Leaf Area Index Estimation Using Chinese GF-1 Wide Field View Data in an Agriculture Region

Xiangqin Wei, Xingfa Gu, Qingyan Meng, Tao Yu, Xiang Zhou, Zheng Wei, Kun Jia, Chunmei Wang2017 · DOI

TL;DR

This paper presents a novel approach for estimating Leaf Area Index (LAI) using data from China's GF-1 Wide Field View (WFV) satellite. By combining radiative transfer modeling with neural networks, they achieve accurate LAI estimations in an agricultural region, demonstrating the potential of GF-1 WFV data for high-resolution vegetation monitoring.

What Problem Does This Paper Address?

Leaf Area Index (LAI) is a crucial parameter for understanding vegetation health, ecosystem processes, and agricultural productivity. While global LAI products exist, they often suffer from low spatial resolution (kilometric scale), making them less suitable for detailed agricultural monitoring. The Chinese GF-1 satellite, with its Wide Field View (WFV) sensor, offers multi-spectral data at a decametric resolution (16m) with high temporal frequency, presenting a valuable opportunity for more localized LAI estimation. However, a specific algorithm tailored for GF-1 WFV data was needed to leverage its potential. The challenge lies in establishing a robust relationship between the satellite's spectral reflectance data and the biophysical parameter of LAI, especially considering the complexities of canopy reflectance.

Methods

The core of this study is the development of an automatic LAI estimation algorithm for GF-1 WFV data. The approach combines physical modeling with machine learning:

  • Radiative Transfer Modeling (PROSAIL): The PROSAIL model was used to simulate the physical relationship between canopy reflectance and LAI. This model accounts for leaf optical properties (PROSPECT model) and canopy structure (SAIL model), along with soil background reflectance. This simulation process generated a comprehensive training dataset of spectral reflectances and corresponding LAI values.
  • Neural Networks (BPNNs): Backpropagation Neural Networks (BPNNs) were employed to learn the complex, non-linear relationships between the simulated canopy reflectances and LAI. The input variables for the BPNNs were the green, red, and near-infrared (NIR) band reflectances from GF-1 WFV data, and the output was the estimated LAI. The network architecture was optimized, and a Levenberg–Marquardt minimization algorithm was used for training.
  • NDVI Thresholding: To distinguish vegetation from non-vegetation pixels and avoid abnormal LAI estimations, a Normalized Difference Vegetation Index (NDVI) threshold of 0.05 was applied. Pixels below this threshold were assigned an LAI of zero.

Datasets

  • GF-1 WFV Data: Multi-spectral data from the GF-1 Wide Field View (WFV) sensor were used. Five GF-1 WFV datasets acquired between June and September 2014 were selected, closely matching the field survey dates. These datasets underwent pre-processing including radiance calibration, atmospheric correction (using FLAASH), and geometric correction.
  • Field LAI Measurements: Ground-truth LAI measurements were collected using an LAI-2000 plant canopy analyzer at 23 representative sample sites within the study area. These measurements were taken on four different dates during the maize growing season (June, July, August, and September 2014) to validate the algorithm's performance across different growth stages.

Key Insights and Ideas

  • Bridging the Gap: The study successfully bridges the gap between low-resolution global LAI products and the need for high-resolution data in agricultural applications by leveraging GF-1 WFV data.
  • Synergy of Physics and Machine Learning: The combination of a physically-based radiative transfer model (PROSAIL) for generating training data and neural networks for inversion offers a powerful and flexible approach to LAI estimation. This avoids the limitations of purely empirical methods.
  • Adaptability of the Algorithm: By training the neural networks on a diverse set of simulated conditions (varying soil and vegetation parameters), the algorithm is designed to be general and adaptable to different land surface conditions.
  • Importance of WFV Data: The results highlight the significant potential of GF-1 WFV data for providing high spatial and temporal resolution LAI information, crucial for dynamic monitoring and management.

Key Results

The developed LAI estimation algorithm demonstrated satisfactory accuracy when validated against field measurements:

  • Accuracy Metrics: The validation using field LAI measurements yielded a coefficient of determination (R²) of 0.818, a Root Mean Square Error (RMSE) of 0.50, and a Nash–Sutcliffe model efficiency coefficient (EF) of 0.797.
  • Spatial and Temporal Patterns: Visual inspection of the LAI estimates showed reasonable spatial and temporal distributions that corresponded well with the known growth stages of maize in the study area, from early emergence to peak growth and decline.
  • Operational Potential: The algorithm was designed to be automatically operated without requiring prior knowledge of land cover or manual parameter adjustments, indicating its potential for operational LAI product generation.

Limitations and Future Work

  • Validation Data: The field LAI measurements were collected from a single agricultural region dominated by maize. The authors suggest that future work should include validation with more diverse vegetation types and broader geographical areas to further assess the algorithm's generalizability.
  • Interpolation Uncertainty: For the September 5th LAI estimation, interpolation from data acquired on August 24th and September 18th was used due to cloud cover. This interpolation might introduce uncertainties, especially given the rapid growth and decline of maize at this stage.
  • Automated Pre-processing: The pre-processing of GF-1 WFV data was conducted manually. Developing an automated pre-processing workflow would significantly simplify and expedite the LAI estimation process.
  • Matching Field and Pixel Data: Accurately matching field measurements with pixel-level estimates remains a challenge due to the non-absolute uniformity of land surfaces. The authors suggest exploring advanced data collection methods, such as unmanned aerial systems, for more precise ground-truth data.

The study concludes that the proposed LAI estimation algorithm for GF-1 WFV data is reliable and has the potential to provide valuable high-resolution LAI datasets for various applications. Future research will focus on validating the algorithm with less uncertain field measurements across diverse land cover types.

Summary updated

BibTeX
@misc{wei2017leaf,
  author = {Xiangqin Wei and Xingfa Gu and Qingyan Meng and Tao Yu and Xiang Zhou and Zheng Wei and Kun Jia and Chunmei Wang},
  title = {{Leaf Area Index Estimation Using Chinese GF-1 Wide Field View Data in an Agriculture Region}},
  year = {2017},
  doi = {10.3390/s17071593},
}
Figure