Search

The unavailability and variability of training samples are the two essential concerns in the training of deep neural network models for image classification. For automated image monitoring systems, these problems are difficult when training a model through supervised learning methods because of the time and effort required. This paper proposes an adaptive solution to this problem by applying online semi-supervised learning to an automated insect pest monitoring system. The method used includes unsupervised pseudo-labelling of insect images and the training of semi-supervised classifier models for insect image recognition. The pseudo-labelling algorithm includes three major components: image labelling, label reconfirmation, and sample cleaning. Experiments were conducted on two unlabelled 1-year insect image datasets to evaluate the efficacy of the proposed method. It was found that the pseudo-labelling algorithm could achieve accuracy up to 0.963, hence enabling automated training data collection. The temporal improvement of the insect recognition performance by including new training data to retrain the classifier model was comparable in performance to the supervised learning approach as evaluated by cluster density, silhouette score, and F1-score. The proposed method was also able to automatically collect quality samples and train models regardless of the complexity of the images, making it a good alternative to replace laborious supervised learning. The proposed method can prevent contamination of a training dataset when images from new locations are collected. The presented techniques may also be used in other continuous learning applications that require automated training data collection and online model update.

Main bottleneck in facilitating integrated pest management (IPM) is the unavailability of reliable and immediate crop damage data. Without sufficient insect pest and plant disease information, farm managers are unable to make proper decisions to prevent crop damage. This work aims to present how an integrated system was able to drive farm managers towards sustainable and data-driven IPM.A system called Intelligent and Integrated Pest and Disease Management (IIt was concluded that novel and efficient strategies can be achieved by using an intelligent IPM system, opening IPM to potential benefits that cannot be easily realized with a traditional IPM program. This is the first work that reports the development of an intelligent strategic model for IPM based on actual automatically collected long-term data. The work presented herein can help in encouraging farm managers, researchers, experts, and industries to work together in implementing sustainable and data-driven IPM. © 2022 Society of Chemical Industry.

To counteract heat stress in dairy cows, a more reliable and efficient method for monitoring the activity of dairy cows and ambient environmental conditions should be developed. This research presents a cost-effective embedded imaging system that is capable of monitoring the drinking behaviour of dairy cows, while ambient temperature and humidity are simultaneously and continuously recorded with the integrated sensor modules. The embedded imaging system was implemented and tested on an experimental dairy farm, with imaging modules installed above the drinking troughs to collect video streams. To estimate the drinking time and frequency of dairy cows, detections of the dairy cow's head over the drinking troughs was performed on video stream using a deep learning convolutional neural network (CNN) model. The F1 score and true positive rate of the cow head detection were 0.987 and 0.983, respectively. The drinking behaviour data and environmental conditions were recorded and analysed to further assess the effects of heat stress on the drinking behaviour of dairy cows. The experimental results show that the daily total length and frequency of drinking bouts of the dairy cows were highly related to the temperature and humidity index (THI). Data from long-term monitoring using the automated imaging system clearly demonstrated that drinking behaviour reflects the effects of heat stress on dairy cows. The proposed monitoring system offers a novel approach for automatic and quantitative assessment of the drinking behaviours of dairy cows.

This work presents an automated insect pest counting and environmental condition monitoring system using integrated camera modules and an embedded system as the sensor node in a wireless sensor network. The sensor node can be used to simultaneously acquire images of sticky paper traps and measure temperature, humidity, and light intensity levels in a greenhouse. An image processing algorithm was applied to automatically detect and count insect pests on an insect sticky trap with 93% average temporal detection accuracy compared with manual counting. The integrated monitoring system was implemented with multiple sensor nodes in a greenhouse and experiments were performed to test the system’s performance. Experimental results show that the automatic counting of the monitoring system is comparable with manual counting, and the insect pest count information can be continuously and effectively recorded. Information on insect pest concentrations were further analyzed temporally and spatially with environmental factors. Analyses of experimental data reveal that the normalized hourly increase in the insect pest count appears to be associated with the change in light intensity, temperature, and relative humidity. With the proposed system, laborious manual counting can be circumvented and timely assessment of insect pest and environmental information can be achieved. The system also offers an efficient tool for long-term insect pest behavior observations, as well as for practical applications in integrated pest management (IPM).

The application of deep neural networks in computer vision has remarkably improved the reliability of automated insect pest identification algorithms. However, to build a robust deep neural network classifier model, sufficient number of image data is necessary. In preparing image data, collection and annotation of large datasets can be labor-intensive and time-consuming. This work aims to reduce the effort and time needed to perform such tasks. In this research, a method for synthesizing insect pest training images in sticky paper trap images and enhancing the performance of insect pest convolutional neural network (CNN) classifier models through a generative adversarial network (GAN) is proposed. Through GAN, images were synthesized to balance and increase the number of training images even with limited data. The fidelity of the synthetic images was tested using t-SNE visualization. To evaluate the developed network, the minimum number of images required to apply GAN-based augmentation was determined. Furthermore, the performance of the insect pest classifier models without augmentation, with traditional augmentation, and with GAN-based augmentation were compared. Testing results show that the classifier models trained with images augmented using GAN achieved an F1-score of 0.95, outperforming the models trained with traditionally augmented images with an F1-score of 0.92. This research shows that by using GAN-based augmentation method, the insect pest CNN classification models can yield a better performance compared to classic augmentation method and reduce the effort needed for data collection. Furthermore, the proposed approach can be easily adapted to various agricultural applications using deep neural networks when image data is limited.

