks can be used to model parameters by removing convolution layers and reducing dimensions.

Maintaining the well-being of individuals is greatly influenced by a healthy lifestyle and balanced diet. The identification and segmentation of food items can be improved by utilizing a mobile-based system in this era of rapid lifestyle changes and technology. This article introduces a novel system that, upon receiving input images, detects and segmentation the food items within the images. The system utilizes deep learning techniques and models, employing the YOLO algorithm. By incorporating regression-based simple methods, the system achieves the capability to detect and categorize food items in a single pass through the network, aiming to enhance accuracy and speed in the detection process. YOLOv7 was employed for food detection and YOLOv5, YOLOv7, and YOLOv8 was utilized for image segmentation. Based on the results, the accuracy, recall, and average precision values for YOLOv7 were 0.844, 0.924, and 0.932, respectively. Furthermore, the instance segmentation performance of YOLOv7 outperformed YOLOv5 and YOLOv8, with precision, recall, and mean average precision values of 0.959, 0.943, and 0.906, respectively. These findings underscore the high accuracy in detecting Iranian foods and the remarkable speed and precision in food image segmentation attainable through advanced deep-learning algorithms. Consequently, this study establishes that accurate detection of Iranian foods can be accomplished through the utilization of sophisticated deep-learning techniques. This research focuses on promoting a healthy lifestyle through intelligent technology and novel deep learning algorithms in Iran.

In many countries, on average, more than 50% of people's food comes from grains, and nearly 70% of the cultivated area of one billion hectares of the world is dedicated to grains. A variety of weeds grow along with cereals in the fields, which can reduce crop yield due to competition for light, water and nutrients. To eliminate weeds accurately and with minimal problems, timely detection with high accuracy and speed is required. be done.
In the field of agriculture, it is controlling and eliminating weeds in grain fields. Weeds are one of the most important factors affecting the production of agricultural products, which are their most important competitors in conventional agriculture, they spray the entire field to eliminate weeds, while weeds appear scattered and patchy in the field. which shows the necessity of using precise agriculture to solve this type of heterogeneity. In addition to causing economic damage, the conventional method of fighting can cause pollution of the environment and even the human food chain. Research shows that the losses caused by pests, diseases and weeds can reach 40% of the global crop every year and it is predicted that this percentage will increase significantly in the coming years. Besides, according to the research of Goktoan et al., the annual cost of weeds for The Australian economy is estimated to be around $4 billion as a loss in agricultural income.

Among the new methods in this field is the use of machine vision technology and related methods such as deep learning object detection algorithms and convolutional neural networks (CNN). The steps related to the implementation of the project include preparing data for training and evaluating networks, using new object detection algorithms, using different convolutional neural networks with different characteristics to extract image features in algorithms, and using the Feature Pyramid Network (FPN) method in object detection algorithms. Was. The output of the networks was evaluated in terms of the number of detections, the exact location of detection and the time of detection in the field. ViTs is based on the Transformer architecture that was originally developed for NLP tasks. Transformers use self-awareness mechanisms that allow the model to capture complex relationships between elements in a sequence. In the case of ViTs, sequence elements are image patches. In using the transformer architecture for visual data, it is dividing the image into small and non-interfering parts. Each patch typically consists of a grid of pixels. These patches are considered the "words" of the image sequence. Spatial embeddings are added to image patches to provide spatial information to the model. Spatial embeddings are necessary because transformers do not have built-in notions of order or spatial relationships. ViTs use multi-series self-awareness mechanisms to capture relationships between different image patches, and the representation of each patch is updated by attention to other patches. Data separation is very important in data watch transformers for two reasons a) the model needs data to learn and b) we need data to measure the model because the model may not be able to extract the information correctly.

The best network in terms of positioning accuracy was the transform model (ViTs) with an average accuracy of 0.95. In addition to this, the network considered in this research managed to recognize 503 of the 535 target weeds, and this means that our network is able to recognize 95% of these weeds. The presented method has been able to reach the highest accuracy compared to other existing methods and has been able to detect existing grasses in a much shorter period of time. Compared to other methods, the reset50 algorithm has been able to detect more than 88%, although its execution time is about 2.5 times that of the proposed method.
In comparing the efficiency of algorithms, execution time is as important as accuracy. By making comparisons and considering 70% of the data as training data and 30% as test data, the presented algorithm has been able to detect the weeds in the field with an accuracy of over 90% in just 13 seconds.

Today, deep learning methods are much more efficient than other methods, so we can use the new methods available in deep learning in the field of agriculture.

Accurate detection of apple leaf diseases is essential to prevent the reduction in both the quantity and quality of crop yield. With advancements in deep learning methods, the diagnosis of these diseases is improving, but data limitations remain a significant barrier. This research evaluates pretrained deep learning models with precise settings and demonstrates that high accuracy in disease detection is possible even with limited data. The selected models perform better than conventional methods and introduce transfer learning as an effective strategy to combat limited data and the diversity of diseases. These models aid farmers and horticulture specialists in identifying and managing apple leaf diseases more efficiently and rapidly. Additionally, these models are not only beneficial for farmers but also for agricultural consultants, students of agricultural sciences, and researchers in this field. Moreover, these methods are adaptable to other diseases and plants, promising advancements in plant disease detection systems in the future. The present study has the potential to revolutionize horticultural processes through optimization and automation, leading to a transformation in plant disease management.

