Hyperspectral Estimation of Winter Wheat Leaf Area Index Based on Continuous Wavelet Transform and Fractional Order Differentiation

Changchun Li, Yilin Wang, Chunyan Ma, Fan Ding, Yacong Li, Weinan Chen, Jingbo Li, Zhen Xiao2021 · DOI

TL;DR

This paper introduces a novel approach for estimating winter wheat's Leaf Area Index (LAI) using hyperspectral data. By applying fractional order differentiation and continuous wavelet transform, the researchers were able to extract more sensitive spectral features, leading to improved LAI estimation models across different growth stages, particularly during flowering and filling.

What Problem Does This Paper Address?

Accurately and efficiently monitoring crop growth is crucial for agricultural management and yield prediction. Leaf Area Index (LAI) is a key indicator of crop health and productivity, representing the total area of leaves per unit ground area. Traditional methods for measuring LAI are destructive, time-consuming, and not scalable for large-area, continuous monitoring. Remote sensing offers a promising alternative, but extracting reliable LAI information from hyperspectral data can be challenging due to noise and the complex relationship between spectral reflectance and LAI. Existing methods often rely on the original spectrum, which may not capture all the subtle spectral variations indicative of LAI changes. This paper aims to overcome these limitations by employing advanced spectral preprocessing techniques to enhance the extraction of LAI-sensitive information.

Methods

The study utilizes two primary spectral preprocessing techniques:

  • Fractional Order Differentiation (FOD): This method generalizes integer-order differentiation, allowing for a more refined analysis of spectral data. By applying fractional orders (specifically, the Grünwald-Letnikov definition), the researchers aimed to denoise the spectral data and highlight subtle spectral features that are more sensitive to LAI variations, potentially overcoming the information loss associated with traditional integer-order differentiation.
  • Continuous Wavelet Transform (CWT): CWT acts like a "mathematical microscope," decomposing the hyperspectral signal into different scales (frequencies) and locations. This allows for the extraction of wavelet energy coefficients, which can capture localized spectral features and reduce noise. The study used CWT to decompose the hyperspectral data and generate 2D wavelet energy coefficients, which were then analyzed for their correlation with LAI.

These preprocessed spectral features were then used to build LAI estimation models using two machine learning algorithms:

  • Optimal Subset Regression (OSR): This method systematically searches for the best combination of independent variables (spectral features) to build a regression model, aiming to maximize the adjusted R-squared value.
  • Support Vector Machine (SVM): A powerful machine learning algorithm known for its robustness and ability to handle complex, non-linear relationships, SVM was used to construct LAI estimation models.

The study evaluated models built using both FOD and CWT features, combined with OSR and SVM, across four distinct growth stages of winter wheat: jointing, booting, flowering, and filling.

Datasets

The study utilized hyperspectral data and corresponding LAI measurements collected from winter wheat grown in the National Precision Agriculture Research and Demonstration Base in Xiaotangshan Town, Changping District, Beijing. The data were collected during the 2018 and 2019 growing seasons. Specific details about the experimental design, including nitrogen fertilizer levels, moisture content, and crop varieties, are provided, suggesting a well-controlled experimental setup.

Key Insights and Ideas

  • FOD Enhances Spectral Sensitivity: Fractional order differentiation proved effective in refining spectral information and highlighting features more strongly correlated with LAI, outperforming raw spectral data in some instances.
  • CWT Captures Multi-Scale Spectral Information: Wavelet energy coefficients derived from CWT provided a rich set of features that captured localized spectral characteristics, proving valuable for LAI estimation.
  • Growth Stage Matters: The optimal estimation method and its performance varied significantly across different growth stages, indicating that a one-size-fits-all approach is not ideal for LAI estimation.
  • Flowering and Filling Stages are Optimal: The study found that the flowering and filling stages offered the best opportunities for accurate LAI estimation, likely due to the significant canopy development and distinct spectral signatures during these periods.
  • SVM and OSR Complementary: Both OSR and SVM showed good performance, with SVM often providing slightly better or comparable results, especially when dealing with the complex spectral information extracted by CWT.

Key Results

The study achieved promising results in LAI estimation:

  • Flowering Stage: The best LAI estimation models were achieved during the flowering stage, with modeling and validation R² values of 0.87 and 0.71 (using CWT + OSR), and 0.87 and 0.57 (using FOD + SVM).
  • Filling Stage: The filling stage also yielded strong results, with modeling and validation R² values of 0.84 and 0.77 (using CWT + OSR), and 0.83 and 0.63 (using FOD + SVM).
  • Overall Performance: Across all growth stages, the study demonstrated that combining advanced spectral preprocessing (FOD and CWT) with robust modeling techniques (OSR and SVM) significantly improved LAI estimation accuracy compared to using raw spectral data. The maximum modeling R² achieved was 0.90 and the maximum validation R² was 0.77.

Limitations and Future Work

The authors acknowledge several limitations and suggest future research directions:

  • Empirical Models: The OSR and SVM models used are empirical and do not explicitly account for the radiative transfer processes of vegetation canopies and the atmosphere. Future work should aim to incorporate these physical processes to improve model stability.
  • Limited Data Scope: The study relied on data from a single experimental area. To enhance the generalizability and transferability of the models, future research should expand the dataset to include more crop species, larger sample sizes, and data from multiple years and locations.

Summary updated

BibTeX
@misc{li2021hyperspectral,
  author = {Changchun Li and Yilin Wang and Chunyan Ma and Fan Ding and Yacong Li and Weinan Chen and Jingbo Li and Zhen Xiao},
  title = {{Hyperspectral Estimation of Winter Wheat Leaf Area Index Based on Continuous Wavelet Transform and Fractional Order Differentiation}},
  year = {2021},
  doi = {10.3390/s21248497},
}
Figure