Greenhouse whitefly (Trialeurodes vaporariorum) is a major insect pest of greenhouse crops. To prevent the damage caused by whiteflies, farmers control the population of whiteflies by spraying pesticides in a regular basis. However, pesticides are costly and may affect the environment and health of farmers. To provide a more efficient way for applying pesticides, this research aims to develop a model for predicting the possible increase in whitefly population in greenhouses using autoregressive integrated moving average (ARIMA) and ARIMA with exogenous variables (ARIMAX). The data used in this work were collected using wireless imaging devices that can monitor the number of whiteflies trapped on sticky paper traps using an automatic insect counting algorithm. The wireless imaging devices were installed in a greenhouse that grew tomato seedlings, which is one of the host plants of whiteflies. The ARIMA and ARIMAX models were compared by setting different combinations of input data. Particularly, ARIMA includes only the whitefly count while ARIMAX includes the whitefly count and environmental data. Based on preliminary testing, the minimum number of input data was found to be at around 60 days to 90 days. ARIMAX was found to be the best model with input data including the increase in whitefly counts, temperature and humidity. In average, the RMSE for 7-day forecasting of the proposed method was found to be around 1.30. To assist farmers in decision-making for pesticide application scheduling, four levels of increase in whitefly count were defined such as Normal, Moderate, High, and Critical, which were determined using K-means clustering algorithm, and testing results on a testing dataset show an F1-scores of 0.86 and 0.42 for Normal and Moderate levels of daily increase in whitefly count.

Dairy cows often suffer from heat stress problems due to extreme levels of humidity and temperature. Heat stress contributes to the decrease in the feed and drink intake, fertility, breathing rate, and milk production of dairy cows. With the use of modern technology, these behaviors can be automatically monitored by farm owners, thereby ensuring the health of the dairy cows. This research presents a lightweight algorithm for cow face recognition tailored for edge computing application. The proposed algorithm was implemented in an automated dairy cow feeding behavior monitoring system made up of embedded imaging devices. By edge computing, the system can be installed in a dairy farm with improved scalability, efficiency, and data security. A lightweight cow face image recognition convolutional neural network (CNN) model was optimized and trained using few-shot learning (FSL) with a testing accuracy of 0.90. A method in FSL, called embedding learning, was used to enable the cow face recognition model to adapt based on newly acquired training samples. Embeddings were generated that represent lower dimension vectors extracted by the model from the cow face images. In the reduced dimension, the L2 distance of each embedding represented the similarity of each image sample with the support of triplet loss. This research overcomes the non-convergence issue in model training through adaptive training methods that can create a similarity space between embeddings. The techniques in this research may also be applied in other fields that require adaptive face recognition methods.

With the advances of computer vision over the recent years, its applications have extended from handling static images to tackling the challenging domain of video analysis, with human activity recognition being among the core research topics. However, recognizing and recording the activity of agricultural workers through surveillance videos is still understudied. By workload efficiency analysis, the quality of agricultural produce can be determined and improved. In this work, an IoT system that utilizes MQTT messaging protocol to stream visual data collected from wireless monitoring nodes was developed. The video streams were displayed in a mobile APP for farm owners to remotely monitor their workplace and crop situation. A spatiotemporal convolutional network was trained to recognize the actions of the workers from the video streams, while the timestamp and duration were recorded in a remote server. Through vision-based analysis, workload data can be seamlessly extracted from the streaming footage, thereby eliminating the need for additional hardware setup, such as wearable sensors. The proposed algorithm was capable of achieving an average F1 score of 96.9%. The workload data analysis results could be beneficial for farm owners in reevaluating and improving their management strategy. Accurate detection and recording of farm activities will allow them to see if a particular activity had been performed appropriately and at the proper time. Furthermore, the usage of resources, including pesticide and water, can be tracked based on the frequency and duration of the activities. The proposed techniques and methods can be employed not only for agricultural applications, but also for other workplace monitoring applications.

The inability to classify insects up to the species level is the current limitation of recently developed automated insect recognition algorithms. Yet, there are situations in integrated pest management (IPM) that call for more precise classification of insects. This research proposes an effective method for automated identification of low-resolution insect images found on sticky paper traps. Insect image samples were collected from sticky paper trap images that were acquired by wireless image sensor nodes installed in an outdoor mango orchard. A tree-based classifier, made up of taxonomically cascaded convolutional neural network (CNN) models, was used to automatically identify the insects on each sticky paper trap from coarse to fine taxonomic levels. The proposed tree-based classification method can classify the insect images up to an F1-score of 0.94, surpassing the performance of an optimized single multi-class image classifier model with an F1-score of 0.87. The proposed method not only boosts the classification performance, but also offers hierarchical predictions that can be used as easily interpretable information. In addition, model hyperparameter optimization, data augmentation, and classification probability threshold tuning can be applied at each level to optimize its performance. This research can be applied to support farmers in selecting IPM components, such as crop variants, pesticides, parasitoids, and more, in order to control the population of different insect species and prevent production loss.