Identifying the growth stages of fruits in orchards is an important factor in improving the quantity and quality of the final product. Having such information helps the growers to apply the appropriate treatment for each growth stage and also to get a proper understanding of the fruit harvesting time, which may change due to changing weather conditions. Therefore, in the present study, color images of golden apples were used to estimate the weeks remaining to the harvest time. The EfficientNetB1 model was used to classify images taken from different weeks of apple fruit development using deep learning technology and convolutional neural networks. The data were divided into three categories: training (60%), validation (20%), and test (20%). Also, two pre-processing processes, i.e. data normalization and data augmentation, were used to obtain better results. Finally, Nadam optimizer and categorical_crossentropy cost function were considered in creating the model. The results showed that the developed model would have a good ability to classify input images. The value of the correlation coefficient (R) for training, validation, and test data was 0.86, 0.88, and 0.87 respectively. Also, the ability of the model to classify different classes was presented using precision, recall, and f1-score parameters for each class, according to which some classes achieved 100% accuracy. Consequently, the obtained results can be used as a platform for the development of harvesting robots, mobile apps, as well as aerial imagery systems using drones, etc. to fulfill various purposes in precision agriculture, and in particular, precision horticulture.

Maize is one of the most important cereal crops worldwide, providing staple food for people globally. Counting maize tassels provides essential information about yield prediction, growth status, and plant phenotyping, but traditional manual approaches are expensive and time-consuming. Recent developments in technology, including high-resolution RGB imagery acquired by unmanned aerial vehicles (UAVs) and advanced machine-learning techniques such as deep learning (DL), have been used to analyze genotypes, phenotypes, and crops.In this study, we modified the YOLOv5s single-stage object detection technique based on a deep convolutional neural network and named it MYOLOv5s. We incorporated BottleneckCSP structures, Hardswish activation function, and two-dimensional spatial dropout layers to increase tassel detection accuracy and reduce overfitting. Our method's performance was compared with three state-of-the-art algorithms: Tasselnetv2+, RetinaNet, and Faster R-CNN. The results obtained from our proposed method demonstrate the effectiveness of MYOLOv5s in detecting and counting maize tassels. 

The High-Intensity Phenotyping Site (HIPS) dataset was collected from the large field at the Agronomy Center for Research and Education (ACRE) of Purdue University, located in West Lafayette, Indiana, USA during the 2020 growing season. A Sony Alpha 7R-III RGB camera mounted on a UAV at a 20m altitude captured high-resolution orthophotos with a pixel resolution of 0.25 cm. The dataset consisted of two replications of 22 entries each for hybrids and inbreds, planted on May 12 using a two-row segment plot layout with a plant population of 30,000 per acre. The hybrids and inbreds in this dataset had varying flowering dates, ranging from 20 days between the first and last variety.This article uses orthophotos taken on July 20th and 24th to train and test the proposed deep network "MYOLOv5s." These orthophotos were divided into 15 images (3670×2150) and then cropped to obtain 150 images (608 × 2048) for each date. Three modifications were applied to the original YOLOv5s to form MYOLOv5s: BottleneckCSP structures were added to the neck part of the YOLOv5s, replacing some C3 modules; two-dimensional spatial dropout layers were used in the defect layer; and the Hardswish activation function was utilized in the convolution structures. These modifications improved tassel detection accuracy. MYOLOv5s was implemented in the Pytorch framework, and the Adam algorithm was applied to optimize it. Hyper-parameters such as the number of epochs, batch size, and learning rates were also optimized to increase tassel detection accuracy.

In this study, we first compared the original and modified YOLOv5s techniques, and our results show that MYOLOv5s improved tassel detection accuracy by approximately 2.80%. We then compared MYOLOv5s performance to the counting-based approach TasselNetv2+ and two detection-based techniques: Faster R-CNN and RetinaNet. Our results demonstrated the superiority of MYOLOv5s in terms of both accuracy and inference time. The proposed method achieved an AP value of 95.30% and an RMSE of 1.9% at 84 FPS, making it about 1.4 times faster than the other techniques. Additionally, MYOLOv5s correctly detected the highest number of maize tassels and showed at least a 17.64% improvement in AP value compared to Faster R-CNN and RetinaNet, respectively. Furthermore, our technique had the lowest false positive and false negative values. The regression plots show that MYOLOv5s provided slightly higher fidelity counts than other methods.Finally, we investigated the effect of score values on the performance of detection-based models and calculated the optimal values of hyperparameters.

