Choi et al. (2025) Probabilistic Deep Learning Framework for Greenhouse Microclimate Prediction with Time-Varying Uncertainty and Covariance Analysis

Identification

• Journal: Agriculture
• Year: 2025
• Date: 2025-11-27
• Authors: Woo-Joo Choi, Myongkyoon Yang
• DOI: 10.3390/agriculture15232461

Research Groups

• Department of Agricultural Machinery Engineering, Jeonbuk National University, Jeonju, Republic of Korea
• Department of Bioindustrial Machinery Engineering, Jeonbuk National University, Jeonju, Republic of Korea
• Institute of Agricultural Machinery & ICT Convergence, Jeonbuk National University, Jeonju, Republic of Korea

Short Summary

This study developed a probabilistic deep learning framework to predict greenhouse microclimate variables with time-varying uncertainty and covariance analysis. The framework, based on a 1D CNN, demonstrated comparable predictive accuracy to deterministic models while providing explainable uncertainty interpretation and robust decision support for greenhouse operators.

Objective

• To develop a probabilistic deep learning framework that quantitatively predicts greenhouse microclimate (internal temperature, relative humidity, and CO2 concentration) three hours later, while simultaneously estimating and interpreting time-varying uncertainty and correlations.

Study Configuration

• Spatial Scale: A smart strawberry greenhouse located in Nonsan, Chungcheongnam-do, Republic of Korea.
• Temporal Scale:
  • Data collection period: 21 September 2023 to 26 April 2024.
  • Data sampling interval: Hourly for microclimate and external environment, 1 minute for control data (resampled to hourly).
  • Prediction horizon: 3 hours into the future.
  • Input sequence length: Optimized to 24 hours.

Methodology and Data

• Models used:
  • Probabilistic Deep Learning Framework: One-dimensional Convolutional Neural Network (1D CNN) and Long Short-Term Memory (LSTM) models.
  • Loss function for probabilistic models: Negative Log Likelihood (NLL).
  • Deterministic models (for comparison): LSTM and 1D CNN with a Multi-Task Learning (MTL) structure.
  • Loss function for deterministic models: Mean Absolute Error (MAE).
• Data sources:
  • Public datasets provided by the Korea Institute of Agriculture Technology Promotion Agency.
  • Greenhouse microclimate data (hourly): Internal temperature (degrees Celsius), internal relative humidity (%), CO2 concentration (ppm).
  • External environment data (hourly): External temperature (degrees Celsius), external relative humidity (%), precipitation (mm), wind speed (m/s), wind direction (degrees), solar radiation (W/m^2), snowfall (cm), ground temperature (degrees Celsius).
  • Control data (1-minute interval): Window and curtain opening statuses (%).
  • Data preprocessing: Handling missing values (replacement with zero for meteorological data, forward fill for control data), outlier detection (seasonal-trend decomposition with 24-hour cycle, residuals exceeding five standard deviations), linear interpolation, resampling to hourly, Min-Max scaling, and sine/cosine transforms for temporal variables.

Main Results

• The 1D CNN-based probabilistic model achieved an average R2 of 0.93, demonstrating predictive performance comparable to deterministic models (no statistically significant difference in RMSE).
• The model exhibited high sharpness and calibration, with an average Negative Log Likelihood (NLL) of 2.08 and a Coverage@90% of 0.901 across the three microclimate variables.
• For internal temperature, relative humidity, and CO2 concentration, the model achieved R2 values of 0.95, 0.94, and 0.92, and RMSE values of 1.4 degrees Celsius, 2.73%, and 43.54 ppm, respectively.
• The estimated covariance matrix successfully interpreted time-varying correlations between microclimate variables, revealing local simultaneous variability that global correlation analysis failed to capture.
• Prediction uncertainty, quantified by standard deviation, was notably higher in time intervals with lower data sampling frequencies.
• Average pairwise correlations were 0.23 for temperature–relative humidity, -0.05 for temperature–CO2 concentration, and -0.00 for relative humidity–CO2 concentration, with extreme values spanning the full theoretical range of -1 to 1.

Contributions

• Building a probabilistic deep neural network for the temporal projection of important greenhouse microclimate variables (internal temperature, relative humidity, and CO2 concentration).
• Integration of a Negative Log Likelihood (NLL) training objective enabling concurrent generation of projections and temporal covariance, thereby encoding model variability.
• Deriving model variability and dependence between nonlinear variables through temporal covariance analysis using Cholesky factorization.
• Introducing analytical methods to quantify the variability and reliability of network output and support the assessment of predictive reliability and resilience.

Funding

• Korea Institute of Planning and Evaluation for Technology in Food, Agriculture and Forestry (IPET)
• Korea Smart Farm R&D Foundation (KosFarm) through Smart Farm Innovation Technology Development Program
• Ministry of Agriculture, Food and Rural Affairs (MAFRA)
• Ministry of Science and ICT (MSIT)
• Rural Development Administration (RDA) (Project Reference: RS-2024-00399854)

Citation

@article{Choi2025Probabilistic,
  author = {Choi, Woo-Joo and Yang, Myongkyoon},
  title = {Probabilistic Deep Learning Framework for Greenhouse Microclimate Prediction with Time-Varying Uncertainty and Covariance Analysis},
  journal = {Agriculture},
  year = {2025},
  doi = {10.3390/agriculture15232461},
  url = {https://doi.org/10.3390/agriculture15232461}
}