NeRF-based 3D reconstruction pipeline for acquisition and analysis of tomato crop morphology

Abstract

Recent advancements in digital phenotypic analysis have revolutionized the morphological analysis of crops, offering new insights into genetic trait expressions. This manuscript presents a novel 3D phenotyping pipeline utilizing the cutting-edge Neural Radiance Fields (NeRF) technology, aimed at overcoming the limitations of traditional 2D imaging methods. Our approach incorporates automated RGB image acquisition through unmanned greenhouse robots, coupled with NeRF technology for dense Point Cloud generation. This facilitates non-destructive, accurate measurements of crop parameters such as node length, leaf area, and fruit volume. Our results, derived from applying this methodology to tomato crops in greenhouse conditions, demonstrate a high correlation with traditional human growth surveys. The manuscript highlights the system's ability to achieve detailed morphological analysis from limited viewpoint of camera, proving its suitability and practicality for greenhouse environments. The results displayed an R-squared value of 0.973 and a Mean Absolute Percentage Error (MAPE) of 0.089 for inter-node length measurements, while segmented leaf point cloud and reconstructed meshes showed an R-squared value of 0.953 and a MAPE of 0.090 for leaf area measurements. Additionally, segmented tomato fruit analysis yielded an R-squared value of 0.96 and a MAPE of 0.135 for fruit volume measurements. These metrics underscore the precision and reliability of our 3D phenotyping pipeline, making it a highly promising tool for modern agriculture.