Bone, muscle, and spine diseases caused by workplace tasks are called work-related musculoskeletal disorders (WMSD). Agricultural workers are most likely to suffer with WMSD due to long working hours and intensive manual work. The purpose of this study is to develop and assess a waist-assistive exoskeleton for agricultural workers to reduce the risk of WMSD and improve work efficiency. An ergonomic evaluation system which consists of wearable inertial measurement unit (IMU) sensors was developed to assess work quality in agricultural tasks. The evaluation system was used to conduct three ergonomic assessments: rapid upper limb assessment (RULA), rapid entire body assessment (REBA), and agricultural whole-body assessment (AWBA). The assessment of these four common agricultural tasks in Taiwan revealed that stooping entailed the highest risk of WMSD. Together with the evaluation system, MonoBearExo, a novel and lightweight wearable exoskeleton, was developed to reduce the risk of WMSD. It was designed to support the waist of agricultural workers when performing stooping and repetitive lifting tasks. The exoskeleton, with a total weight of only 5.5 kg, is driven by a hybrid-powered system which consists of electrical rotation motors and a mechanical energy storage system. The system has a maximum output torque of 53.3 Nm for supporting bending motion via cables, and driving both legs. The implemented motion recognition model for precise heavy object lifting detection can achieve an FR1R-score of 0.92; it was found to be suitable for several individuals. Compared to recently developed waist-assistive exoskeletons, the proposed design has long battery life and offers improved comfort in working conditions.

Currently, global warming is worsening, causing the difficulty of cultivating crops in open fields, and leading to unstable quality of crops. Plant factories provide a well-controlled growth environment for precisely cultivating plants. However, uneven temperature distributions (UTDs) still occur at each cultivation shelf in plant factories, which decreases the yields (fresh weight) of plants. In this study, a wireless sensor network (WSN)-based automatic temperature monitoring and fan-circulating system for precision cultivation in plant factories is proposed, and it is built upon the technologies of WSN, ordinary kriging spatial interpolation, and automation control, to precisely find the UTD areas of cultivation shelves. Once a UTD area occurs, the fan-circulating system can be triggered immediately to automatically trace the area and circulate the air. This action can effectively improve the air flow in the cultivation zone, providing optimal growth conditions for plants. The proposed system has been deployed in two plant factories that grew Boston lettuces, and a series of performance evaluation experiments were conducted. The experimental results indicate that the fresh weight of the harvested lettuces increases by 61–109% when employing the proposed system that efficiently and significantly decreases the variation of the temperature in the cultivation zone.

Pollen foraging efficiency provides vital information for the behavioral research on honey bees. The pollen production of beehives can be measured by manually weighing the pollen collected from pollen traps. For long-term pollen foraging monitoring, this approach is both inefficient and laborious. This study presents an efficient method for automatically monitoring the pollen foraging behavior and environmental conditions through an embedded imaging system. The imaging system uses an off-the-shelf camera installed at the beehive entrance to acquire video streams that are processed using the developed image processing algorithm. A lightweight real-time object detection and deep learning-based classification model, supported by an object tracking algorithm, was trained for counting and recognizing honey bee into pollen or non-pollen bearing class. The F1-score was 0.94 for pollen and non-pollen bearing honey bee recognition, and the precision and recall values were 0.91 and 0.99, respectively. For foraging trip counting algorithm, the mean average percent errors of the pollen bearing honey bee count and the total incoming honey bee count were 8.45 ± 2.72% and 10.55 ± 2.10%, respectively. An experiment was performed to test the performance of the imaging system in continuous monitoring of honey bee pollen foraging behavior as well as to investigate the effect caused by weather factors. The incoming and outgoing honey bee count were recorded and used to calculate indices based on the hourly and daily recorded counts for further analyses. The experimental results and analyses revealed that the daily pollen foraging trip ratio was 24.5 ± 3.5%; a single beehive collected about 49.1 ± 11.0 g of pollen per day. The pollen foraging trip count increased with increasing temperature and light intensity, and decreased with increasing relative humidity, rain level and wind speed. A significant reduction of pollen foraging activities was observed in heavy rainfall or gentle breeze conditions. This study not only quantitatively presents the effect of environmental factors on pollen foraging behavior, but also demonstrates the efficacy of the proposed imaging system. The automated imaging system can be applied as an efficient and reliable tool for researchers to gain deeper insights into honey bee foraging behavior, and help beekeepers achieve beehive management.

In recent years, people have started to pay more attention to ergonomics in different fields of work, especially in agricultural workplaces in which most tasks cause discomfort or a decline in work efficiency. In particular, repeated and intensive waist motions may lead to back injuries, which is a very common complication caused by agricultural operations. This study proposes a system that combines the usage of inertial measurement units (IMUs) and an exoskeleton robot to automatically assess work performance and assist workers while performing several agricultural tasks. The system was optimized based on its weight, endurance, convenience, and cost. By placing three IMUs on the subject‘s trunk and legs, the system is able to control the exoskeleton robot to move the wearers‘ legs while supporting their backs. In order to evaluate the waist assistance provided by the system, ergonomic assessment tests were used, such as Rapid Upper Limb Assessment (RULA) and electromyography (EMG). In our research, by using the exoskeleton we built, the muscle activity of the Erector Spinae decreased by about 17.3% for lifting10 kg objects during stoop lifting and 20.4% for lifting 10 kg during semi-squat lifting, while the muscle activity of Biceps Femoris decreased by 18.2% and 15.2%, respectively, for the same tasks as above. This work can be applied not only to agricultural tasks, but also for rehabilitation, construction work, and other related work fields.

