Ma et al. (2026) Prediction Method of Canopy Temperature for Potted Winter Jujube in Controlled Environments Based on a Fusion Model of LSTM–RF

Identification

• Journal: Horticulturae
• Year: 2026
• Date: 2026-01-12
• Authors: Shufan Ma, Lin Kou, Sheng Huang, Ying Fu, F. Y. Zhang, Xianpeng Sun
• DOI: 10.3390/horticulturae12010084

Research Groups

• College of Horticulture, North West Agriculture and Forestry University, Xianyang 712100, China
• Key Laboratory of Horticultural Engineering in Northwest Facilities, Ministry of Agriculture, Xianyang 712100, China
• Northwest A&F University & Xi’an Jiaotong University Agricultural Equipment Research Institute, Xianyang 712100, China

Short Summary

This study developed a data-driven model to forecast canopy temperature for potted winter jujube in controlled environments. The proposed LSTM–RF fusion model achieved superior prediction performance (R2 = 0.974, MAE = 0.844 °C, RMSE = 1.155 °C) compared to benchmark models, providing reliable and interpretable insights for precision irrigation.

Objective

• To achieve high-precision prediction of canopy temperature for facility-cultivated winter jujube by developing a novel integrated forecasting framework that synergistically combines a Long Short-Term Memory (LSTM) network and a Random Forest (RF) model.

Study Configuration

• Spatial Scale: Potted winter jujube (Ziziphus jujuba Mill. ‘Dongzao’) plants cultivated in a large-span plastic greenhouse (Rougu Nongyuan Agricultural Park, Yangling District, Xianyang City, Shaanxi Province, China). Experiments used cylindrical pots with a volume of 0.1413 m³.
• Temporal Scale: Data collected from 1 April to 7 September (approximately 5 months and 7 days). All variables were recorded at 30-minute intervals, enabling nowcasting (single-step prediction).

Methodology and Data

• Models used:
  • Proposed: LSTM–RF fusion model (Long Short-Term Memory network serially integrated with Random Forest algorithm).
  • Benchmark models: Transformer, TimesNet, standalone Random Forest (RF), standalone Long Short-Term Memory (LSTM).
  • Data preprocessing: Isolation Forest (IF) for outlier detection, Spearman correlation analysis for feature selection, Robust Principal Component Analysis (RPCA) for thermal image dimensionality reduction.
• Data sources:
  • Thermal infrared images: FLIR AX8 camera (7.5–13 µm spectral range, 320 × 240 pixels resolution, captured at 30 min intervals from 1.2 m distance and 30° above canopy).
  • Environmental sensors: Greenhouse air temperature, air humidity, solar radiation, wind speed, soil temperature, and soil humidity (all collected at 30 min intervals).
  • Data collection period: 1 April to 7 September, yielding 7680 sets of data.

Main Results

• The proposed LSTM–RF model demonstrated superior canopy temperature prediction performance with a determination coefficient (R2) of 0.974, a mean absolute error (MAE) of 0.844 °C, and a root mean square error (RMSE) of 1.155 °C.
• This performance significantly outperformed benchmark models, including standalone LSTM (R2=0.941, MAE=0.985 °C, RMSE=1.747 °C), RF (R2=0.956, MAE=0.956 °C, RMSE=1.505 °C), Transformer (R2=0.916, MAE=1.396 °C, RMSE=1.806 °C), and TimesNet (R2=0.849, MAE=1.573 °C, RMSE=2.135 °C).
• SHAP-based interpretability analysis identified the "thermodynamic state of air" driver group (air temperature, air humidity, vapor pressure deficit) and LSTM-derived latent temporal features as the most influential factors impacting canopy temperature prediction.
• The canopy system's thermal memory was found to diminish beyond approximately one hour (three 30-minute time steps), indicating the characteristic timescale of canopy temperature response.
• Ablation experiments confirmed the critical role of the "thermodynamic state of air" driver group, LSTM-derived latent features, and short-term lagged inputs for achieving high predictive performance.

Contributions

• Explores the integration of Random Forest with Long Short-Term Memory networks for greenhouse crop canopy temperature prediction, a novel combination in this specific context.
• Provides a comprehensive comparison and analysis of multiple deep learning algorithms, demonstrating the superior predictive performance of the proposed hybrid model.
• Identifies and evaluates key features influencing canopy temperature dynamics in protected winter jujube cultivation, offering a novel approach for managing water stress and optimizing irrigation strategies.
• Establishes a reliable, interpretable, data-driven foundation for real-time water stress monitoring and intelligent irrigation scheduling in controlled horticulture systems.

Funding

• National Key R&D Program of China (2024YFD2001001-02)
• Key Research and Development Plan of Shaanxi Province (S2024-YF-ZDCXL-ZDLNY-015)
• Key Research and Development Plan of Shaanxi Province (2024NC2GJHX12)

Citation

@article{Ma2026Prediction,
  author = {Ma, Shufan and Zhang, Yingtao and Kou, Lin and Huang, Sheng and Fu, Ying and Zhang, F. Y. and Sun, Xianpeng},
  title = {Prediction Method of Canopy Temperature for Potted Winter Jujube in Controlled Environments Based on a Fusion Model of LSTM–RF},
  journal = {Horticulturae},
  year = {2026},
  doi = {10.3390/horticulturae12010084},
  url = {https://doi.org/10.3390/horticulturae12010084}
}