Comparative evaluation of 3D data acquisition approaches for rail-driven field plant phenotyping platforms

The selection of sensors for a high-throughput plant phenotyping platform is crucial for its subsequent development. It impacts the control mode, data storage and transmission, phenotype analysis algorithm and accuracy. This paper compared and evaluated the three-dimensional (3D) data acquisition performance of LiDAR, Multi-View Stereo (MVS) reconstruction, and depth image synthesis in five growth stages of maize canopies. The study found that LiDAR was the most stable and least affected by the environment. Additionally, it had the highest plant height estimation accuracy, with an average R 2 of 0.80 across all five stages. However, LiDAR is greatly affected by the stationarity of the platform and the noise of the resulting maize point cloud can be significant. The sensor required for MVS mode is low-cost, has minimal influence on platform stationarity, and allows for convenient point cloud synthesis and colour information. However, it is greatly affected by the lighting environment, resulting in a certain degree of distortion in the obtained point cloud. Additionally, it has the highest pre-processing complexity. Depth point cloud has the highest synthesis efficiency and the lowest data pre-processing complexity, making it suitable for online pre-processing and analysis. However, the initial data obtained is large and its stability is low due to its susceptibility to environmental factors. The point cloud acquired by MVS and Depth are clearer than LiDAR, making it easier for plant segmentation. This study provides a valuable foundation for the development of a high-throughput plant phenotyping platform and sensor selection.