This paper proposes an online unsupervised deep learning approach for automatically collecting new image data for continuously training more accurate image classifier models of a multi-class insect identification algorithm. Once images are collected by the sticky paper trap image monitoring system, insect images are cropped from the images and screened out based on an unsupervised data collection technique. The proposed unsupervised data collection technique uses reference models trained with supervision for the collection of images based on different classification probability threshold settings. After testing 15 months of data, it was found that reference models trained with 8 months of image data could consistently perform unsupervised data collection with minimal collection error. After 7 months of the unsupervised online training cycle, the technique was able to improve the average F1-score of the classifier models from 0.88 to 0.926, close to the 0.94 F1-score achieved with supervision. The proposed technique reduces the effort required to train new image classifier models, as it does not require the traditional manual collection of new images. The presented technique can be applied to not only insect identification, but also other image monitoring applications.

Observation of animal behavior in dairy farms is necessary for precision livestock farming as well as to prevent ailments that may affect the herd. Heat stress in dairy farms has been reported to be a serious problem that leads to the drastic decline in milk production and fertility of dairy cows. In recent studies, wearable devices were used to monitor the behavior of individual dairy cows. However, such kinds of contact devices may affect the behavior of the dairy cows for several reasons. Therefore, designing a non-contact system that can monitor the behavior of dairy cows is recommended. In this work, an embedded imaging system that can monitor the feeding behavior of dairy cows was developed. The system includes cameras that were fixed in front of the feeding area to acquire cow face images. Cow face detection and recognition were performed on the acquired images using a deep convolution neural network (CNN) to record feeding behavior and identify individual cow faces. The cow face detector has an F1-score of 0.971 based on validation with a static image testing dataset. Meanwhile, the cow face recognition model was also validated with an average F1-score of 0.85 on 19 different cows. Finally, the predictions of feeding time were compared with the manual observation with the 0.7802 without the cows having lower recall from the recognition results. In the future, combining the feeding time and temperature and humidity index can be a health indicator of individual cows.

Hyperspectral imaging has proven to be an effective non-destructive method for assessing strawberry foliar anthracnose.The incubation stage, in which symptoms are not yet visible, can be distinguished.Several hyperspectral imaging analysis methods were investigated using 3 duplicate sets of experiments.Significant wavelengths for strawberry foliar anthracnose are 551, 706, 750 and 914nm. Hyperspectral imaging provides comprehensive spectral and spatial information about observed objects. This technology has been applied to many fields, such as geology, mining, surveillance and agriculture. Strawberry qualities have been examined using hyperspectral imaging in several studies. However, none of the previous literature presented a non-destructive method for diagnosing the infection stages of anthracnose, a devastating disease for strawberries. This study examined strawberry foliar anthracnose using three different hyperspectral imaging analyzing methods: spectral angle mapper (SAM), stepwise discriminant analysis (SDA) and self-developed correlation measure (CM). Three different infection stages, including healthy, incubation and symptomatic stages, were investigated using these methods. The incubation stage is a stage at which the symptoms are still not yet visible. The three infection stage classification results were promising, with a classification accuracy of approximately 80%. For two infection stage classification (healthy and symptomatic stages), an average accuracy of high 80% was attained. In fact, an average accuracy of 93% was achieved by SDA for two-stage classification. This study not only proves the feasibility of hyperspectral imaging for diagnosing strawberry foliar anthracnose infection, but also identifies a smaller set of significant wavelengths at which similar classification performance was accomplished. For significant wavelength selection, partial least squares (PLS) regression is an standard wavelength selection method and it was applied to be compared with SDA and CM. Wavelengths of 551, 706, 750 and 914nm formed the multispectral imaging observing bands that showed an accuracy of 80% when classifying the three infection stages. Therefore, using either hyperspectral or multispectral imaging to detect anthracnose infected foliar areas is more practical and efficient than classic destructive methods. In particular, early detection (the incubation stage), something that cannot be seen via naked eyes, reaches 80% classification accuracy with both SDA and CM. Strawberry farmers could profit greatly from this technology.

A mosquito monitoring system using the wingbeat pattern was proposed.A robust optical sensing device for wingbeat profiling was developed.Three algorithms for mosquito classification were assessed using three species of living mosquitoes.The average classification accuracy of the proposed algorithm is above 85%. We developed an automatic mosquito classification system which consists of an infrared recording device for profiling the wingbeat of the in-flight mosquito species and a machine learning model for classifying the gender, genus, and species of the incoming mosquitoes by the signatures of their wingbeats.The recording device is a set of infrared emitters and receivers, which are attached to the wall of an apparatus. When the winged subject enters the apparatus, its flapping wings block the infrared beam from the emitters intermittently such that the receivers convert the wingbeat to the electrical waveform. To classify the incoming subjects, we proposed a machine learning method, which is the Gaussian mixture model trained using the expectation-maximization algorithm (EM-GMM), and compared it with the previously proposed algorithms, including the artificial neuron network model (ANN) and the nearest neighbor model.To assess the performance of the system, we used the living male and female Aedes albopictus, Aedes aegypti and Culex quinquefasciatus. The results show that the accuracies of the proposed system are above 80% on identifying the gender and genus of the mosquitoes, with the precisions above 80% and 70%, respectively. The results also suggest that the EM-GMM algorithm outperforms the other two algorithms on the accuracy and precision of the classification of the classes of mosquitoes. In addition to the evaluation of the performance of the system, we also found that certain classes of mosquitoes share similar wingbeat characteristics, which implies that the distinctive wingbeat characteristics should be considered for the optimal accuracy of the classification of the insects of interest.

