Bridging domain gaps in agricultural 3D point cloud classification using adversarial domain adaptation

Area of Science:

• Agricultural Technology
• Computer Vision
• Machine Learning

Background:

• Traditional 2D imagery limits 3D sensing in agriculture.
• A gap exists in applying 3D technologies for plant monitoring and classification.
• Domain shift challenges models trained on controlled vs. real-world data.

Purpose of the Study:

• To develop an adversarial unsupervised domain adaptation framework for 3D point cloud classification in agriculture.
• To address the domain shift between controlled (Crops3D) and real-world (Pheno4D) datasets.
• To enable robust plant phenotyping using 3D sensing.

Main Methods:

• Utilized a PointNet-based feature extractor.
• Employed a domain discriminator with a Gradient Reversal Layer (GRL).
• Incorporated an entropy minimization objective for confident predictions on unlabeled target data.

Main Results:

• Achieved 97% classification accuracy on the target domain (Pheno4D).
• Demonstrated strong per-class F1 scores despite significant dataset differences.
• Evaluated real-time performance and edge device deployment feasibility.

Conclusions:

• The proposed 3D domain adaptation framework significantly enhances plant classification in precision agriculture.
• The method shows potential for creating more generalizable plant phenotyping models.
• 3D domain adaptation is crucial for advancing agricultural monitoring technologies.