Integration of machine learning and 3D plant biophysical models for acceleration of agricultural system design, management, and crop improvement
Brian N Bailey @ UC Davis
Abstract:
Helios is a flexible, high-performance plant modeling framework that explicitly represents the 3D geometry and interactions between plants and their mircoenvironment, and is implemented as an open-source C++ API. The core engine of Helios manages geometry and data associated with 3D models, and acts as the model coupler. Model and utility routines are implemented as plug-ins to the core engine, which include, e.g., models of radiation transfer, energy balance (for computing temperature and water use), photosynthesis, stomatal conductance, soil water transport, in addition to utilities for processing aerial and terrestrial LiDAR data, and procedurally constructing randomized 3D plant geometries. Many of these plug-ins utilize graphics processing units (GPUs) to perform computations in parallel to maximize the modeled range of scales and level of physical realism.
This presentation will highlight a number of ongoing applications of Helios relevant to agricultural production and breeding, including:
Generation of auto-annotated simulated sensor data (e.g., RGB-D, multispectral, thermal, LiDAR) for efficient training of machine learning models for physiologically-based phenotyping
3D computer-aided design and management of efficient and sustainable agricultural systems
An irrigation system CAD tool integrating irrigation hydraulics, soil hydrology, and plant demand
Computer-aided mechanical harvester design with explicit 3D representation of harvester-plant-fruit interactions
Novel algorithms and implementations for processing LiDAR data into useful 3D model inputs and traits
Bio:
Dr. Brian Bailey brings an interdisciplinary research background that spans the fields of engineering, computer science, and plant biology. He received his PhD in Mechanical Engineering in 2015, before working at the USDA-ARS in Corvallis, OR as a Research Engineer. Since joining the Department of Plant Sciences at the University of California, Davis in 2016, his research program has focused on merging his interests in biophysics and plant physiology in order to develop next-generation 3D plant models for computer-aided design, management, and breeding of efficient agricultural systems. Applications of these modeling tools have focused on canopy design, irrigation design and management, disease management, and integration with machine learning models for high-throughput physiologically-based phenotyping.
Summary:
Motivation: CAD for agricultural crops
Design parameters (e.g. for vineyards):
Location
Cultivar/roostock
Training (trellis, head height, pruning)
Row spacing/orientation
Irrigation system
Micrometeorology
Regulations
Crop modeling is quite popular
Process-based crop models: APSIM, DSSAT
ODEs that capture the rates of development
Don’t include the biological drivers of development
ODEs created empirically from observed trends, require lots of data across various locations, seasons, varieties, management practices
Not spatially explicit
Poor models of perennial crops
Helios
Detailed 3D model of the plant
Detailed geometry resolved down to sub-leaf/branch scale
Can be used to model a wide range of natural, agricultural, agro-photovoltaic systems
Includes many plugins: Visualizer, solar radiation, photosynthesis, procedural plant model generator
Working on plugins: soil water transport, irrigation hydraulic design, carbohydrate transport
Observation: It is important to capture the illumination across each leaf (varies based on other leafs and sun angle)
Using leaf-averaged illumination doesn’t produce accurate results
Can use LIDAR surveys to parameterize sub-leaf model
Illumination onto leafs inferred via backward ray tracing (efficiency samples illumination onto all types of leaves, including scattering)
Outputs of HELIOS are validated using LIDAR data
HELIOS can predict the exact way each plant will look
Developed an automated leaf angle and area measurement model
Applications:
Pest propagation, infestation
Evaluation/design of different trellising & shade cloth strategies
Computer-aided design of harvesters, especially for bruisable fruits
Project GEMINI (collaboration with Mineral)
Genotype x Environment x Management interactions
Efforts:
3D Biophysical Crop Modeling
Pre-breeding and Genomics
AI-Enabled Phenotyping
Crops: Sorghum, Cowpea, Common Bean (focus is sub-Saharan Africa)
Using the Mineral rover
Multi-band video: RGB, 735hnm 850hm, thermal infrared
Since HELIOS can produce video, can use real video to tune the model’s parameters
Not using LIDAR, so giving up 3D structure information and LIDAR can penetrate through the canopy somewhat
Integration with AgML: Open-source Agriculture ML infra
Access to datasets, benchmarks, pre-trained models