Shape variation within a flower breeding line is a topic of considerable interest to horticulturalists. Sinningia speciosa (Florist's Gloxinia) is a species that presents diversified floral shapes. This study aimed to quantitatively assess floral shape variation in an F2 cross of S. speciosa between a zygomorphic wild variety and an actinomorphic peloric cultivar via geometric morphometrics. The result indicated symmetric variation and tube dilation accounted for the major variance of floral shape changes. We further tested whether these shape variations can be correlated to any inherited genetic variation. ANOVA analysis detected candidate CYCLOIDEA (SsCYC) marker allele that showed strong associations with variations of corolla symmetry and tube dilation. This study demonstrated that using geometric morphometrics might considerably enhance the detection of phenotypic and genetic association on complex floral shape variation.

STFT-ranking feature is efficient for multi-channel EMG signal analysis.STFT-ranking feature can characterize relationships between multi-channel signals.Recognition accuracy over 90% was achieved applying the STFT-ranking feature.The performance of STFT-ranking feature is superior to conventional features.STFT-ranking feature can be applied to other multi-channel signals applications. Electromyography (EMG) is widely applied for neural engineering. For motion pattern recognition, many features of multi-channel EMG signals were investigated, but the relationships between muscles were not considered. In this study, a novel STFT-ranking feature based on short-time Fourier transform (STFT) is proposed. The novelty of STFT-ranking features is considering and covering the relationship information between EMG signals and multiple muscles in a motion pattern. With an exoskeleton robot arm, two series of motion patterns corresponding to the shoulder and elbow in the sagittal plane were investigated. EMG signals from six muscles were acquired in arm motion patterns when participants worn the robot arm. Four types of feature combinations, including seven conventional features, were compared with the STFT-ranking feature. The principal component analysis (PCA) and support vector machine (SVM) were used to build the motion recognition model. With the STFT-ranking feature, the recognition performance (93.9%) is superior to the conventional features (33.3-90.8%). The recognition variation is smaller (SD=4.3%) than the other features tested (SD=5.9-13.8%). These achievements will contribute to the advancement of control method of exoskeleton robots or power orthoses based on multi-channel EMG signals in the future. Based on the principle of STFT-ranking feature, the method also has potential for other multi-channel signal applications, such as electroencephalography (EEG) signal processing, speech recognition, and acoustic analysis.

Photovoltaic (PV) generation systems have been shown to have a promising role for use in high electric-load buildings, such as the closed plant factory which is dependent upon artificial lighting. The power generated by the PV systems can be either directly supplied to the buildings or fed back into the electrical grid to reduce the high economic costs and environmental impact associated with the traditional energy sources such as nuclear power and fossil fuels. However, PV systems usually suffer from low energy-conversion efficiency, and it is therefore necessary to improve their performance by tackling the energy loss issues. The maximum power point tracking (MPPT) control technique is essential to the PV-assisted generation systems in order to achieve the maximum power output in real time. In this study, we integrate the previously proposed direct-prediction MPP method with a perturbation and observation (P&O) method to develop a new hybrid MPPT method. The proposed MPPT method is further utilized in the PV inverters in a PV system installed on the roof of a closed plant factory at National Taiwan University. The tested PV system is constructed as a two-stage grid-connected photovoltaic power conditioning (PVPC) system with a boost-buck full bridge design configuration. A control scheme based on the hybrid MPPT method is also developed and implemented in the PV inverters of the PVPC system to achieve tracking of the maximum power output of the PV system in real time. Based on experimental results obtained in an outdoor environment over a period of one year, the hybrid MPPT method not only decreases energy loss but also increases power utilization. These results demonstrate the applicability of the PV system to a closed plant factory for saving energy consumption and reducing CO 2 emissions.

Multichannel electromyography (EMG) signals are one of the common methods used in human motion pattern recognition. In exoskeleton robot control, EMG signals are measured during dynamic or isometric muscle contractions. Various types of contraction can cause EMG signals to vary, affecting recognition performance. A motion pattern recognition model using EMG signals from either dynamic or isometric muscle contractions has not yet been fully investigated. In this study, a novel feature extraction method, using the short-time Fourier transform ranking (STFT-ranking) feature, was employed to determine multichannel EMG signals. The performance of the novel feature and conventional features for motion pattern recognition using EMG signals, which included time-domain and frequency-domain features, was compared during dynamic and isometric muscle contractions. Experiments were conducted using an exoskeleton robotic arm to aid users in generating EMG signals of designated motion patterns. Among the features tested, the STFT-ranking feature yielded an accuracy rate exceeding 90% when the EMG signals used in the training and validation feature data sets were of the same type of muscle contraction. After examining the STFT-ranking feature projected onto the PCA space, the STFT-ranking feature was determined to offer more satisfactory performance than the other features tested for motion pattern recognition, because the feature data it collected from various motion patterns were more separable. The experimental results also revealed that it is preferable that EMG signals from the same type of muscle contraction, whether dynamic or isometric, are consistently used in both the training and validation (control) phases. Inconsistent EMG signals in the training and validation phases yielded a negative effect on motion pattern recognition performance. The methodology developed in this study has potential applications in exoskeleton robot control and rehabilitation.

