High throughput phenotyping of energy cane using uncrewed aircraft system (UAS) and machine learning

Abstract

Accurate estimation of biomass in energy cane is essential for cultivar selection in breeding programs and biomass supply forecasting in bioenergy production. This study evaluated the integration of Uncrewed Aircraft System (UAS) based Light Detection and Ranging (lidar) and Red-Green-Blue Structure-from-Motion (RGB-SfM) photogrammetry to enhance biomass prediction for high-throughput phenotyping (HTP). Seven cultivars were monitored between December 2023 and July 2024 at an experimental field in Weslaco, Texas. Structural metrics such as percentile-based heights, canopy volume, and interaction variables were extracted and used to train four machine learning models: Random Forest (RF), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and a one-dimensional Convolutional Neural Network (1D-CNN). Ensemble tree algorithms consistently outperformed CNN, with XGBoost and LightGBM providing the most stable and interpretable predictions. For single sensor inputs, lidar models (R² = 0.70–0.78; RMSE = 1.36–1.60 kg/m²) generally outperformed RGB-SfM (R² = 0.69–0.73; RMSE = 1.52–1.63 kg/m²), though RGB-SfM performed competitively with XGBoost. Fused models combining lidar and RGB-SfM features achieved the highest accuracies (XGBoost: R² = 0.80, RMSE = 1.31 kg/m²; LightGBM: R² = 0.79, RMSE = 1.36 kg/m²), mitigating the underestimation of high-biomass plots in RGB-SfM and the slight overestimation of lidar at the upper tail. Cultivar specific analysis confirmed TH16–22 as the top performer, followed closely by Ho02–113 and TCP10–4928, demonstrating the capacity of UAS HTP to support breeding decisions. These findings confirm the biological relevance of percentile-based height metrics (particularly the 75th percentile), canopy volume, and their interactions for biomass accumulation and underscore the value of sensor fusion in reducing systematic bias. This study provides systematic demonstration of lidar and RGB-SfM fusion for biomass estimation in energy cane, establishing a scalable and non-destructive approach that advances high-throughput phenotyping and supports the development of sustainable bioenergy cropping systems. • Ensemble models (RF, LightGBM) outperform CNN in UAS phenotyping at plot scale. • Canopy volume and percentile heights identified as key biomass predictors. • Fusion of UAS lidar and RGB-SfM improves energy cane biomass estimation. • Sensor fusion mitigates RGB underestimation and lidar overestimation biases. • UAS-HTP supports energy cane cultivar screening for bioenergy applications.

Plant phenotyping relevance

UAS lidar・RGB-SfMによるバイオマス推定と、特徴抽出・機械学習・センサ融合の検証が研究の中心であり、再利用可能な高スループット表現型計測手法を評価している。

Code and data availability

公開論文であることは確認できましたが、現在の公式API・許可済み取得経路では本文を自動取得できませんでした。

No evidence-backed public reproduction asset is currently recorded.