Tree-based machine learning models to predict maize growth, nitrogen uptake, and yield using UAV remote sensing

Abstract

Unmanned aerial systems (UAS) equipped with multispectral sensors enable within-season crop phenotyping; however, conventional vegetation index–based approaches often lack accuracy in predicting crop growth and yield. This study evaluated the performance of three machine learning (ML) algorithms, support vector machine (SVM), random forest (RF), and extreme gradient boosting (XGBoost), for predicting maize leaf area index (LAI), relative chlorophyll content (SPAD), aboveground biomass (AGB), plant height, nitrogen (N) uptake, and grain yield, and compared their performance with stepwise multiple linear regression (SMLR). Multispectral imagery (five spectral bands) was collected across multiple growth stages during the 2021 and 2022 growing seasons near College Station, Texas, under varying split N-fertilizer applications. Band reflectance and eight vegetation indices were used as model inputs. All ML models outperformed SMLR in predicting LAI, SPAD, AGB, N uptake, and plant height. SVM, RF, and XGBoost showed comparable performance for LAI (R² = 0.82–0.83), AGB (R² = 0.86–0.92), and plant height (R² = 0.95–0.97). However, XGBoost exhibited overfitting, resulting in lower validation accuracy for SPAD (R² = 0.54) and N uptake (R² = 0.68) compared with SVM (R² = 0.73 and 0.80, respectively). Grain yield prediction accuracy increased with crop maturity, with reproductive-stage imagery producing the highest accuracy across models (R² = 0.88–0.98), where all ML models outperformed SMLR. However, at the V6 growth stage, SVM and RF did not perform as well as SMLR. Overall, integrating ML with UAS-based multispectral imagery improved prediction accuracy across key maize phenotypic and nutrient traits, providing a robust framework for non-destructive, high-throughput phenotyping and precision N management.

Plant phenotyping relevance

UASマルチスペクトル画像と機械学習により、作物の形態・生理・収量関連形質を推定する方法を比較評価しており、フェノタイピング手法が研究の中心である。

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.