Monitoring the Maize Canopy Chlorophyll Content Using Discrete Wavelet Transform Combined with RGB Feature Fusion

Wenfeng Li, Kun Pan, Yue Huang, Guodong Fu, Wenrong Liu, Jiyu He, Weihua Xiao, Yi Fu, Jin Yu GuoAgronomy 2025 · DOI

TL;DR

This paper introduces a novel method for monitoring maize chlorophyll content using RGB images from drones, combining color, texture, and Discrete Wavelet Transform (DWT) features. Their approach, particularly with the Stacked Ensemble Learning (SEL) model, significantly improves accuracy over methods using only color and texture, offering a more efficient and scalable solution for precision agriculture.

What Problem Does This Paper Address?

Accurately monitoring crop health is crucial for optimizing agricultural yields and ensuring food security. Chlorophyll content (CHL) is a key indicator of plant health, but traditional methods like chemical analysis are destructive and time-consuming, while spectroscopic methods can be sensitive to environmental conditions. Existing remote sensing approaches using Unmanned Aerial Vehicles (UAVs) have shown promise, but there's a need for more robust and accurate methods, especially using readily available RGB imagery. This paper aims to bridge this gap by exploring how to extract richer information from RGB images, specifically by integrating advanced signal processing techniques like Discrete Wavelet Transform (DWT) with traditional color and texture features.

Methods

The researchers employed a multi-faceted approach:

  • Data Acquisition: High-resolution RGB images of maize canopies were captured using a DJI Phantom 4 RTK drone at different growth stages (jointing, tasseling, and grouting).
  • Feature Extraction:
    • Color and Texture Features: RGB images were converted to grayscale, and various color (e.g., R, G, B, Normalized Red, Normalized Green, Normalized Blue) and texture features (e.g., entropy, energy, correlation, contrast) were extracted. Image segmentation techniques were used to isolate maize plants from the background.
  • Discrete Wavelet Transform (DWT): A two-level decomposition of the images was performed using the db1 wavelet basis function. This process decomposes the image into different frequency sub-bands (LL, LH, HL, HH) and extracts energy features from these sub-bands. DWT is chosen for its ability to capture both global trends and local details, and to reduce noise.
  • Feature Fusion: Datasets were created using only color and texture features, and a more comprehensive dataset was generated by fusing wavelet features with color and texture features.
  • Machine Learning Models: Three machine learning models were trained and evaluated:
  • Backpropagation Neural Network (BP): A standard neural network model.
  • Stacked Ensemble Learning (SEL): An ensemble method that combines multiple base models (in this case, SVR and LightGBM) to improve predictive performance.
  • Gradient Boosting Decision Tree (GBDT): A powerful ensemble learning algorithm.
  • Model Evaluation: The models' performance was assessed using metrics like R² (coefficient of determination), RMSE (root mean square error), and NRMSE (normalized root mean square error) by comparing predicted CHL values against measured SPAD values.

Datasets

The study utilized a dataset of RGB images of maize canopies collected in 2023 from experimental sites in Mangshi Daxingzhai and Mabozi Village, Dehong Prefecture, Yunnan Province, China. The data was collected across three key growth stages: jointing, tasseling, and grouting. A total of 270 samples were collected for the full birth stage, with 216 used for training and 54 for testing the models. No new benchmark datasets were introduced; rather, existing field data was used for this specific experiment.

Key Insights and Ideas

  • DWT unlocks richer information: The integration of Discrete Wavelet Transform (DWT) features significantly enhances the monitoring accuracy of maize chlorophyll content compared to using only color and texture features. This highlights DWT's capability to extract subtle structural and textural information that is crucial for understanding plant health.
  • Feature fusion is key: Combining different types of features (color, texture, and wavelet) leads to more robust and accurate models. This demonstrates the power of a holistic approach to feature extraction in remote sensing.
  • SEL model excels: The Stacked Ensemble Learning (SEL) model consistently outperformed BP and GBDT, showcasing its effectiveness in handling complex relationships and achieving higher predictive accuracy.
  • RGB imagery's potential: The study reinforces the idea that even standard RGB imagery, when processed with advanced techniques like DWT and fused with other features, can yield highly accurate results for crop monitoring, making it a cost-effective alternative to more specialized sensors.
  • Feature importance matters: The red light parameter (R) was identified as the most important feature, followed by entropy (ENT) and the wavelet feature LH. This provides valuable insights into which aspects of the imagery are most indicative of chlorophyll content.

