Articles
A novel 3D Gaussian splatting approach for leaf area index estimation of tomato plants
Article number
1433_37
Pages
291 – 298
Language
English
Abstract
The monitoring of plant morphological characteristics, such as the leaf area index (LAI) and stem thickness, has become a key point of attention in greenhouse horticulture, as it is crucial for refining cultivation practices, optimizing resource management, and enhancing overall crop productivity.
In this study, a novel method to assess the LAI of tomato plants was developed that combines the advantages of both direct and indirect LAI estimation methods.
This method uses the 3D Gaussian splatting technique, by which a soft point cloud of so-called splats is derived from 2D images of the plants under study.
Through post-processing of this point cloud, various significant morphological characteristics, including the LAI, can be extracted.
To validate this methodology, first, a digital setup featuring a 3D mesh of a generic plant was devised, which allowed for controllable manipulation of the LAI. Of this setup, photorealistic synthetic images were rendered and used as input into the Gaussian splatting method.
Then, following splat filtering based on position, colour, size, and opacity, a 3D representation of the plant was reconstructed.
For LAI quantification, the point cloud was voxelized, and the voxels were classified as either leaf or non-leaf.
Through top-down summation of the leaf voxel mask, a 2D LAI map was then generated, and the average of this map provided the LAI estimate.
After choosing an appropriate voxel size that preserved the morphological properties of the plant, this method yielded a deviation of less than 3% from the ground truth LAI (based on the initial 3D model). Following experiments conducted on an actual plant, our proposed approach, utilizing video captured by a smartphone camera, aligns with the destructive LAI measurement of the identical plant.
In this study, a novel method to assess the LAI of tomato plants was developed that combines the advantages of both direct and indirect LAI estimation methods.
This method uses the 3D Gaussian splatting technique, by which a soft point cloud of so-called splats is derived from 2D images of the plants under study.
Through post-processing of this point cloud, various significant morphological characteristics, including the LAI, can be extracted.
To validate this methodology, first, a digital setup featuring a 3D mesh of a generic plant was devised, which allowed for controllable manipulation of the LAI. Of this setup, photorealistic synthetic images were rendered and used as input into the Gaussian splatting method.
Then, following splat filtering based on position, colour, size, and opacity, a 3D representation of the plant was reconstructed.
For LAI quantification, the point cloud was voxelized, and the voxels were classified as either leaf or non-leaf.
Through top-down summation of the leaf voxel mask, a 2D LAI map was then generated, and the average of this map provided the LAI estimate.
After choosing an appropriate voxel size that preserved the morphological properties of the plant, this method yielded a deviation of less than 3% from the ground truth LAI (based on the initial 3D model). Following experiments conducted on an actual plant, our proposed approach, utilizing video captured by a smartphone camera, aligns with the destructive LAI measurement of the identical plant.
Authors
J. Westra, D. Boesten, S. Nieboer, S. Aerts, J. Bolte
Keywords
smart horticulture, digital twin, plant growth monitoring
Online Articles (41)