Progga et al. (2025) A chip-based radio frequency sensor for soil moisture measurements: A machine learning and deep learning calibration approach

Identification

• Journal: Journal of Agriculture and Food Research
• Year: 2025
• Date: 2025-12-19
• Authors: Jannatul Ferdaous Progga, Xiaomo Zhang, Srabana Maiti, Xin Sun, Shuvashis Dey, Sulaymon Eshkabilov, Xiaoyu Feng
• DOI: 10.1016/j.jafr.2025.102591

Research Groups

• Department of Agricultural and Biosystems Engineering, North Dakota State University, Fargo, USA
• Department of Electrical and Computer Engineering, North Dakota State University, Fargo, USA
• Feng Research Laboratory, North Dakota State University, Fargo, USA
• RF-Connect Laboratory, North Dakota State University, Fargo, USA

Short Summary

This study developed and evaluated machine learning (ML) and deep learning (DL) calibration models for a novel chip-based radio frequency (RF) soil moisture sensor using diverse soil samples. The Convolutional Neural Network (CNN) model achieved the highest accuracy (R² = 0.78, RMSE = 1.92 % volumetric moisture content) for generalized soil moisture estimation.

Objective

• To develop a generalized and robust calibration framework using various advanced ML and DL techniques for a novel chip-based RF soil moisture sensor.
• To develop and compare different ML-based calibration models for predicting volumetric moisture content (VMC) across different soil types.
• To optimize and validate model accuracy and robustness using k-fold cross-validation for reliable moisture estimation.
• To analyze feature importance and evaluate model generalizability under reduced input scenarios to identify key variables for field deployment.

Study Configuration

• Spatial Scale: Laboratory experiments conducted at North Dakota State University, Fargo, USA. Nine diverse soil samples were used: three commercial (Pure Sand, Mixed Sand, Garden Soil) and six field-collected from Carrington, ND, and Casselton, ND, USA. Experiments involved 300 mL soil samples in 500 mL glass beakers, with the sensor inserted at 5 cm depth. A total of 261 data points were collected.
• Temporal Scale: Experiments were conducted in 2024, with soil samples collected in October 2024. Measurements for each moisture level were averaged from continuous readings at 12 readings per minute after a 10-minute equilibration period.

Methodology and Data

• Models used: Linear Regression Model (LRM), Polynomial Regression Model (PRM), Artificial Neural Network (ANN), One-dimensional Convolutional Neural Network (CNN), Dense Neural Network (DNN), Long Short-Term Memory (LSTM). Random Forest Regression (RFR) was used for feature importance analysis.
• Data sources:
  • Chip-based RF sensor readings (Received Signal Strength Indicator - RSSI, in dBm).
  • Laboratory measurements of soil physicochemical properties: Volumetric Moisture Content (VMC, in %), Electrical Conductivity (EC, in μS/cm), pH, Bulk density (in g/cm³), and Soil textural classes.
  • VMC was calculated from gravimetric moisture content and bulk density.

Main Results

• Feature Importance: Bulk density was the most influential predictor (importance score 0.61), followed by the raw RF signal (RSSI) (0.20), EC (0.08), and pH (0.06). Categorical soil textural classes had minimal predictive value (<0.02).
• Model Performance (Full Feature Set, Single Run):
  • CNN achieved the highest accuracy: R² = 0.78, adjusted R² = 0.75, RMSE = 1.92 % VMC.
  • DNN followed closely: R² = 0.76, adjusted R² = 0.75, RMSE = 2.02 % VMC.
  • PRM was the best traditional ML model: R² = 0.75, adjusted R² = 0.74, RMSE = 2.04 % VMC.
  • All models except LRM met the agricultural application criteria of R² > 0.65 and RMSE < 3.5 % VMC; only CNN met the more stringent RMSE < 2.0 % VMC.
• Impact of Feature Reduction: ANN and DNN models demonstrated the greatest resilience to reduced input features, maintaining strong predictive performance. CNN and LSTM were more sensitive to feature limitations.
• K-fold Cross-Validation (5-fold): Nonlinear models (PRM, ANN, CNN, DNN) showed overall strong performance. While CNN and DNN achieved the highest mean R² (0.67), PRM (mean R² = 0.66 ± 0.06) and ANN (mean R² = 0.65 ± 0.08) demonstrated greater stability and consistency across folds (narrower confidence intervals and lower standard deviations).

Contributions

• This study presents the first comprehensive investigation into developing a generalized calibration model for novel chip-based RF soil moisture sensors using advanced ML and DL methodologies.
• It provides a strong proof of concept for integrating low-cost, chip-based RF sensing with data-driven modeling to deliver a highly accurate, scalable, and generalized solution for real-time soil moisture estimation in precision agriculture.
• The developed CNN model achieved high accuracy (RMSE = 1.92 % VMC) across nine highly heterogeneous soil types without requiring soil-specific calibration, validating the low-cost RF sensor as a robust alternative to commercial sensors.
• The demonstrated resilience of ANN and DNN models to feature reduction suggests a viable pathway for cost-effective system simplification in field deployment, requiring only core sensor readings and a few essential soil parameters.

Funding

• U.S. Department of Agriculture (Agreement No. 58-3060-3-022)
• U.S. Department of Agriculture, Agricultural Research Service (Agreement No. 59-5082-4-001)
• U.S. Department of Agriculture, National Institute of Food and Agriculture, Hatch Multistate project (accession number, 7005554)
• North Dakota State University (funds from North Dakota NASA EPSCoR)
• North Dakota Agricultural Experiment Station

Citation

@article{Progga2025chipbased,
  author = {Progga, Jannatul Ferdaous and Zhang, Xiaomo and Maiti, Srabana and Sun, Xin and Dey, Shuvashis and Eshkabilov, Sulaymon and Feng, Xiaoyu},
  title = {A chip-based radio frequency sensor for soil moisture measurements: A machine learning and deep learning calibration approach},
  journal = {Journal of Agriculture and Food Research},
  year = {2025},
  doi = {10.1016/j.jafr.2025.102591},
  url = {https://doi.org/10.1016/j.jafr.2025.102591}
}