Spring Maize Yield Prediction and Optimal Phenological Stage Assessment in the Junggar Basin Based on UAV and Stacked Ensemble Learning

Abstract

Abstract Accurate prediction of maize yield is crucial for improving field management and enabling timely yield estimation. To improve the accuracy and determine the optimal timing of field-scale spring maize yield estimation in the Junggar Basin, this study focuses on spring maize in this region. In 2023, UAV-based multispectral images were acquired at three key growth stages: jointing, filling, and milk stages. Eighteen spectral features significantly correlated with yield were selected. Spring maize yield prediction models were constructed using XGBoost, CatBoost, RF, DT, SVR, GP, LR, and a stacked ensemble learning model, respectively, revealing differences in prediction accuracy across growth stages. Finally, SHAP was used for model interpretability analysis. The results show that: (1) The milk stage achieved the highest prediction accuracy (R² = 0.761, MAE = 0.067 kg·m⁻², RMSE = 0.089 kg·m⁻², MAPE = 5.055%), outperforming the jointing and early grain-filling stages, thereby resolving the uncertainty regarding the optimal timing for UAV-based yield estimation of spring maize in the Junggar Basin. (2) Compared with traditional machine learning algorithms, the stacked ensemble model exhibited stronger robustness and generalization ability. This study provides a technical reference for timely yield estimation and field management of spring maize in irrigated areas of the Junggar Basin, supporting regional food production stability.

Plant phenotyping relevance

UAVマルチスペクトル画像からトウモロコシ収量を推定し、複数生育段階・機械学習モデルの精度と頑健性を比較して最適推定時期を評価しており、表現型推定手法が研究の中心である。

Code and data availability

The paper's UAV multispectral imagery, 200-point yield sampling data, and stacking ensemble analysis code are not publicly deposited. The Data availability statement says only that data will be made available on request; no public repository, URL, or code deposit is provided.

No evidence-backed public reproduction asset is currently recorded.