As plant diseases could cause agricultural production and economic loses, there is a need of fast and effective plant disease detection and assessment methods. Non-destructive methods have gained popularity among these methods as they do not affect plant growth while examining plant health conditions. Not only plant diseases can be detected but also production can be improved with proper quality controls. Hyperspectral imaging is one of the non-destructive examination techniques which have been widely applied in agriculture. Hyperspectral image analysis has been applied to different problems including plant disease detection and assessments. It provides not only spatial image but also spectral information of the observed object. This research has aimed to compare two hyperspectral image analysis methods: stepwise discriminant analysis (SDA) and spectral angle mapper (SAM) and the proposed Simple Slope Measure (SSM) method in strawberry foliage Anthracnose assessment. Anthracnose is one of the most devastating diseases for strawberries. Anthracnose disease can affect the whole plant and may result in 100 percent fruit loss from crown and fruit rot. Hence, an early detection of the Anthracnose disease will be beneficial to ensure production and quality of strawberries. This research has shown that the three different Anthracnose infection status (healthy, incubation and symptomatic) could be separated by the methods examined. The performance of these disease assessment models were evaluated and compared. The examination outcomes prove the feasibility to assess strawberry foliage Anthracnose nondestructively and as early as the symptoms not visible to naked eyes. As soon as early detections of the Anthracnose disease are achievable in the strawberry field, the damage to strawberry production due to the spread of Anthracnose disease could be reduced.

Mobile robots have already been widely applied in agriculture fields. In order to perform complex tasks such as localization, mapping and path planning, mobile robots need to understand their environment. Therefore, the perception ability is important to robots. In the past twenty years, Simultaneous Localization and Mapping has been a popular algorithm to solve localization and mapping problems. In agricultural applications, robots need to overcome bumpy surfaces and should avoid inducing destruction to plant growths. Hence, the assumptions of robot poses need to be represented by six degree of freedom, not just flat surface movements. In order to estimate the 6D poses, 3D information was acquired by our custom-made of 3D laser scanner. The ICP algorithm then merged different scenes together to reconstruct the whole environment. However, errors are accumulated in the ICP algorithm, so a global optimization algorithm was applied to deal with this accumulated error. We had successfully reconstructed two large scale environments that contain millions of points. The method we proposed here has a RMSE of 55.94 centimeters. It is believed that this system has different applications such as agriculture field investigation, virtual reality and geographic information system.

Plant factories are closed cultivation spaces where environmental conditions are controllable. In order to find the most suitable environmental conditions for plant growth, plant features can provide a good indication. Traditional direct measurement methods are simple but destructive and laborious. In this research, an automatic plant growth measurement system, which is composed of a stereo vision system and a weight measurement system, was developed. The growth of Boston lettuces were monitored and measured with our automatic system. The camera mounted on a sliding rail was moved by a linear actuator to extend the field of view of the planting bed. Using image processing methods and stereo vision techniques, panoramic images of the plating bed were constructed and plant geometric features such as projected leaf area, maximum and minimum radii and plant volume were calculated. At the same time, plant weights were continuously recorded by the weight measurement instrument. Both the vision-based system and weight measurement system are based on non-destructive methods, so the plants features can be monitored throughout the whole growth cycle without affecting its growth. The plant growth measurement system developed in this research can obtain a large amount of plant features easily comparing to traditional destructive measurement methods. Besides, the relationship between the projected leaf area and the fresh weight of Boston lettuces was also discussed in this research. The results have proved the feasibility and functionality of this integrated system. This system would provide valuable information for plant factory managements and agricultural research applications.

In plant factories, plants are usually cultivated in nutrient solution under a controllable environment. Plant quality and growth are closely monitored and precisely controlled. For plant growth evaluation, plant weight is an important and commonly used indicator. Traditional plant weight measurements are destructive and laborious. In order to measure and record the plant weight during plant growth, an automated measurement system was designed and developed herein. The weight measurement system comprises a weight measurement device and an imaging system. The weight measurement device consists of a top disk, a bottom disk, a plant holder and a load cell. The load cell with a resolution of 0.1 g converts the plant weight on the plant holder disk to an analogue electrical signal for a precise measurement. The top disk and bottom disk are designed to be durable for different plant sizes, so plant weight can be measured continuously throughout the whole growth period, without hindering plant growth. The results show that plant weights measured by the weight measurement device are highly correlated with the weights estimated by the stereo-vision imaging system; hence, plant weight can be measured by either method. The weight growth of selected vegetables growing in the National Taiwan University plant factory were monitored and measured using our automated plant growth weight measurement system. The experimental results demonstrate the functionality, stability and durability of this system. The information gathered by this weight system can be valuable and beneficial for hydroponic plants monitoring research and agricultural research applications.

