A combined model of shoot phosphorus uptake based on sparse data and active learning algorithm

Tianli Wang, Yi Zhang, Haiyan Liu, Fei Li, Dayong Guo, Ning Cao, Yubin ZhangFrontiers in Plant Science 2025 · DOI

TL;DR

This paper introduces a novel combined model for predicting shoot phosphorus uptake (SPU) in maize using hyperspectral remote sensing data. By integrating two-dimensional correlation spectroscopy (2DCOS) and an active learning approach, the model achieves high accuracy even with limited field data, offering a promising tool for sustainable agricultural management and global soil phosphorus monitoring.

What Problem Does This Paper Address?

Excessive phosphorus (P) fertilizer application is a major environmental concern, leading to soil degradation and water pollution. Accurate monitoring of crop P uptake (SPU) is crucial for optimizing fertilizer use and promoting sustainable agriculture. However, obtaining large-scale, on-site SPU data is challenging due to the labor-intensive nature of traditional methods. Existing remote sensing approaches often struggle with data scarcity and the complexity of extracting relevant spectral information, especially for different plant parts. This paper aims to bridge this gap by developing a robust and efficient method for predicting SPU using hyperspectral data, even when field measurements are sparse.

Methods

The authors propose a multi-faceted approach:

  • Two-Dimensional Correlation Spectroscopy (2DCOS): This technique is used to analyze spectral data from different growth stages (filling and milk stages) and to enhance spectral features. Specifically, they employ:
  • 1Der-2DCOS: For analyzing spectral changes related to P concentration in maize leaves.
  • filling-milk-2T2DCOS: To couple spectral data from the filling and milk stages for a more comprehensive analysis.
  • Active Learning (NRBO-AL): To address the challenge of limited field data, an active learning strategy based on the Newton-Raphson-based optimizer (NRBO-AL) is utilized. This method iteratively selects the most informative simulated data points to refine the training dataset, improving model generalization and prediction accuracy with minimal sampling effort.
  • Hybrid Modeling: A hybrid approach is adopted, combining data-driven methods with physical models (PROSAIL-5B). This leverages the strengths of both, allowing for more accurate predictions of plant biomass and nutrient content.
  • Combined Prediction Model: The final SPU prediction is achieved by combining predictions for different plant parts:
  • Leaf P uptake is predicted using a hybrid model informed by active learning.
  • Stem P uptake is predicted using a polynomial regression model.
  • Grain P uptake is predicted using an empirical model.

Datasets

The study utilized hyperspectral data collected from maize plants in an experimental zone in Gaolong Town, Yanshi District, Luoyang City, Henan Province, China. The experiments involved different P fertilization treatments. The paper does not introduce new benchmark datasets but rather uses field data collected for this specific study.

Key Insights and Ideas

  • 2DCOS for Robust Spectral Analysis: The study highlights the power of 2DCOS techniques (especially 1Der-2DCOS and filling-milk-2T2DCOS) in extracting meaningful spectral relationships, even from small datasets. This is crucial for understanding complex plant-nutrient interactions.
  • Active Learning for Data Efficiency: The NRBO-AL method proves highly effective in optimizing simulated datasets, significantly reducing the need for extensive field sampling while maintaining or improving prediction accuracy. This is a game-changer for applications where data collection is costly and time-consuming.
  • Plant Part-Specific Prediction: Recognizing that different plant parts have varying P uptake dynamics, the authors developed a strategy to predict SPU by modeling leaves, stems, and grains separately. This offers a more nuanced and accurate representation of overall plant P uptake.
  • Filling Stage as Optimal Prediction Window: The research identified the filling stage as the most suitable period for pre-season SPU prediction, offering valuable insights for timely agricultural management decisions.

