Abstract:
With the growing prevalence of challenges in global food systems, including restricted access to fresh produce in remote areas and issues faced within traditional cultivation practices, outlooks on controlled environment agriculture (CEA) offer prospective solutions through precision methods of localized food production. Controlled environments within crop production systems provide opportunities to observe and cultivate plants with diverse characteristics, or phenotypes, making it suitable for specialized applications from trait selection for cultivar development and biopharming to food production in extreme and resource scarce environments. Subsequently, the ability to accurately predict and represent the phenotypic expression of a plant and its trajectory at various points in its growth cycle demonstrates a fundamental step towards advancing simulation and modeling research within CEA. In this talk we discuss the predictions and representations of plant growth patterns in simulated and controlled environments that are important for addressing various challenges in plant phenomics research and focus on the spatiotemporal modeling of plant traits and the integration of dynamic environmental interactions. We provide an examination of deterministic, probabilistic, and generative modeling approaches, emphasizing their applications in high-throughput phenotyping and simulation-based plant growth forecasting. Key topics include regressions and neural network-based representation models for the task of forecasting, limitations of existing experiment-based deterministic approaches, and the need for dynamic frameworks that incorporate uncertainty and evolving environmental feedback.
Bios:
Mark Lefsrud
Dr. Mark Lefsrud is an Associate Professor at McGill University and leads the Biomass Production Laboratory. His upbringing on a farm and work in the oil fields of Alberta, Canada combined with his B.Sc. and M.Sc. in Agricultural and Bioresource Engineering and a Ph.D. in Plant Physiology gives him a very strong background in the fields of agriculture, biology, and engineering. His research program deals with the development of bioprocesses and improvements in plant growth and environmental energy usage.
The laboratory is focused on four areas: 1) The development and improvement of new sources of biomass (food, fibre and/or fuel); 2) The improvement of energy efficiency of greenhouses and plant growth environments (light (LEDs) and heating systems); 3) The development quality practices for cannabis production; and 4) Development of monitoring techniques for plants and microorganisms using machine vision, nutrient monitoring, proteomics and metabolomics. His overall research philosophy is a holistic one in which focus on individual facets of an issue leads to a solution to the problem as a whole.
Mohamed Debbagh
Mohamed Debbagh is a Ph.D candidate at McGill University under the co-supervision of Professors Mark Lefsrud and Shangpeng Sun. He holds a B.Eng and M.Sc in Bioresource Engineering and Precision Agriculture and is driven by finding solutions to challenges faced within multi-scale food systems. His industry background covers sensor development, sensor fusion methods and robotics for various controlled environment producers. His Ph.D research focuses on the application of probabilistic computer vision and sensor fusion techniques for assessing plant development to improve crop throughput in controlled environment agriculture.
Summary:
Focus: Plant Phenomics in Controlled Environment Agriculture
Trait discovery and quantification
Genotype-phenotype mapping
Trait-environment interactions
Phenomics:
Data-centric
Complex data structures based on high-throughput phenotyping
Quantification at static states
Limitations:
Little information about the dynamic behavior of plant growth and its causal drivers
Makes it difficult to predict impacts of interventions or ultimate plant phenology / crop yield
Objective: predict future spatial attributes of plant given preceding time series of images
Typical approach:
Captured sensor data
Extraction of key traits (leaf area, count, etc.)
Train a development model from this data to forecast plant development
Characteristics: experiment-centric, fixed set of parameters, deterministic
Limitations:
Insight into plant growth is shallow
Doesn’t incorporate auto-regressive factors
Does not enable long-term forecasts
Poor uncertainty quantification and difficult to incorporate new information
Alternative: Probabilistic modeling
Characteristics: model-centric, generative, handing long-term objectives, probabilistic description of space of possible futures
Experiment:
Population of basil plants over 2 months under 4 different growth conditions
500k images of growth
Sensors: environmental, nutrient availability, lighting/irrigation schedules
10 min sampling frequency
Initial approach: Recurrent Convolutional neural network
Discrete in time, which limits ability to capture internal plant dynamics
Proposed model: Conditional Generative Model
Conditional Variational Autoencoder
Input: plant images conditioned on sequential information and sensor data
LSTM encoder: encoding of the prior sequence context
VAE encoder-decoder, conditioned on encoded sequence context vector
Decoder predicts next image
Produces a time series of images, can keep sampling time series from the model to create a full idea of the probability distribution
Loss term balances regularization (internally consistent accuracy) and reconstruction (more deterministic output)
More focus on reconstruction produces image time series that are point clouds that aggregate over many futures
More focus on regularization creates many different predictions that are internally more realistic and consistent
Possible to
Interpolate between time points to create a more temporally refined time series
Control treatment parameters and predict outcomes of different treatments
Challenges:
Lack of quantitative metrics for assessing quality of predicted morphology
Metrics
Generation quality
Frame coherence
Qualitative Observations by plant experts: evaluate quality of branching structures, root shapes, etc.
Data Requirements & Interpretability
As we increase the number of trials and growing conditions the amount of data increases quickly
Model training would be better if data represented experimental interventions, which are more independent and can produce a better model
Limited availability of good spatiotemporal datasets
Prospects for future work:
Improvement for generative models
Incorporating of knowledge-informed mechanistic models
Establishment of quantitative morphological metrics
Establishment of spatiotemporal crop datasets
Applications:
Controlled Hydroponics Imaging Platform
Exploring the divergence of plant behavior across diverse conditions (e.g. plants on the moon)
Identifying and explaining plant signals based on various stimuli
Sim2Real framework to control plant growth via reinforcement learning