This study aimed to develop an imaging system for monitoring and analyzing the in-and-out activity of honey bees as they pass through the entrance of a beehive. As such, the daily in-and-out activity of a beehive could be automatically recorded with minimum interference with the regular behavior of the bees. The components of the imaging system include: (1) an infrared LED light source; (2) an infrared CCD camera and a personal computer for image acquisition and processing; and (3) an acrylic passageway to temporarily confine the bees within the image-capturing range of the camera. In operation, this imaging system was affixed to the outside of a beehive at the entry/exit point. To identify each forager of bee workers individually, circular character-encoding tags were attached to the dorsum of the bees' thoraxes. To locate individual honey bees in a video frame, a circular Hough transform was used to detect the presence of the circular tags. A black positioning dot on the tag was used to identify the orientation of the characters in order to facilitate the reading of the symbols on the circular tag. The extracted character symbols were further segmented and a support vector machine (SVM) classifier was deployed to recognize the characters and identify the individual honey bee. The system developed in this study was used in experimentation to identify each of the tagged bees and to record the timing of the entries and exits. The character symbol recognition and identification accuracy rates of the system were about 98% and 86%, respectively. Based on the in-and-out records, daily foraging behavior of honey bees were analyzed and presented. These experimental results demonstrate that this imaging system is feasible and can be used as an efficient tool to study honey bee behavior.

During the ripening process, fruit and vegetables undergo physico-chemical changes which affect their color, texture, aroma, and sugar content, etc. For tomatoes, color change is the main index of ripening stages, and it is associated with metabolized lycopene. In order to detect the internal changes of sugar and lycopene contents in tomatoes during ripening, chemical shift imaging (CSI) and high performance liquid chromatography (HPLC) analyses were employed. Imaging experiments on tomatoes were performed in a Bruker 7T MRI system using the gradient-echo sequence for scanning. Spatial distribution of the internal sugar and lycopene contents of tomatoes were, for the first time, revealed and compared before and after the ripening process. The CSI analysis showed that a linear relationship existed between CSI intensity and concentrations of sugar solutions in the range of 0.1–1.0 M. The CSI results indicate that there was no consistent trend in the variation of sugar signal intensity during ripening while the lycopene content significantly increased in the outer pericarp and columella of the tomatoes at the red ripe stage, as compared to the mature green stage. This observation for the sugar and lycopene signal intensities was further confirmed by HPLC analyses. The non-destructive CSI technique has potential as an efficient tool for the physiological analysis of fruit and vegetables.

We herein introduce an automated three-dimensional (3D) locomotion tracking and pose reconstruction system for rodents with superior robustness, rapidity, reliability, resolution, simplicity, and cost. An off-the-shelf composite infrared (IR) range camera was adopted to grab high-resolution depth images (640×480×2048 pixels at 20Hz) in our system for automated behavior analysis. For the inherent 3D structure of the depth images, we developed a compact algorithm to reconstruct the locomotion and body behavior with superior temporal and solid spatial resolution. Since the range camera operates in the IR spectrum, interference from the visible light spectrum did not affect the tracking performance. The accuracy of our system was 98.1±3.2%. We also validated the system, which yielded strong correlation with automated and manual tracking. Meanwhile, the system replicates a detailed dynamic rat model in virtual space, which demonstrates the movements of the extremities of the body and locomotion in detail on varied terrain.

Highlights? A new automatic and effective fruit quarantine system is presented. ? Guava, pitaya, pear, apple, peach, mango and Sunkist are used for experiments. ? Optimal settings for X-ray tube are presented. This paper presents a new automatic and effective quarantine system for detecting pest infestation sites in agricultural products, e.g. fruits. This work integrated mechanical design, mechatronics instrumentation, X-ray and charge-coupled device (CCD) image acquisition devices, LabVIEW-based analysis and control software, and image diagnosis algorithms into the automatic X-ray quarantine scanner system. Based on the LabVIEW development platform, a friendly graphical user interface (GUI) was designed for assisting the operations of quarantine scanner system. To enhance the accuracy and efficiency of pest quarantine process, a control scheme for performing start-up procedure of the system, parameter setting and calibration of the X-ray source and line-scan sensor, and automatic inspection for pest were developed. A novel pest infestation detector consisted of image processing algorithms were also proposed to aid the operator in identifying possibly infested fruits. The image processing procedures include contrast enhancement, median filtering, mathematical morphology operators, and adaptive thresholding by statistical z-test for identifying the infested sites of fruit on an X-ray image. Experimental results show that the X-ray quarantine scanner and pest infestation detector are able to locate the infested sites with highly successful rate up to 94% on the 4th day after eggs implanted. Furthermore, both intact and egg-implanted fruits were used to evaluate the sensitivity, specificity, accuracy, and precision of the proposed system. The evaluation results are respectively 96.8%, 98.6%, 97.7%, and 98.7%, which are significantly better than traditional visual inspection.

Fish length measurement of commercial catches is an important data input to many fish stock assessments to appraise stock status and to provide management advice for sustainable fisheries. However, as the process of measurement is usually laborious or complex, it is difficult to operate on longline fishing vessels on deck. This study proposes an effective technique for measuring a tuna's snout to fork length (SNFL) in digital images or photographs taken on the deck of tuna fishing vessels by observers using hand-held camera. Photographic factors, such as direction angle, top-view angle, and distance to fish body were examined in terms of estimation error. Hough transform (HT) and projective transform (PT) were applied to conduct line detection in automatic mode and to correct for projective distortion of the fish images, respectively. Experimental results showed that the estimation error was within 5% when the image was taken with a direction angle of 315-0^o or 135-225^o, and with a top-view angle of more than 45^o. The PT correction reduced the estimation error especially in small top-view angle samples, with an average improvement of 64% and 48% in manual mode and in automatic mode, respectively. The proposed method was validated with the results showing that, after 30min of practice, an operator can successfully make a measurement within a minute while the average (+/-std. dev.) estimation error was 4.5+/-4.4% using the developed software program. With slight modification, the developed approach could be adapted to various applications concerning size measurement of plants or animals using photometric approach.

