Evaluation and Optimization of Prediction Models for Crop Yield in Plant Factory

Yaoqi Peng, Yudong Zheng, Zengwei Zheng, Yong HePlants 2025 · DOI

TL;DR

This paper introduces a novel approach for predicting crop yield in plant factories by precisely capturing crop canopy images and removing background noise. They developed a method to accurately measure crop canopy area and then evaluated 28 prediction models, identifying a Wide Neural Network as the most optimal for its high accuracy, speed, and small model size, making it suitable for real-time applications.

What Problem Does This Paper Address?

Traditional agriculture faces significant challenges like land scarcity, climate change, and environmental pollution. Plant factories offer a controlled environment solution, but accurate crop yield prediction within these settings is crucial for optimizing growth and resource management. Existing methods often rely on expensive sensors or complex setups, limiting their practicality. This study aims to bridge this gap by developing a low-cost, efficient method for crop yield prediction in plant factories, focusing on accurate canopy image analysis and model selection.

Methods

The core of this research lies in its image processing pipeline and model evaluation.

  • Crop Canopy Image Acquisition: High-resolution images of black leaf Chinese cabbage were captured using a smartphone from a consistent top-down perspective, approximately 80 cm above the crop. A ruler was included in the frame for scale calibration.
  • Background Interference Removal: Manual background removal was employed to isolate the crop canopy.
  • Crop Canopy Projection Area (CCPA) Recognition: A combination of color features, morphological operations (erosion and dilation), and an active contour model was used to precisely identify and extract the crop canopy boundary.
  • Spatial Resolution Calibration: The ruler in the images was used to establish a pixel-to-real-world measurement mapping. Outliers were removed using the interquartile range (IQR) method to determine a spatial resolution of 0.078 mm/pixel.
  • Crop Canopy Area Calculation: The recognized canopy outline was used to compute the physical area of the crop canopy.
  • Yield Prediction Model Evaluation: 28 different prediction models were trained and evaluated using metrics like Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), R-squared (R²), prediction speed, training time, and model size.

Datasets

The study utilized a dataset of 60 black leaf Chinese cabbage canopy images collected from a plant factory facility. These images were captured over a period of approximately 30 days, covering various growth stages from seedling to maturity. The dataset was used for both training and validation of the prediction models. No new benchmark dataset was introduced; rather, existing images were processed and utilized.

Key Insights and Ideas

  • Smartphone as a Viable Imaging Tool: The research demonstrates that a standard smartphone can be effectively used for high-quality crop canopy imaging in plant factories, offering a low-cost alternative to specialized equipment.
  • Robust Image Processing Pipeline: The combination of morphological operations and active contour models proved highly effective in accurately segmenting crop canopies, even with potential background noise and variations in lighting.
  • Importance of Spatial Calibration: Precisely calibrating the image resolution using a physical ruler is critical for accurate area calculation, which directly impacts yield prediction.
  • Wide Neural Network's Superiority: The study highlights that among 28 models, the Wide Neural Network model stands out for its balanced performance across multiple critical metrics, especially for real-time applications.

Key Results

  • CCPA Recognition Accuracy: The crop canopy projection area (CCPA) recognition achieved a high coefficient of determination (R²) of 0.98.
  • Spatial Resolution: A spatial resolution of 0.078 mm/pixel was determined for the captured images.
  • Optimal Model Performance: The Wide Neural Network model was identified as the most optimal.
  • Predictive Accuracy: Achieved an R² of 0.95, an RMSE of 27.15 g, and a MAPE of 11.74%.
  • Prediction Speed: Demonstrated a remarkable speed of 60,234.9 observations per second.
  • Model Size: A compact size of 7039 bytes, making it suitable for efficient deployment.
  • Model Comparison: The study provided a comprehensive comparison of 28 models, illustrating trade-offs between accuracy, speed, and size.

Limitations and Future Work

  • Color Feature Reliance: The current crop contour recognition algorithm relies heavily on color features, which can be a limitation when dealing with significant color variations in leaves or challenging lighting conditions.
  • Future Work Suggestions: The authors suggest integrating additional features like shape and texture into the recognition algorithm to improve robustness and accuracy, especially in more complex environmental conditions. Further research could also explore optimizing the integration of multiple features for enhanced performance.

Summary updated

Abstract

This study focuses on enhancing crop yield prediction in plant factory environments through precise crop canopy image capture and background interference removal. This method achieves highly accurate recognition of the crop canopy projection area (CCPA), with a coefficient of determination (R<sup>2</sup>) of 0.98. A spatial resolution of 0.078 mm/pixel was derived by referencing a scale ruler and processing pixel counts, eliminating outliers in the data. Image post-processing focused on extracting the canopy boundary and calculating the crop canopy area. By incorporating crop yield data, a comparative analysis of 28 prediction models was performed, assessing performance metrics such as MSE, RMSE, MAE, MAPE, R<sup>2</sup>, prediction speed, training time, and model size. Among them, the Wide Neural Network model emerged as the most optimal. It demonstrated remarkable predictive accuracy with an R<sup>2</sup> of 0.95, RMSE of 27.15 g, and MAPE of 11.74%. Furthermore, the model achieved a high prediction speed of 60,234.9 observations per second, and its compact size of 7039 bytes makes it suitable for efficient, real-time deployment in practical applications. This model offers substantial support for managing crop growth, providing a solid foundation for refining cultivation processes and enhancing crop yields.
BibTeX
@inproceedings{peng2025evaluation,
  author = {Yaoqi Peng and Yudong Zheng and Zengwei Zheng and Yong He},
  title = {{Evaluation and Optimization of Prediction Models for Crop Yield in Plant Factory}},
  booktitle = {Plants},
  year = {2025},
  doi = {10.3390/plants14142140},
}
Figure