Cabbage yield estimation with multi‐view convolutional neural networks from UAV‐derived point clouds

Three‐dimensional (3D) data from remote sensing, including point clouds, provide detailed canopy structural information that can enhance both estimation accuracy and model generalization. The main objective of this study was to evaluate the accuracy of per‐plant yield estimation of cabbage ( Brassica oleracea var. capitata L.) using a multi‐view convolutional neural network approach based on 3D projection images from multiple viewpoints. In addition, the generalization ability of this approach was examined using test datasets from unseen fields. Four types of models were established in this study. Two models used multi‐view images with eight oblique views, either with or without an additional top‐down view. The other two were baseline models for comparison: a single‐image convolutional neural network using orthomosaic crops and a simple pixel‐based linear model. The models were evaluated on datasets from the same field as training and validation, as well as on two independent fields. The multi‐view models achieved Lin's concordance correlation coefficient of 0.74 ± 0.013 and 0.73 ± 0.017 on the dataset from the same field as training and validation, which exceeds the single‐image models. Although prediction errors increased in independent fields compared with random‐split test dataset, multi‐view models outperformed the baselines, demonstrating robustness under extrapolative conditions. Gradient‐weighted Class Activation Mapping results suggested that the multi‐view models focused on regions different from those highlighted by the single‐image model. These results highlight that multi‐view models not only improve yield estimation accuracy but also enhance generalization, supporting broader applications such as stress detection and growth monitoring in precision agriculture.