Laouz et al. (2026) A Novel Deep Learning Framework Based on Heterogeneous Temporal Data Harmonization for Irrigation Water Amount Prediction

Identification

• Journal: Arabian Journal for Science and Engineering
• Year: 2026
• Date: 2026-04-04
• Authors: Hamed Laouz, Soheyb Ayad, Labib Sadek Terrissa, Samir Merdaci, Noureddine Zerhouni
• DOI: 10.1007/s13369-026-11267-1

Research Groups

• Computer Science Department, University of Biskra, Biskra, Algeria (Hamed Laouz, Soheyb Ayad, Labib Sadek Terrissa)
• Agronomy Department, University of El Oued, El Oued, Algeria (Samir Merdaci)
• AS2M Department, Franche Comté University, Besançon, France (Noureddine Zerhouni)

Short Summary

This paper proposes a novel deep learning framework, combining a Convolutional Neural Network (CNN) for heterogeneous temporal data harmonization and a Multilayer Perceptron (MLP) for prediction, to accurately forecast daily irrigation water amounts, achieving a Mean Absolute Error of 0.5 L/m² and an R² score of 0.82.

Objective

• To develop a deep learning framework that predicts the daily irrigation water amount for plants by harmonizing heterogeneous temporal climate and soil data, specifically addressing the challenge of varying data recording intervals (per-minute and daily).

Study Configuration

• Spatial Scale: Controlled greenhouse environment (Autonomous Greenhouse Challenge, Netherlands).
• Temporal Scale: Input data includes per-minute (5-minute interval) climate parameters and daily soil/irrigation parameters; the model predicts daily irrigation water amounts.

Methodology and Data

• Models used: A hybrid CNN-MLP model. CNN is used for feature extraction and dimensionality reduction from per-minute data, and MLP is used for the final regression prediction.
  • Baseline models for comparison: Support Vector Regressor (SVR), Random Forest Regressor (RFR), Decision Tree Regressor (DTR), Gradient Boost Regressor (GBR), Recurrent Neural Network (RNN), Long-Short Term Memory (LSTM), Broad Learning System (BLS), Transformer, Multi-Attention LSTM.
  • Dimensionality reduction methods for comparison: Principal Component Analysis (PCA), t-Distributed Stochastic Neighbor Embedding (t-SNE), ISOMAP.
• Data sources: Autonomous Greenhouse Challenge (AGC) public dataset (first edition, 2018) from the winning team, specifically the "GH Climate" (per-minute) and "Irrigation" (daily) datasets.

Main Results

• The proposed CNN-MLP model achieved a Mean Absolute Error (MAE) of 0.510 L/m², a Root-Mean-Squared Error (RMSE) of 0.728 L/m², and an R² score of 0.8122.
• This approach significantly outperformed classical data harmonization methods (e.g., daily averaging) and various baseline machine learning and deep learning models across all evaluation metrics. For instance, the best baseline using daily average (GBR) achieved an MAE of 0.506 L/m², RMSE of 0.826 L/m², and R² of 0.62.
• The CNN-MLP model also showed superior performance compared to advanced deep learning models like BLS (MAE = 0.521 L/m², RMSE = 0.741 L/m², R² = 0.79) and other dimensionality reduction techniques such as ISOMAP (MAE = 0.601 L/m², RMSE = 1.043 L/m², R² = 0.5376).
• Statistical evaluation using t-tests confirmed that the observed improvements were statistically significant (p-values < 0.05).
• The training time for the proposed approach was 21 seconds, with an inference time of 0.08 seconds, demonstrating reasonable computational efficiency.

Contributions

• Introduction of a novel smart irrigation framework that integrates CNN-based feature extraction with MLP regression, specifically designed for predicting irrigation water requirements.
• Development of a dedicated data harmonization strategy utilizing a CNN-based feature extractor to effectively transform irregular per-minute plant air data into uniform daily-scale features without significant information loss.
• Demonstrated superior predictive performance compared to existing literature, including traditional machine learning models, advanced deep learning architectures, and conventional data harmonization and feature reduction techniques.

Funding

• Algerian Directorate General for Scientific Research and Technological Development (DGRSDT) under the National Research Projects (PNR) program.
• Technical Institute for the Development of Saharan Agronomy (ITDAS) of Biskra.

Citation

@article{Laouz2026Novel,
  author = {Laouz, Hamed and Ayad, Soheyb and Terrissa, Labib Sadek and Merdaci, Samir and Zerhouni, Noureddine},
  title = {A Novel Deep Learning Framework Based on Heterogeneous Temporal Data Harmonization for Irrigation Water Amount Prediction},
  journal = {Arabian Journal for Science and Engineering},
  year = {2026},
  doi = {10.1007/s13369-026-11267-1},
  url = {https://doi.org/10.1007/s13369-026-11267-1}
}