Key Results

  • Improved Accuracy with Fusion: Models using the fused dataset (wavelet + color + texture) consistently showed higher accuracy than those using only color and texture features.
  • For the fused dataset, the SEL model achieved the highest accuracy with an R² of 0.792, RMSE of 3.319, and NRMSE of 15.34%.
  • Compared to the color and texture-only dataset, the fused dataset led to an increase in R² by 0.018 to 0.049, a decrease in RMSE by 0.132 to 0.574, and a decrease in NRMSE by 0.44% to 1.8%.
  • Model Performance:
    • SEL model: R² = 0.792, RMSE = 3.319, NRMSE = 15.34% (with fused data).
  • BP model: R² = 0.728, RMSE = 3.911, NRMSE = 15.24% (with fused data).
  • GBDT model: R² = 0.756, RMSE = 3.730, NRMSE = 15.45% (with fused data).
  • Feature Importance: The red light parameter (R) was the most important feature (20.19%), followed by entropy (ENT) (12.35%) and wavelet feature LH (10.14%).

Limitations and Future Work

The authors acknowledge several limitations:

  • RGB Spectrum Limitation: Reliance on RGB data restricts the analysis to the visible spectrum (400-700 nm). This means the method cannot evaluate agronomic variables that require near-infrared or thermal spectral ranges, such as water stress or nutrient deficiencies.
  • Model Validation: The robustness of the SEL model has not been validated with multispectral or hyperspectral data.
  • Environmental Variability: Fluctuations in field conditions, like changing light intensities, can affect results.

Future work suggested includes:

  • Cross-validation with other sensors: Incorporating hyperspectral and thermal infrared sensors to enhance model reliability and broaden applicability.
  • Expanding sample size and metrics: Assessing feature fusion more comprehensively by increasing the sample size and using additional evaluation metrics.
  • Addressing environmental heterogeneity: Developing methods to better account for spatial and temporal variability in field conditions.

Summary updated

Abstract

To evaluate the accuracy of Discrete Wavelet Transform (DWT) in monitoring the chlorophyll (CHL) content of maize canopies based on RGB images, a field experiment was conducted in 2023. Images of maize canopies during the jointing, tasseling, and grouting stages were captured using unmanned aerial vehicle (UAV) remote sensing to extract color, texture, and wavelet features and to construct a color and texture feature dataset and a fusion of wavelet, color, and texture feature datasets. Backpropagation neural network (BP), Stacked Ensemble Learning (SEL), and Gradient Boosting Decision Tree (GBDT) models were employed to develop CHL monitoring models for the maize canopy. The performance of these models was evaluated by comparing their predictions with measured CHL data. The results indicate that the dataset integrating wavelet features achieved higher monitoring accuracy compared to the color and texture feature dataset. Specifically, for the integrated dataset, the BP model achieved an R2 value of 0.728, an RMSE of 3.911, and an NRMSE of 15.24%; the SEL model achieved an R2 value of 0.792, an RMSE of 3.319, and an NRMSE of 15.34%; and the GBDT model achieved an R2 value of 0.756, an RMSE of 3.730, and an NRMSE of 15.45%. Among these, the SEL model exhibited the highest monitoring accuracy. This study provides a fast and reliable method for monitoring maize growth in field conditions. Future research could incorporate cross-validation with hyperspectral and thermal infrared sensors to further enhance model reliability and expand its applicability.
BibTeX
@inproceedings{li2025monitoring,
  author = {Wenfeng Li and Kun Pan and Yue Huang and Guodong Fu and Wenrong Liu and Jiyu He and Weihua Xiao and Yi Fu and Jin Yu Guo},
  title = {{Monitoring the Maize Canopy Chlorophyll Content Using Discrete Wavelet Transform Combined with RGB Feature Fusion}},
  booktitle = {Agronomy},
  year = {2025},
  doi = {10.3390/agronomy15010212},
}
Figure