High-throughput estimation of sugarcane phenotypic traits using UAV multispectral data under high-density planting conditions.

Abstract

High-throughput phenotyping using unmanned aerial vehicle (UAV)-based imagery offers substantial potential for improving sugarcane breeding efficiency. This study utilized UAVs-equipped multispectral sensors to capture high-resolution imagery of 652 sugarcane varieties under high-density planting condition, enabling the development of predictive models for key phenotypic traits including plant height, leaf length, leaf width, and relative chlorophyll content (SPAD value). A comprehensive feature extraction process yielded 100 vegetation indices, 7 texture indices, and canopy height parameters derived from the UAV imagery. To develop robust predictive models, we implemented three feature processing strategies—correlation-based filtering (COR), stepwise regression selection (SWR), and principal component analysis (PCA)—in conjunction with five machine learning algorithms: Lasso Regression (LASSO), Ridge Regression (Ridge), Support Vector Machine Regression (SVM), Random Forest (RF), and Gradient Boosting Regression Trees (GBR). Two ensemble methods, Bayesian Model Averaging (BMA) and Stacked Generalization, were also employed. Results demonstrated that LASSO performed best among traditional machine learning models, whereas the Stacking ensemble method, which integrated predictions from all individual algorithms, achieved the highest prediction accuracy (the coefficient of determination ( R 2 ) = 0.77; root mean squared error ( RMSE ) = 12.99 cm for plant height). Additionally, K-means clustering partitioned the sugarcane varieties into two distinct clusters (A and B; p ≤ 0.001). Notably, cluster-specific models trained on PCA-processed features demonstrated exceptional predictive accuracy during validation, achieving R 2 values of 0.94, 0.91, 0.87, and 0.90 for plant height, leaf length, leaf width, and SPAD value, respectively. This research presents an integrated framework combining optimized feature processing, population clustering, and ensemble learning to enhance trait prediction in large-scale UAV-based phenotyping for sugarcane breeding.

Plant phenotyping relevance

UAVマルチスペクトル画像からサトウキビの複数形質を推定する予測モデルと統合的な表現型解析フレームワークを開発・検証しており、表現型取得・抽出手法が研究の中心である。

Code and data availability

The paper states that its data and code (UAV multispectral phenotyping data and analysis code for sugarcane trait prediction) are available for non-commercial academic use via an FTP server (ftp://supercane.synology.me). However, this URL is not among the allowed_urls, so it cannot be cited as a public actionable asset

No evidence-backed public reproduction asset is currently recorded.