Session: 16-01-01: Government Agency Student Poster Competition

Paper Number: 150603

150603 - Procedural Generation of 3d Maize Plant Models 

Breeding studies in agriculture are currently performed using traditional field experiments or greenhouse trials, which are very resource-intensive. 3D computational models of plants could provide a viable, cost-effective alternative. The flexibility of virtual models allows for quick and efficient experimentation with various plant configurations and growth scenarios. However, one of the main bottlenecks is creating a virtual 3D geometry of plants that could be easily modified to match the genotype being studied. To overcome this challenge, we create procedural models of maize plants that can be easily modified with minimal user invention. We specifically develop methods to automatically generate procedural models of Maize plant leaves that match point cloud data captured using LiDAR scanners.

Our procedural model consists of using Non-Uniform Rational B-Spline (NURBS) surfaces to represent the leaves of the Maize plant. We use a two-step optimization process to perform this process automatically -- a zeroth order strategy to robustly find the basin of attraction, followed by a first/second order method for rapid convergence to the optima. In the first step, we use particle swarm optimization (PSO) to optimize 32 parameters that define a 3x6 NURBS surface. These control point coordinates are used to create the NURBS surface, whose quality of fit is then evaluated by minimizing the distance between the generated surface points and the ground truth point cloud data. The optimization criterion combines the Chamfer distance and a weighted Hausdorff distance metric. Following PSO, the best output, comprising 18 control points, is refined using a differentiable NURBS programming framework called NURBS-Diff. The NURBS-Diff framework is designed to optimize NURBS surfaces through gradient descent methods. This framework leverages automatic differentiation to compute the gradients of the surface with respect to input NURBS parameters, enabling gradient-based optimization. We achieve detailed and accurate leaf models by treating each leaf separately during both PSO and NURBS-Diff optimizations. These optimized leaves are then assembled to reconstruct the entire maize plant. This approach ensures precise control over each plant part, facilitating further analysis and simulations.

Preliminary results demonstrate the effectiveness of combining PSO with NURBS-Diff. Using only PSO does not fully align the NURBS surface with the ground truth, but subsequent optimization with NURBS-Diff achieves a better fit. Similarly, NURBS-Diff alone cannot produce good results since the initial values are not close to the final surface. However, starting with a PSO-generated surface close to the ground truth leads to high-quality outcomes. This combined approach significantly improves quality of fit, indicating improved accuracy of the NURBS surfaces compared to the point cloud data. The procedural generation of these models provides a robust framework for manipulating and experimenting with virtual plant structures.

Presenting Author: Mozhgan Hadadi Iowa State University of Science and Technology

Presenting Author Biography: I received my BSc in Biosystem mechanical engineering from the University of Tehran, Iran, in 2017. I joined the ComPM Lab(Computational Physics and Mechanics Laboratory) in the Fall of 2022 to pursue my studying in Ph.D. degree, and I am co-advised by Dr. Baskar Ganapathysubramanian and Dr. Adarsh Krishnamurthy. My research focuses on procedural models in agriculture and differentiable models in machine learning and deep learning.

Authors:

Mozhgan Hadadi Iowa State University of Science and Technology
Mehdi Saraeian Iowa State University of Science and Technology
Aditya Balu Iowa State University of Science and Technology
Talukder Zaki Jubery Iowa State University of Science and Technology
Yawei Li Iowa State University of Science and Technology
Patrick Schnable Iowa State University of Science and Technology
Adarsh Krishnamurthy Iowa State University of Science and Technology
Baskar Ganapathysubramanian Iowa State University of Science and Technology