Key Results

  • The combined prediction model achieved a high prediction accuracy with an R² of 0.87, RMSE of 2.39 kg/ha, and relative percentage difference of 2.71%. This represents a significant improvement over simple empirical models.
  • The 2DCOS methods demonstrated robustness and effectiveness in extracting spectral trait relationships, enabling efficient prediction even with small sample sizes.
  • The hybrid model, constrained by the NRBO-AL active learning method, effectively filtered localized simulation data, leading to improved prediction accuracy.
  • The study demonstrated that mature leaf P concentration can be used as a proxy for grain P concentration in certain prediction scenarios.

Limitations and Future Work

  • Stem P Uptake Prediction: The prediction of stem P uptake using polynomial regression showed limitations, with concentrated predicted values and a relatively poor prediction effect. This is attributed to the complex and diverse relationship between leaf and stem P uptake, which is influenced by growth stage and environmental factors not fully accounted for in this study.
  • Interannual Planting Effects: The study acknowledges that interannual planting, a long-term cultivation factor, can influence P uptake and utilization, but this aspect was not deeply explored.
  • Spatial Variability: Applying the model to large-scale farmland faces challenges due to the relatively large spatial variability of P observed in small-scale crops.
  • Future Work: The authors suggest future research should focus on:
  • Training and validating models using satellite remote sensing data, considering appropriate pixel sizes and accurate sample localization.
  • Evaluating the P uptake prediction models for large-scale farmland to assess their accuracy in practical applications.
  • Further investigating the complex relationship between leaf and stem P uptake, considering various growth stages and environmental factors.
  • Collecting ground measurement samples and satellite images from different regions and crops to calibrate the model for broader applicability and continuous updating.

Summary updated

Abstract

The soil ecosystem has been severely damaged because of the increasingly severe environmental problems caused by excessive application of phosphorus (P) fertilizer, which seriously hinders soil fertility restoration and sustainable farmland development. Shoot P uptake (SPU) is an important parameter for monitoring crop growth and health and for improving field nutrition management and fertilization strategies. Achieving on-site measurement of large-scale data is difficult, and effective nondestructive prediction methods are lacking. Improving spatiotemporal SPU estimation at the regional scale still poses challenges. In this study, we proposed a combination prediction model based on some representative samples. Furthermore, using the experimental area of Henan Province, as an example, we explored the potential of the hyperspectral prediction of maize SPU at the canopy scale. The combination model comprises predicted P uptake by maize leaves, stems, and grains. Results show that (1) the prediction accuracy of the combined prediction model has been greatly improved compared with simple empirical prediction models, with accuracy test results of <i>R</i> <sup>2</sup> = 0.87, root mean square error = 2.39 kg/ha, and relative percentage difference = 2.71. (2) In performance tests with different sample sizes, two-dimensional correlation spectroscopy i.e., first-order differentially enhanced two-dimensional correlation spectroscopy (1Der-2DCOS) and two-trace 2DCOS of enhanced filling and milk stages (filling-milk-2T2DCOS)) can effectively and robustly extract spectral trait relationships, with good robustness, and can achieve efficient prediction based on small samples. (3) The hybrid model constrained by the Newton-Raphson-based optimizer's active learning method can effectively filter localized simulation data and achieve localization of simulation data in different regions when solving practical problems, improving the hybrid model's prediction accuracy. The practice has shown that with a small number of representative samples, this method can fully utilize remote sensing technology to predict SPU, providing an evaluation tool for the sustainable use of agricultural P. Therefore, this method has good application prospects and is expected to become an important means of monitoring global soil P surplus, promoting sustainable agricultural development.
BibTeX
@inproceedings{wang2025combined,
  author = {Tianli Wang and Yi Zhang and Haiyan Liu and Fei Li and Dayong Guo and Ning Cao and Yubin Zhang and Yubin Zhang and Yubin Zhang},
  title = {{A combined model of shoot phosphorus uptake based on sparse data and active learning algorithm}},
  booktitle = {Frontiers in Plant Science},
  year = {2025},
  doi = {10.3389/fpls.2024.1470719},
}
Figure