The MYOLOv5s technique outperformed other state-of-the-art models in detecting maize tassels, achieving the highest precision, recall, and average precision (AP) values.The MYOLOv5s method had the lowest root mean square error (RMSE) value in the error counting metric, demonstrating its accuracy in detecting and counting maize tassels.We evaluated the correlation between predicted and ground-truth values of maize tassels using the R2 score, and for the MYOLOv5s method, the R2 score was approximately 99.28%, indicating a strong correlation between predicted and actual values.The MYOLOv5s method performed exceptionally well in detecting tassels, even in highly overlapping areas. It accurately distinguished and detected tassels, regardless of their proximity or overlap with other objects.When compared to the counting-based approach TasselNetv2+, our proposed MYOLOv5s method showed faster inference times. This suggests that the MYOLOv5s method is computationally efficient while maintaining accurate tassel detection capabilities.

Today, the implementation of precision agriculture to manage and control citrus pests can effectively optimize pesticide use, reducing adverse environmental effects and ensuring human health. But managing individual trees on a large scale is a big challenge. Therefore, using a machine vision system seems necessary to monitor and identify pests in different partsof the trees at different times. In this study, an Unmanned Aerial Vehicle (UAV) equipped with a camera was used to identify pests in other parts of the citrus orchard. For the optimal selection of the UAV linear speed, three speeds in the range of 10, 20, and 30 cm/s were considered. After framing and formatting, the recorded videos were trained in three pretrained models: AlexNet, VGG-16, and GoogleNet. Three optimization algorithms were used in the network training process: SGDm, RMSProp, and Adam. The evaluation results showed that the AlexNet model, with the help of SGDm algorithm, had the best performance in terms of detection accuracy. The highest pest detection accuracy was 96.43% at a velocity of 10 cm/s, so increasing the linear velocity to 30 cm/s reduced the detection accuracy by 13%. The results of this study show that using a combination of UAV technology and artificial intelligence methods can help professionals and farmers manage and control citrus orchard pests.

Soil is one of the most important sources of production in agriculture. Therefore, with the determination of soil and its important characteristics, proper management and sustainable use of agricultural lands can be achieved. The current study aimed to predict the soil texture using a machine vision system and deep convolutional neural network (DCNN) algorithm. The proposed CNN model was composed of two blocks, including convolutional layers, max pooling layers, a dropout layer, batch normalization layers, fully connected layers, and a support vector machine classifier. This model was trained and tested on the images of different soil samples (11 types of soil texture and a total of 790 soil sample images). The data is prepared by a machine vision system and a smartphone camera (Galaxy A8). Using the confusion matrix, important statistical parameters such as accuracy, precision, specificity, sensitivity, and area under the curve were obtained at 99.65%, 98.75%, 99.8%, 98.75, and 99.27%, respectively. The suggested model successfully and correctly classified the soil sample images with 98.1% accuracy. The obtained results indicated that this study's implemented deep learning model can be a proper alternative to costly and time-consuming laboratory methods for determining soil texture.

Identifying fish species is critical for aquaculture and fishery industries, managing aquatic stocks and environmental monitoring of aquatics. In this study, deep learning neural network as a non-destructive and real-time approach was developed and used to identify four economically important species of carp family including common carp, grass carp, bighead carp and silver carp. For this purpose, the architecture of pre-trained VGG19 (Visual Geometry Group-19) was updated by pooling, fully-connected, normalization and dropout layers. 409 images were used for training and evaluating the developed model. The mean value of accuracy, precision, sensitivity, specificity and AUC parameters was calculated as 98.39, 96.87, 96.87, 98.96, and 97.92%, respectively. The obtained high level of accuracy is due to the ability of the proposed deep model in constructing a hierarchy of self-learned features which was consistent with the hierarchy of fish identification keys.

One of the methods of assessing the performance of seed drills may be to compare the performance with the crop population growth. Pixels of crop emergence zone appear to have similar characteristics concerning image parameter variations between soil and crop. The use of deep learning methods based on convolution neural networks to map regions of interest in the image seems appropriate. In this regard, a total of 2720 images of early-growth cereals were obtained from a field. 212 images with different backgrounds were selected and annotated to feed and train a neural network model. Raw images were defined as inputs and maps of manually marked growth points as network outputs. In order to calculate the network cost, the predicted output of the network was compared with the pre-marked pixel map. Prediction errors were then back-propagated and the network parameters updated. Examination of the initial network output showed that the trained network had responded incorrectly to plant tips, weeds and plant remains as plant growth points. To overcome these errors and improve network performance, a penalty function was defined for the mistaken predicted points. The network was trained with three penalty rates and evaluated with nine Soft Max thresholds. According to the network output, images were arranged in terms of plant density. In order to evaluate the model in different ranges, images from each particular range were selected at random. These images were fed to the model and their outputs compared with the truth. For the ranges where approximately 94% of the total field images existed, the average harmonic accuracy of the precision index and the recall index was estimated to be over 80%, indicating good model performance. The results showed that the model can provide acceptable feedback on sowing performance and improve farm management and efficiency in the next steps.