Li et al. (2025) Multi-source data fusion for estimating potato transpiration under water stress using machine learning models

Identification

• Journal: Agricultural Water Management
• Year: 2025
• Date: 2025-11-20
• Authors: Yida Li, Yuxin Wang, Yuqi Zhang, Liuyang Wang, Man Zhang, Han Li
• DOI: 10.1016/j.agwat.2025.109987

Research Groups

• Key Laboratory of Smart Agriculture System Integration, Ministry of Education, China Agricultural University, Beijing 100083, China
• Key Laboratory of Agricultural Information Acquisition Technology, Ministry of Agriculture and Rural Affairs, China Agricultural University, Beijing 100083, China

Short Summary

This study developed a multi-source data fusion framework to estimate daily cumulative potato transpiration under varying water stress by integrating image-derived canopy indices (Crop Water Stress Index and Relative Leaf Area Index) with meteorological measurements, demonstrating that this integration significantly enhances model performance and that optimal model choice depends on environmental stability.

Objective

• To accurately extract potato canopy Relative Leaf Area Index (RLAI) and Crop Water Stress Index (CWSI) based on image registration and semantic segmentation techniques.
• To systematically investigate the impact of multi-source input combinations of environmental and plant factors on the estimation accuracy of daily potato transpiration (T).
• To evaluate the applicability and performance of machine learning models, including Random Forest Regression (RFR), Back-Propagation Neural Network (BPNN), and Long Short-Term Memory (LSTM), in estimating potato transpiration under varying water stress conditions and environmental heterogeneity.

Study Configuration

• Spatial Scale: Plant-scale experiments conducted in 10 L pots within a controlled greenhouse environment at the National Precision Agriculture Research Base, Beijing, China (40°18′N, 116°45′E).
• Temporal Scale: Data collected over two distinct periods: November 25 to December 4, 2022, and March 21 to March 30, 2024, focusing on daily cumulative transpiration.

Methodology and Data

• Models used:
  • Machine Learning: Random Forest Regression (RFR), Back-Propagation Neural Network (BPNN), Long Short-Term Memory (LSTM).
  • Image Processing: PSPNet (for semantic segmentation, compared with Mask2Former and Segformer), improved Brain Storm Optimization (BSO) and Powell local search algorithm (for thermal-visible image registration).
  • Physiological Index Calculation: Idso (1982) method for CWSI, Li et al. (2020) method for RLAI.
• Data sources:
  • Image Data: Thermal infrared images (384 × 288 pixels, Guide IPT384 thermal camera) and visible light images (1280 × 720 pixels, Pixel XYZ binocular USB camera), acquired every 20 minutes.
  • Meteorological Variables: Air temperature (-40–80 °C, accuracy: ±0.3 °C), relative humidity (0–99.9 % RH, accuracy: ±3 % RH), photosynthetic photon flux density (0–65,535 Lux, accuracy ±5 % FS), and CO₂ concentration (0–10,000 ppm, accuracy ±40 ppm ±3 %), measured every 2 minutes via a LoRa wireless sensor network.
  • Plant Weight: Real-time weight measurements (maximum capacity 30 kg, accuracy ±1 g) from a weighing platform, logged every 2 minutes.
  • Substrate Moisture: Dielectric permittivity variations tracked every 2 minutes using a self-developed system with triaxial probes, TDR200 analyzer, and CR1000 datalogger.
  • Experimental Treatments: Six datasets collected over two years (2022 and 2024) under three water stress treatments: severe stress (T1, 20–35 % field capacity), moderate stress (T2, 45–55 % field capacity), and normal irrigation (T3, 65–75 % field capacity).
  • Time Features: Normalized hour and minute indicators.

Main Results

• Integrating CWSI and RLAI with meteorological variables significantly improved model performance, increasing R² values by 1.77–18.44 % for RFR, 3.44–11.87 % for BPNN, and 0.44–18.42 % for LSTM, while reducing RMSE and MAE.
• Under stable environmental conditions in 2022, RFR achieved the best accuracy (R² = 0.8851–0.9654, RMSE = 2.60–9.63 g, MAE = 1.83–6.06 g, RPD = 2.96–5.49).
• Under more variable environmental conditions in 2024, LSTM demonstrated superior performance (R² = 0.9187–0.9898, RMSE = 14.36–21.02 g, MAE = 10.92–14.62 g, RPD = 3.64–10.54).
• All models achieved optimal estimation accuracy under severe water stress (T1) due to smaller fluctuations in cumulative transpiration and more defined stomatal regulation.
• PSPNet exhibited superior performance for semantic segmentation of canopy and wet reference surface with an mIoU of 95.9, mPrecision of 97.87, and mDice of 97.87.
• The improved BSO_Powell algorithm for image registration achieved high precision, demonstrating comprehensive performance advantages over manual feature point selection (MI 0.9158, NMI 0.2639, MSSIM 0.5646).

Contributions

• Developed and validated a robust multi-source data fusion framework for estimating daily cumulative potato transpiration by integrating image-derived canopy indices (CWSI, RLAI) with meteorological data.
• Quantified the significant enhancement in machine learning model accuracy for transpiration estimation when incorporating crop physiological indicators, highlighting their crucial role alongside environmental variables.
• Provided empirical evidence for selecting appropriate machine learning models based on environmental stability, recommending RFR for stable conditions and LSTM for fluctuating environments.
• Generated high-precision, plant-scale daily cumulative transpiration data under controlled water stress conditions, serving as valuable benchmark truth values for model development.
• Contributed to the foundation for precision irrigation management in potato production by demonstrating improved reliability of transpiration estimations.

Funding

• National Natural Science Foundation of China (Grant No. 32171893)
• 2115 Talent Development Program of China Agricultural University, China

Citation

@article{Li2025Multisource,
  author = {Li, Yida and Wang, Yuxin and Zhang, Yuqi and Wang, Liuyang and Zhang, Man and Li, Han},
  title = {Multi-source data fusion for estimating potato transpiration under water stress using machine learning models},
  journal = {Agricultural Water Management},
  year = {2025},
  doi = {10.1016/j.agwat.2025.109987},
  url = {https://doi.org/10.1016/j.agwat.2025.109987}
}