AI-Driven Background Segmentation for High-Throughput 3D Plant Scans

Accurate background segmentation in 3D plant phenotyping is crucial for reliable trait assessment but remains challenging. Current methods are either excessively complex, developed for a different domain, or lead to data loss (coordinate-based). This paper addresses these issues by introducing an AI-driven approach using a Multi-Layer Perceptron (MLP) model, leveraging RGB, spatial (XYZ), and near-infrared (NIR) data to enhance precision. The method was evaluated on high-throughput phenotyping data, achieving a classification accuracy of 0.9993, significantly reducing false positives and false negatives compared to coordinate-based segmentation. The proposed approach improved segmentation, particularly in early growth stages and for prostrate species, where traditional methods often fail. The model’s impact on leaf area estimation was validated against destructive measurements, demonstrating substantial accuracy improvements, especially for species with small and prostrate canopies. Additionally, the model exhibited strong generalization capabilities when applied to an external 3D dataset, confirming its reusability beyond plant phenotyping tasks. Integrating this simple method into phenotyping pipelines will enhance efficiency and accuracy in high-throughput trait estimation, supporting advancements in plant science and precision agriculture.