A photosynthetic rate prediction model using improved RBF neural network

Liuru Pu, Yuan-Fang Li, Pan Gao, Haihui Zhang, Jin HuScientific 2022 · DOI

TL;DR

Researchers have developed a high-precision model to predict plant photosynthesis by incorporating "light quality" (spectral composition) alongside traditional variables like temperature and intensity. By optimizing a Radial Basis Function (RBF) neural network with a Quantum Genetic Algorithm (QGA), the team achieved a near-perfect prediction accuracy ($R^2 = 0.996$), significantly outperforming standard artificial intelligence approaches.

What Problem Does This Paper Address?

In precision agriculture and greenhouse management, regulating the light environment is the primary lever for improving crop yield and quality. To do this effectively, engineers need a "photosynthetic rate prediction model"—a mathematical way to tell how a plant will respond to specific environmental changes.

However, existing models face two major hurdles:

  1. Dimensionality Gap: Most models focus only on light intensity and temperature, ignoring "light quality" (the specific ratio of wavelengths, such as red vs. blue light), which profoundly impacts plant physiology.
  2. Algorithmic Limitations: Photosynthesis is a highly non-linear biological process. Current models often suffer from low prediction accuracy or slow modeling speeds, making them difficult to use for real-time environmental regulation.

Methods

The authors propose a hybrid approach called QGA-RBF:

  • RBF Neural Network: The core is a Radial Basis Function network, chosen for its excellent non-linear mapping capabilities and fast learning speeds compared to standard back-propagation networks.
  • The "Spread" Optimization: The performance of an RBF network depends heavily on a parameter called "spread." If it's too small, the model overfits; too large, and it fails to capture fast changes.
  • Quantum Genetic Algorithm (QGA): This is the paper's primary innovation. Standard Genetic Algorithms (GAs) can get stuck in "local optima" (good-enough solutions that aren't the best). QGA uses quantum bit representation and quantum gates to explore the solution space more effectively, quickly finding the global optimal "spread" for the neural network.

The model uses three inputs—temperature, light intensity, and light quality—to predict one output: the net photosynthetic rate.

Datasets

The researchers generated a custom dataset using "golden embryo 2" cucumber seedlings:

  • Equipment: They used a LI-6800 portable photosynthesis system to record data under controlled conditions.
  • Variables: The dataset included 720 groups of measurements across five temperatures (18–30°C), 16 levels of light intensity (0–1600 μmol/m²·s), and nine red/blue light mass ratios (10% to 90% blue light).
  • Preprocessing: The Dixon criterion was used to eliminate outliers (reducing the set to 675 groups), and data was normalized to a [0, 1] range to improve training stability.

Key Insights and Ideas

  • Light Quality is a Critical Dimension: The study proves that light quality isn't just a secondary factor. Adding it as a model input increased the determinant coefficient ($R^2$) from 0.981 to 0.996.
  • Blue Light as a Stress Buffer: A fascinating biological insight emerged: increasing the proportion of blue light can mitigate the negative effects of temperature stress. Blue light influences stomatal opening and osmotic regulation, helping the plant maintain photosynthesis even when temperatures are sub-optimal.
  • The Quantum Advantage: The QGA-RBF approach proved more robust than standard GA-optimized models. It converged to a minimum mean square error of 0.273 within just 15 generations without becoming "volatile" or getting stuck in local plain areas.

Key Results

The QGA-RBF model demonstrated superior performance across several metrics:

  • Near-Perfect Correlation: The model achieved an $R^2$ of 0.996, a straight slope of 1.000, and a tiny intercept of 0.061, indicating the predicted values almost perfectly match measured data.
  • Low Error: The Mean Absolute Error (MAE) was only 0.337 μmol/(m²·s), significantly lower than the standard RBF (0.874) or GRNN (1.371) models.
  • Efficiency: Despite the added complexity of quantum optimization, the computing time was a mere 0.040 seconds, making it viable for real-time agricultural control systems.

Limitations and Future Work

  • Species Specificity: The model was trained and validated specifically on cucumber seedlings. While the QGA-RBF method is generalizable, the specific weights of the model would likely need to be retrained for different crop species.
  • Fixed CO2 Levels: The experiment held CO2 concentration constant at 400 μL/L. In actual greenhouse environments, CO2 fluctuates. Future iterations could incorporate CO2 concentration as a fourth input dimension for even greater realism.
  • Hardware Integration: While the paper establishes the mathematical framework, the next step would be integrating this model into automated greenhouse "smart" controllers to test its efficacy in a live production environment.

Summary updated

Abstract

A photosynthetic prediction rate model is a theoretical basis for light environmental regulation, and the existing photosynthetic rate prediction models are limited by low modeling speed and prediction accuracy. Therefore, this paper analyses effects of light quality on photosynthesis rate, and proposes a method based on Radial basis function (RBF) optimized by Quantum genetic algorithm (QGA) to establish photosynthetic rate prediction model. We selected "golden embryo<sup>2</sup> formula 98-1F1" cucumber seedlings as experimental material and used LI-6800 to record the photosynthetic rates under different temperatures, light intensities and light quality. Experimental data is used to train and test the proposed model. The determinant coefficient of the model between the predicted and the measured values is 0.996, the straight slope of linear fitting is 1.000, and the straight intercept of linear fitting is 0.061. Moreover, the proposed method is compared with 6 artificial intelligence algorithms. The comparison results also validate that the proposed model has the highest accuracy compared with other algorithms.
BibTeX
@inproceedings{pu2022photosynthetic,
  author = {Liuru Pu and Yuan-Fang Li and Pan Gao and Haihui Zhang and Jin Hu},
  title = {{A photosynthetic rate prediction model using improved RBF neural network}},
  booktitle = {Scientific},
  year = {2022},
  doi = {10.1038/s41598-022-12932-9},
}
Figure