Data-Driven Methods for the Estimation of Leaf Water and Dry Matter Content: Performances, Potential and Limitations

Bin Yang, Hui Lin, Yuhao He2020 · DOI

TL;DR

This paper compares three data-driven methods for estimating leaf water (EWT) and dry matter (LMA) content using spectral data. The authors found that combining vegetation indices with machine learning (ML-VI-based method) significantly improved accuracy over using vegetation indices alone or raw spectral reflectance with ML, suggesting this hybrid approach is promising for precise agriculture and vegetation monitoring.

What Problem Does This Paper Address?

Accurately measuring leaf water (equivalent water thickness, EWT) and dry matter (leaf mass per area, LMA) is crucial for understanding plant health, predicting crop yields, and managing agriculture precisely. While remote sensing offers a non-destructive way to estimate these traits, existing data-driven methods have limitations. Physical models can be computationally expensive, and previous data-driven approaches often focused on canopy-level measurements or didn't fully leverage the strengths of both vegetation indices (VIs) and machine learning (ML). Specifically, there was a lack of direct comparison between different data-driven methods at the leaf level, and it was unclear how well ML models perform when using VIs as inputs versus raw spectral data, especially considering that some spectral regions might be less informative or even detrimental for certain traits.

Methods

The study compares three data-driven methods:

  • VI-based method (M1): Builds a statistical relationship (linear or exponential) between selected vegetation indices (VIs) and EWT/LMA.
  • ML-reflectance-based method (M2): Uses machine learning techniques (KNN, PLSR, SVR, RF) with raw leaf reflectance data across the 900-2400 nm spectral range as input to estimate EWT/LMA.
  • ML-VI-based method (M3): Combines VIs (selected for their sensitivity to EWT or LMA) as inputs to the same ML techniques to estimate EWT/LMA.

The ML techniques (KNN, PLSR, SVR, RF) were chosen for their popularity and effectiveness in regression tasks. Parameter optimization for these ML models was performed using K-fold cross-validation. To assess generalization, three sampling strategies were employed: using one dataset for training and the other for validation (S1, S2), and mixing both datasets for training and validation (S3). Performances were evaluated using Root Mean Square Error (RMSE) and coefficient of determination (R²).

Datasets

The study utilized two publicly available, independent datasets:

  • LOPEX (Leaf Optical Properties EXperiment): Collected in Italy in 1993, comprising 320 fresh leaf samples from 45 different species.
  • ANGERS (Anjou-Nantes-Green-Environment-Research-Service): Collected in France in 2003, comprising 276 leaf samples from 43 different species.

Both datasets contain synchronous measurements of leaf optical properties (reflectance and transmittance in the 400-2500 nm range) and leaf biochemical constituents (EWT and LMA). The 900-2400 nm spectral range was selected for analysis due to its higher sensitivity to EWT and LMA.

Key Insights and Ideas

  • VIs are powerful, but ML can enhance them: While VIs are effective, especially when combined with ML, simply using raw spectral data with ML doesn't always yield superior results, particularly for LMA.
  • The "sweet spot" for LMA estimation is tricky: The 900-1300 nm spectral region, often used for LMA estimation, might contain less useful information or be confounded by water absorption, impacting ML model performance when using the full spectral range.
  • Hybrid ML-VI approach is the winner: Combining the sensitivity of VIs to specific biochemical traits with the nonlinear modeling capabilities of ML techniques (ML-VI-based method) offers the most accurate and robust estimation for both EWT and LMA.
  • Data quality and sampling strategy matter: The performance of data-driven methods is sensitive to the training/validation data split and potential noise or spectral resolution differences between datasets, highlighting the importance of generalization.

Key Results

  • ML-VI-based method (M3) outperformed others: This method consistently provided better accuracy (lower RMSE) for both EWT and LMA estimation across all sampling strategies compared to the VI-based (M1) and ML-reflectance-based (M2) methods.
  • Significant error reduction with M3: Compared to M2, M3 reduced RMSE by an average of 5.7% for EWT and a substantial 41.5% for LMA. Compared to M1, M3 reduced RMSE by 1.8% for EWT and 12.4% for LMA.
  • SVR was a strong performer: Support Vector Regression (SVR) was identified as a particularly effective ML technique, especially for EWT estimation, often yielding the best results within M2 and M3.
  • MSI and NDMI were key VIs: The Moisture Stress Index (MSI) for EWT and the Normalized Dry Matter Index (NDMI) for LMA were found to be the most sensitive VIs in the VI-based method.
  • R² values for EWT were generally higher: EWT estimation consistently achieved higher R² values (often >0.8) across methods compared to LMA estimation, where R² values were frequently below 0.8.

Limitations and Future Work

  • Generalization challenges: The performance of data-driven methods is heavily dependent on the training data. Discrepancies in noise levels, spectral resolution, and leaf species between datasets can limit generalization. The study suggests strategies like using expert knowledge and simulated data to improve this.
  • Wavelength selection and new VIs: The optimal wavelengths for VIs can vary by vegetation type. Future work should focus on identifying more consistent bands and defining new VIs that can capture more nuanced biochemical information.
  • Canopy-level extension: The current study focused on leaf-level measurements. Extending these data-driven methods, particularly the ML-VI approach, to the canopy level is a crucial next step. This will require accounting for additional confounding factors like canopy structure, potentially using canopy reflectance models.

Summary updated

BibTeX
@misc{yang2020data,
  author = {Bin Yang and Hui Lin and Yuhao He},
  title = {{Data-Driven Methods for the Estimation of Leaf Water and Dry Matter Content: Performances, Potential and Limitations}},
  year = {2020},
  doi = {10.3390/s20185394},
}

Loading related content...

Figure