Non-Invasive Turkey Body Weight Monitoring and Prediction via Deep Visual Time Series Analysis
PI: Enrico Casella (Animal Science)
Plan for funding tuition for graduate students, or the remainder of the researcher’s salary for postdoc and research faculty: PI’s start-up funds and/or departmental/college support.
Introduction
Accurately monitoring the growth and predicting the final yield of poultry are critical for optimizing farm efficiency and profitability while improving the American food security challenged by the threats of avian influenza on the entire food system. Current practices to monitor growth often rely on periodic manual weighing of a sample of birds to estimate flock averages. This method is labor-intensive, time-consuming, and can introduce stress to the animals. Furthermore, relying solely on flock averages can mask individual variations in growth, hindering the ability to identify and address potential issues early on and impacting flock uniformity, which is a key factor in processing efficiency and product quality. The availability of continuous visual data streams, such as overlapping color and depth images, offers a non-invasive and potentially more granular alternative for monitoring animal development. This project aims to develop a novel hybrid deep learning model that leverages longitudinal visual data, potentially combined with historical flocklevel time series information, to estimate current body weight, predict future body weight trajectories, and ultimately forecast final carcass weight in turkeys. By moving beyond reliance on labor-intensive sampling and flock averages, this approach has the potential to enable more precise and timely insights into individual animal development and flock uniformity, without requiring individual animal identification by the model during inference.
Project Description
This project will explore the feasibility and efficacy of using a hybrid deep learning architecture for visual feature extraction coupled with time series analysis capability, such as recurrent neural networks (RNNs) or Transformer architectures to predict turkey body weight and carcass weight. The model will be trained on a unique dataset comprising three months of synchronized color and depth images captured at a rate of one frame per second, along with corresponding body weight measurements taken 5 times a week, and final carcass weights. The model should also take as input 4-channel images to exploit the combined benefits of color images (three channels) and depth images (one channel).
The core of the project involves developing a model that can: (1) extract relevant visual features from the color and depth images that correlate with body weight; (2) learn temporal patterns and dependencies in these visual features over time; and (3) integrate these learned representations to estimate current body weight and predict future growth trends. Given the lack of individual animal tracking by the model during inference, the time series component will primarily focus on capturing the overall flock dynamics and how visual features evolve at a group level. The availability of historical flock-level data may further enhance the model’s predictive capabilities.
An ICDS junior researcher would be responsible for:
• Developing and implementing a custom deep learning model (e.g., CNN + LSTM, RNN, Transformer) using a suitable deep learning framework (e.g., PyTorch, TensorFlow).
• Designing an appropriate architecture to process the image data and the time series of visual features.
• Training and validating the model on the provided dataset of turkey images, body weights, and carcass weights.
• Evaluating the model’s performance in terms of accuracy for current weight estimation, future weight prediction, and final carcass weight prediction.
• Compare a predictive model using exploiting historical data as a time series with one that only relies on instantaneous information.
• Investigating the potential benefits of incorporating historical flock-level time series data.
• Exploring the interpretability of the learned features and temporal patterns.
Expertise/skills of interest
• Deep Learning, Computer Vision, Time Series Analysis.
• Experience with convolutional neural networks (CNNs), recurrent neural networks (RNNs) (e.g., LSTMs, GRUs), and/or Transformer networks.
• Programming experience using Python and deep learning frameworks such as PyTorch or TensorFlow.
• Familiarity with data preprocessing, model training, and evaluation techniques.
• Experience with handling image and time series data.
Expectations:
• Graduate student or postdoc with at least some experience and/or training in Computer Science, Data Science, Computer Engineering, or a related discipline.
• Write well-organized and documented Python code achieving the aims of the research, along with documentation.
• Writing the manuscript to publish results, or at least, create graphs and tables to clearly show the obtained results to an academic audience.
• Participate in weekly meetings and provide regular project updates to the faculty advisor. Engage in group meetings as scheduled.
Goal
To determine the feasibility and accuracy of using a hybrid deep learning model fed with longitudinal visual data to effectively estimate current turkey body weight, predict future body weight trajectories, and forecast final carcass weight, without relying on individual animal identification during inference. The project aims to generate preliminary results that demonstrate the potential of this approach for improved poultry monitoring and management, potentially leading to future funding proposals focused on precision agriculture and animal welfare.
Specific Objectives:
• Develop a minimal working example of a custom hybrid deep learning model that couples visual feature extraction with time-series data to process sequences of color and depth images to predict current turkey body weight.
• Evaluate the model’s performance in predicting future body weight at different time horizons using the longitudinal image data.
• Assess the model’s ability to predict final carcass weight based on the visual growth patterns observed throughout the three-month period.
• Optional: explore the relative contributions of color and depth information to the model’s predictive accuracy.
• Optional: assess whether foundation models that estimate depth from color images (e.g. Depth Anything) can perform similarly to real depth images.
Engagement:
Dr. Casella is an Assistant Professor of Data Science for Animal Systems in the Department of Animal Science. He is also a co-hire of the Institute of Computational and Data Sciences (ICDS).