Length frequency data are essential to the stock assessments of albacore ( Thunnus alalunga ) which has been either fully- or over-exploited in many regions of the world. To obtain ‘visible’ and verifiable length data, a photo-based length measurement approach was previously developed in which fishers take photos of fish adjacent to a plastic calibration board with colored squares using ordinary digital camera technology. Here we assess the feasibility of implementing the photo-based approach using direct comparisons with measured lengths from observer programs. Data from the Taiwanese longline albacore fishery are used to assess the utility of the photo approach in terms of accuracy (with considerations of on-board photographing and at-land length estimation) and cost (comparing with port-sampling and observer programs). Results indicate potentially significant cost savings compared with the other two programs, with acceptable levels of estimation error.

Chemical shift imaging (CSI) relies on a strong and homogeneous main field. Field homogeneity ensures adequate coherence between the precessions of individual spins within each voxel. Variation of field strength between different voxels causes geometric distortion and intensity variation in chemical shift images, resulting in errors when analyzing the spatial distribution of specific chemical compounds. A post-processing method, based on detection of the spectral peak of water and baseline subtraction with Lorentzian functions, was developed in this study to automatically correct spectra offsets caused by field inhomogeneity, thus enhancing the contrast of the chemical shift images. Because this method does not require prior field plot information, it offers advantages over existing correction methods. Furthermore, the method significantly reduces geometric distortion and enhances signals of chemical compounds even when the water suppression protocol was applied during the CSI data acquisition. The experimental results of the water and glucose phantoms showed a considerable reduction of artifacts in the spectroscopic images when this post-processing method was employed. The significance of this method was also demonstrated by an analysis of the spatial distributions of sugar and water contents in ripe and unripe bananas.

This research involved a magnetic resonance imaging (MRI) technique that was applied to examine water distribution and migration in single rice kernels during the tempering process. The imaging experiments were performed in a Bruker 9.4T MRI system. Three-dimensional spin-echo (SE) imaging sequences were optimized by adjusting the scanning parameters of echo time (TE) and repetition time (TR) to obtain images with maximum contrast. The MR images showed that the moisture distribution in the rice kernel is non-uniform and compartmental. The embryo region exhibited much higher MR signal intensity than the starchy endosperm portion. The tempering process was analyzed with spatial-temporal signal intensities of the endosperm following the drying process of the rice kernel. The transient change of the signal intensities in the endosperm was well fitted with a double exponential function suggesting that both convection and diffusion contributed to the reduction of the moisture gradient within the rice kernel during tempering. This hypothesis was further supported by the experimental data of the insulated rice kernel whose convective mass transfer was excluded. The experimental results revealed that MR imaging of rice kernels could be used as an efficient tool to examine the mechanisms of moisture migration within cereal grains.

Chang, S-K., Lin, T-T., Lin, G-H., Chang, H-Y., and Hsieh, C-L. 2009. How to collect verifiable length data on tuna from photographs: an approach for sample vessels. – ICES Journal of Marine Science, 66: 907–915. Length frequencies are essential data for fish stock assessments, particularly for longer-lived species. They are usually provided by commercial vessels, or by port sampling, observers, or sample-vessel programmes, but each of these has limitations. Collection by sample vessels might be the most balanced way if the data quality is verifiable. We introduce a photograph-based length-measurement approach for sample vessels to photograph fish images with a calibration board, using a regular digital camera to obtain length estimates that can be verified after the images are transformed to reduce errors of perspective distortions. We analyse this approach under ideal conditions, develop a set of objective criteria for choosing acceptable photographs from observers, and compare estimated lengths for bigeye tuna (Thunnus obesus) based on this approach with lengths measured by observers. The criteria can serve as guidelines for photographing: if images are captured following these guidelines, the approach shows the potential for obtaining cheaply a large quantity of length estimates that deviate around 3% (up and down) on average from the actual measurements taken by observers.

The purpose of this study was to evaluate defects in trees using an ultrasonic tomographic technique. In order to simulate different degrees and kinds of defects, a hole was chiseled in the central and side areas of trunk cross-sections and then the ultrasonic velocities were measured in multiple directions (eight equidistant ultrasonic test points and 56 independent travel time measurements) and analyzed using ultrasonic equipment. Results revealed that the highest coefficient of determination was found between the residual disc diameter ratio and the ultrasonic velocity in the virtual radial direction. Strong correlations between the drop in the relative ultrasonic velocity and residual disc diameter ratio were observed in the virtual radial direction. Moreover, the ratio of ultrasonic velocity along paths D (a three-point interval between two measured points) to B (a one-point interval between two measured points) decreased with increase in the ratio of the hole-to-disc area. However, values became more diverse for specimens with a hole chiseled in the side area as the ratio of the hole-to-disc area increased. In situ experimental results showed that differences in the mapped colors and ultrasonic velocities between the hole area and its surrounding area were obvious, and when the ratio of the hole area to the cross-sectional area exceeded 2.8%, the hole was detectable by ultrasonic tomography. Therefore, the location and size of the hole in the tree trunk cross-section could be detected by this ultrasonic tomographic technique.