Estimation of Daylily Leaf Area Index by Synergy Multispectral and Radar Remote-Sensing Data Based on Machine-Learning Algorithm

Mingzhou Hu, Jingshu Wang, Peng Yang, Ping Li, Peng He, Rutian Bi

TL;DR

This paper introduces a machine learning approach to accurately estimate the Leaf Area Index (LAI) of daylilies by combining optical and radar satellite data. By using Random Forest feature selection, they identified key spectral and backscatter features that significantly improve LAI estimation accuracy, outperforming models that rely on single data sources.

What Problem Does This Paper Address?

Accurately measuring Leaf Area Index (LAI) is crucial for understanding plant health, predicting crop yields, and managing agricultural fields. Traditional methods of measuring LAI are time-consuming, labor-intensive, and destructive, making them impractical for large-scale monitoring. While remote sensing offers a promising alternative, single-source data (either optical or radar) has limitations. Optical data can be affected by clouds and rain, and its spectral signals can become saturated at high LAI values. Radar data, while less affected by weather, provides less spectral information and can be influenced by soil background. This study aims to overcome these limitations by synergistically combining optical and radar remote-sensing data to achieve more robust and accurate LAI estimation for daylilies.

Methods

The researchers employed a machine learning framework to estimate daylily LAI. The core of their approach involved:

  • Feature Engineering: They calculated various vegetation indices (VIs) from Sentinel-2 optical data and extracted backscatter coefficients (VV, VH, VV-VH, etc.) from Sentinel-1 radar data.
  • Random Forest (RF) Feature Selection: A key innovation was using the Random Forest algorithm's feature importance score. This method helps identify the most relevant features for LAI estimation, reducing data redundancy and potentially improving model performance and interpretability. The RF algorithm itself was chosen for its robustness against noise and overfitting, and its ability to handle complex, non-linear relationships.
  • Machine Learning Models: Three common machine learning algorithms were used for LAI inversion:
  • Partial Least Squares Regression (PLSR): A linear regression model that handles multicollinearity.
  • Artificial Neural Network (ANN): A biologically inspired model capable of learning complex patterns.
  • Random Forest Regression (RFR): An ensemble method that builds multiple decision trees.
  • Model Comparison: The performance of models built using single-source data (optical VIs only, or radar backscatter only) and multisource data (combined VIs and backscatter) was compared. Feature selection was applied to all model configurations.

Datasets

The study utilized the following datasets:

  • Ground Truth Data: Field-measured LAI data for daylilies collected in Yunzhou District, Datong City, Shanxi Province, China, during July 2023. Measurements were taken using a LI-COR LAI-2000C plant canopy analyzer. A total of 47 sampling sites were used, with LAI values ranging from 2.37 to 7.84.
  • Remote Sensing Data:
    • Sentinel-1A SAR data: Interferometric Wide Swath (IW) Ground Range Detected (GRD) product, acquired on July 17, 2023, including VV and VH polarization modes.
  • Sentinel-2A multispectral data: L2A-level images acquired on July 19, 2023, covering the study area.

Key Insights and Ideas

  • Synergy is Key: Combining optical and radar data significantly outperforms single-source approaches for daylily LAI estimation. This highlights how different data types capture complementary information about plant structure and physiology.
  • Feature Selection Matters: Random Forest feature importance is a powerful tool for identifying the most informative spectral indices and radar backscatter features. This not only improves model accuracy but also provides insights into what aspects of the daylily canopy are most sensitive to LAI changes.
  • Specific Features Shine: The study found that vegetation indices like LSWI, GDVI, and NDYI, along with radar backscatter coefficients VV and VH, were particularly important. This suggests that canopy water content (LSWI), general vegetation health (GDVI), flower color (NDYI), and canopy structure/moisture (VV, VH) are critical indicators of daylily LAI.
  • RFR's Robustness: The Random Forest Regression (RFR) algorithm demonstrated superior performance compared to PLSR and ANN, likely due to its inherent ability to handle non-linearities, reduce overfitting, and its relative insensitivity to noisy data.

Key Results

  • Optimal Model: The RFR model using a combination of the top five features (LSWI, GDVI, NDYI, VV, and VH – termed RFR-VI-BC05) achieved the best LAI estimation accuracy.
  • Accuracy Metrics:
    • Training Data: For the RFR-VI-BC05 model, Pearson's r was >0.95, RMSE < 0.3, and NRMSE was 10.25%.
  • Validation Data: The RFR-VI-BC05 model achieved a Pearson's r of 0.711, an RMSE of 0.498, and an NRMSE of 9.10%.
  • Improvement over Single-Source: The synergistic RFR-VI-BC05 model showed a significant improvement in Pearson's r (increased by 0.104) and reductions in RMSE (by 0.123) and NRMSE (by 2.26%) compared to models using only optical or radar data after feature selection.
  • Spatial Distribution: The study also mapped the spatial distribution of daylily LAI, revealing higher values in the northwest of the study area and lower values in the southeast, which was correlated with the age and scale of daylily cultivation.

Limitations and Future Work

  • Limited Dataset Size: The authors acknowledge that the study used a limited dataset size, and further validation across different phenological phases is recommended.
  • Vigorous Growth Stage Focus: The analysis primarily focused on the vigorous growth stage of daylilies. Future work should investigate LAI estimation across various growth stages.
  • SAR Parameter Exploration: While VV and VH polarization modes were used, future research could explore optimizing LAI estimation by holistically utilizing other SAR-derived parameters beyond conventional vegetation indices.
  • Advanced Applications: The derived LAI inversions have implications for precision agriculture, suggesting potential for targeted fertilization, irrigation optimization, and the development of 3D virtual crop growth models.

Summary updated

Abstract

Rapid and accurate leaf area index (LAI) determination is important for monitoring daylily growth, yield estimation, and field management. Because of low estimation accuracy of empirical models based on single-source data, we proposed a machine-learning algorithm combining optical and microwave remote-sensing data as well as the random forest regression (RFR) importance score to select features. A high-precision LAI estimation model for daylilies was constructed by optimizing feature combinations. The RFR importance score screened the top five important features, including vegetation indices land surface water index (LSWI), generalized difference vegetation index (GDVI), normalized difference yellowness index (NDYI), and backscatter coefficients VV and VH. Vegetation index features characterized canopy moisture and the color of daylilies, and the backscatter coefficient reflected dielectric properties and geometric structure. The selected features were sensitive to daylily LAI. The RFR algorithm had good anti-noise performance and strong fitting ability; thus, its accuracy was better than the partial least squares regression and artificial neural network models. Synergistic optical and microwave data more comprehensively reflected the physical and chemical properties of daylilies, making the RFR-VI-BC05 model after feature selection better than the others ( r = 0.711, RMSE = 0.498, and NRMSE = 9.10%). This study expanded methods for estimating daylily LAI by combining optical and radar data, providing technical support for daylily management.
BibTeX
@inproceedings{hu2025estimation,
  author = {Mingzhou Hu and Jingshu Wang and Peng Yang and Ping Li and Peng He and Rutian Bi},
  title = {{Estimation of Daylily Leaf Area Index by Synergy Multispectral and Radar Remote-Sensing Data Based on Machine-Learning Algorithm}},
  booktitle = {Agronomy},
  year = {2025},
  doi = {10.3390/agronomy15020456},
}
Figure