Monitoring crop leaf area index using improved global structure-from-motion and multi-feature data fusion on a phenotyping robot

Abstract

Reconstruction of crop three-dimensional (3D) point clouds is essential for monitoring phenotypic parameters, like plant height and leaf area index (LAI), which is a critical phenotype predictor for smart crop breeding. The main 3D reconstruction technologies include image-based approaches, laser scanning, and depth camera methods. Among these methods, image-based structure-from-motion (SfM) is widely used due to its low cost and high accuracy. However, field crop canopy image data for high-resolution point cloud construction are often large-scale, unordered, and uncalibrated. Conventional SfM methods struggle with 3D reconstruction due to high computational costs and long processing times, delaying phenotypic analysis. To address this issue, we developed an improved global SfM algorithm, which increases the point cloud reconstruction speed by an average of 1.39 times compared to traditional incremental SfM methods and by more than 10 % on average compared to two mainstream global SfM algorithms. In addition, we integrated three types of predictors, point cloud features, color indices and texture features, through multi-feature data fusion and machine learning. A random forest algorithm for the prediction of LAI for a combined data set of four different crops, and using all three categories of predictors, achieved higher monitoring accuracy compared to using a single feature category (R²=0.78 vs R²=0.71–0.74). This new method, which includes an improved global SfM algorithm and a three-predictor fusion-based LAI monitoring approach, offers an efficient and reliable solution for precise crop phenotyping and continuous growth monitoring in complex field environments, enabling accurate assessment of crop morphology and developmental dynamics.

Code and data availability

公開本文の所在を確認できませんでした。非公開または購読が必要な可能性があります。

No evidence-backed public reproduction asset is currently recorded.