Recipient Organization
SOUTHERN ILLINOIS UNIV
(N/A)
CARBONDALE,IL 62901
Performing Department
(N/A)
Non Technical Summary
This project includes two main parts: education and research, and its discipline code is "A". The multidiscipline research part addresses an animal science challenge; Sub-acute ruminal acidosis (SARA). It is a common metabolic disease in an intensive cattle farm production system. Cattle with SARA experience significant drop in their performance (e.g. feed intake, milk production and weight gain), health (laminitis, organs inflammation, and tissues damages) and welfare. It has been estimated that the annual losses in the U.S. dairy industry due to SARA is $1 billion. The early detection of SARA is important as it would allow cattle and dairy producers to correct the cause for this metabolic disease (e.g. ration composition, soring, slug feedings,..etc) and therefore minimize the disease impact on animals. The current techniques used to early detect SARA are traditional and most of them have several limitations (e.g. accuracy and applicability) that limit their use under a traditional farming system. Artificial Intelligence (AI) and computer vison (CV) have incredible potential in developing myriad of applications for various sectors such as agriculture. In the research activity of this project, AI and CV will be used to help in developing models that are capable to early detect SARA using the gasses composition naturally emitted by ruminant animals. Therefore, this project will 1) use gas infrared cameras (IR) to collect datasets of digital images of emitted gasses (CO2 and CH4) under different rumen pH conditions using as in vitro system as a model for ruminant animals, and 2) develop deep convolution neural networks (CNN) to generate deep SARA detection models (saraDeepG2M). These models will be used to develop low-cost solutions that can early detect SARA in real-time. The education part includes two activities to organize educational workshops to: 1) incorporate the AI in the animal sciences for animal science students, and 2) disseminate the recent and future AI techniques in the animal sciences to the relevant societies.
Animal Health Component
(N/A)
Research Effort Categories
Basic
100%
Applied
(N/A)
Developmental
(N/A)
Goals / Objectives
This research hopes to employ recent and future technologies for cattle management and production. Artificial Intelligence and CV have incredible potential to transform key areas of business and develop myriad of applications for various sectors including agriculture. In recent years, scientists have been exploring the use of AI and CV in animal farming systems to improve the efficiency of animals' production, improve animals' health and reduce their impact on the environment. This proposal has three main objectives: 1) developing multidiscipline joint-research activities to promote cattle health and production, 2) conducting educational activities to promote the incorporation of AI in cattle farming and animal science educational programs, and 3) disseminating the recent and future AI techniques to the relevant societies and agencies.Developing multidiscipline joint-research activities to promote cattle health and production:Developing a real-time, efficient, and automated CV and DL models for the early detection of SARA. We will be using the emitted gas composition (CH4 and CO2) from the rumen, using an in vitro system, as variables to develop CV and DL algorithmic models capable of identifying SARA at an early stage using an automated real-time system.Study the emitted gas composition and ratio (CH4 and CO2) under different rumen pH levels with the aim to identify the trade-point between gas composition and the development of SARA.Conducting educational activities: The advances in AI offer significant opportunities to explore how cattle production systems may benefit from these recent technologies. This project will conduct two training workshops to promote AI development skills to animal science students, agriculture teachers, and extension personals.Disseminating these technologies to agriculture societies (e.g., dairy and beef cattle livestock) and extension agencies.
Project Methods
The emitted gasses will be measured using six single flow continuous fermenter system. Each fermenter will be incubated with 700 ml of rumen liquor, collected from two fistulated cows fed a diet consisting of 50% forages and 50% concentrate mix. All fermenters will be fed the same diet fed to the fistulated cows at the rate of 45 grams/day (DM basis) divided equality into three feedings. Fermenters will however vary in their pH level averaging 6.5 (treatment 1), 6.2 (treatment 2), 5.9 (treatment 3), 5.6 (treatment 4), 5.3 (treatment 5), and 5.0 (treatment 6). Fermenters pH will be maintained by the automatic infusion of 3 M HCl or 5 M NaOH as described by. The pH in fermenters will be monitored and recorded continuously using a digital pH meter device inserted in each fermenter. The experiment will have 4 periods and each experimental period will be 15 days long where the first 12 days are for diet (pH) adaptation while the last three days for data collection. The digital images of the emitted gasses (CH4 and CO2) from each fermenter gas port will be collected during the last 3 days of each period by an infrared optical gas imaging camera (IR). The IR camera will collect dataset during the first 4-hours after both the morning and afternoon feedings on the last 3 days of each period. The total time to collect the data is 3 days × 4 hours × 2 times (morning and afternoon) are equals to 24 hours. The IR will collect 20-min video per hour. The expected video frame rate is ~2 frames/sec. Each frame has a dimension of 240×320 grayscale image. In each experiment about 115K frames will be recorded. The expected total number of frames after N experiments is N ×115K frames. In addition, laser gasses detectors (LMD) will be used to measure gases (CH4, and CO2) concentration values from distance of 0.5-2m and define M size classes (including zero) of gases with different flow rates. The collected values from the LMD will be used to label the frames/images taken by the IR camera. Each diet (pH level) will be tested for N incubation periods, while gasses reading will be stored in a high-performance computer to be used in the data analysis, and model training stage. To make the training process accurate and faster, a background subtraction process will be applied to each frame. It creates training images with gasses plume only and no other objects. The collected dataset will be divided into three parts (training data 70%, validation data 20% and test dataset 10%).We will then design and develop several convolutional neural networks (CNN) to train the collected datasets. Python and Pytorch framework will be used to design and develop DL algorithms for detecting and localizing gasses plumes such as two-stages techniques RCNN, Fast RCNN, Faster RCNN, and one-stage techniques such as YOLO and SSD. In addition, RoI pooling layer will be used to reshape them into fixed size to fed them into a fully connected layer. Finally, SoftMax layer will be used to predict the class of the proposed region and the offset values for the bounding box. As well, YOLO (You Only Look Once) will be used to detect gasses plume where it uses single CNN to predict the bounding boxes and the class probabilities for these boxes (gasses measurement). Training processing time depends on the dataset size, CNN network, and the training server capabilities (memory, GPUs, and CPU).