Research on acquiring maize aboveground structure data using NeRF-Based 3D reconstruction and point cloud segmentation

Abstract

Abstract Maize（Zea mays L.） is an important crop, and improving its productivity is required even under challenging conditions such as labor shortages and uncertain climate fluctuations. One approach to enhancing yield is utilizing crop data for cultivation management and yield prediction. However, efficient acquisition of such data remains constrained by various limitations. In this study, we developed a non-contact and labor-efficient method for crop data acquisition by generating 3D models of maize at the ripening stage using Neural Radiance Fields (NeRF). Segmentation was performed on the obtained point clouds to estimate plant height, leaf area, and leaf angle. The coefficients of determination (R2) were 0.903, 0.954, and -0.521, respectively, demonstrating high accuracy for plant height and leaf area even at the ripening stage, while reducing the time required for data acquisition by 93% compared to manual measurements. Nevertheless, some manual operations−such as removing kernels and separating overlapping leaves−were still necessary, and full automation was not achieved. The main sources of error were identified as reconstruction errors in the base during scale adjustment, excessive removal of leaf sheaths, and the curvature of individual plants. Furthermore, we examined how measurement accuracy was influenced by factors such as the time of day and cultivar. The proposed method is expected to contribute to the practical implementation of a labor -saving 3D measurement technique that supports yield prediction and growth diagnosis in maize.

Plant phenotyping relevance

NeRFによる3D再構成と点群セグメンテーションを開発し、トウモロコシの草丈・葉面積・葉角度を推定して精度と誤差要因を検証しており、表現型取得法が研究の中心である。

Code and data availability

The article describes NeRF-based maize 3D reconstruction and point cloud segmentation with phenotype measurements, but contains no data availability statement, no public dataset or image deposit, and no author code repository. The only URLs mentioned are generic tooling (Anaconda) and citations to prior work (arXiv NeR

No evidence-backed public reproduction asset is currently recorded.