The Retrieval of 30-m Resolution LAI from Landsat Data by Combining MODIS Products

Jianmin Zhou, Shan Zhang, Hua Yang, Zhiqiang Xiao, Feng GaoRemote 2018 · DOI

TL;DR

This paper presents a novel approach to generate high-resolution (30-m) Leaf Area Index (LAI) maps by combining moderate-resolution MODIS data with high-resolution Landsat imagery. They developed two methods to extract training samples from MODIS products and used Support Vector Regression (SVR) to retrieve LAI, achieving good agreement with field measurements.

What Problem Does This Paper Address?

Leaf Area Index (LAI) is a crucial parameter for understanding vegetation health and ecosystem processes. While global LAI products exist, their spatial resolution (often 500m or coarser) is too coarse for detailed regional studies, such as monitoring crop growth or analyzing canopy structure at a fine scale. Traditional methods for retrieving high-resolution LAI rely on field measurements for training, which are difficult to obtain in sufficient quantity and quality. This paper addresses the challenge of generating high-resolution LAI by leveraging readily available, albeit lower-resolution, MODIS data to create robust training samples for a high-resolution retrieval using Landsat data.

Methods

The core of this research lies in developing effective methods to extract high-quality training samples from MODIS products for use with a Support Vector Regression (SVR) algorithm. Two primary methods were employed:

  • Homogeneous and Pure Pixel Filter Method (Method A): This method first identifies agricultural land areas and then filters MODIS LAI pixels based on quality control flags to ensure they were derived from the main algorithm and are not saturated. Subsequently, it uses the coefficient of variation (CV) of Landsat reflectance within MODIS pixels to select homogeneous and pure pixels. Two datasets were generated using this method:
  • Dataset A1: Upscaled Landsat reflectance (aggregated to 500m) and MODIS LAI.
  • Dataset A2: MODIS reflectance and MODIS LAI.
  • Pixel Unmixing Method (Method B): This method aims to decompose mixed MODIS pixels into their constituent endmembers. It uses a linear unmixing model to derive 30-m resolution LAI and reflectance from MODIS data, leveraging 30-m land cover maps. The unmixed MODIS LAI and reflectance of agricultural land pixels were then used as training samples.

An SVR algorithm was trained using these three datasets (A1, A2, and B) to build inversion models. These models then used Landsat surface reflectance as input to retrieve 30-m resolution LAI.

Datasets

  • MODIS Products:
    • MODIS land-cover type product (MCD12Q1)
    • MODIS LAI products (MCD15A2H)
    • MODIS reflectance products (MOD09A1)
  • Landsat Data:
    • Landsat 8 OLI and Landsat 5 TM/ETM+ surface reflectance products at 30-m spatial resolution.
  • Field Measurements:
    • LAI measurements from the 2002 Soil Moisture Experiment (SMEX02) in Des Moines, Iowa, USA.
  • LAI measurements from Baoding, Hebei Province, China.

No new benchmark datasets were introduced; instead, the study focused on generating training samples from existing global products.

Key Insights and Ideas

  • Leveraging MODIS for High-Resolution Training: The study demonstrates that MODIS LAI and reflectance products, despite their coarser resolution, can be effectively used to generate high-quality training samples for retrieving high-resolution LAI from Landsat data. This circumvents the limitations of relying solely on scarce field measurements.
  • Pixel Unmixing for Enhanced Resolution: The pixel unmixing method (Method B) showed particular promise by directly deriving 30-m resolution training data, leading to slightly better retrieval accuracy. This suggests that de-mixing coarse-resolution data can effectively bridge the spatial resolution gap.
  • SVR's Suitability for Nonlinear Relationships: The use of Support Vector Regression (SVR) proved effective in capturing the complex, nonlinear relationships between vegetation indices, reflectance, and LAI, which is crucial for accurate remote sensing retrievals.
  • Importance of Training Sample Quality: The study highlights that while both methods yielded good results, the quality and representativeness of the training samples are paramount for accurate LAI retrieval. Using multi-year data and quality control flags for MODIS LAI contributed to robust training sets.

Key Results

  • All three developed SVR inversion models (based on datasets A1, A2, and B) achieved good agreement with field measurements.
  • Accuracy Metrics:
    • For the SMEX02 site, R² values ranged from 0.79 to 0.82, and RMSE values were between 0.65 and 0.73.
  • For the Baoding site, R² values were around 0.77-0.79, and RMSE values were around 0.41-0.49.
  • The pixel unmixing method (dataset B) generally yielded slightly better results (higher R², lower RMSE) compared to the homogeneous and pure pixel filter method (datasets A1 and A2).
  • Within Method A, using MODIS reflectance and LAI (dataset A2) performed slightly better than using upscaled Landsat reflectance and MODIS LAI (dataset A1).
  • The retrieved LAI time series from Landsat data were consistent with MODIS LAI trends and showed good agreement with spatial vegetation distribution.

Limitations and Future Work

  • LAI Saturation: The study observed underestimation of LAI when values exceeded 3, particularly in July. This is attributed to the saturation of the relationship between surface reflectance and LAI at higher values and the potential under-representation of high LAI samples in the training data (due to MODIS main algorithm limitations).
  • MODIS Main Algorithm Bias: The reliance on MODIS quality control flags meant that samples derived from the main algorithm were prioritized, potentially excluding high LAI values where the backup algorithm might have been used.
  • Scaling Effects: Differences in spatial resolution between MODIS and Landsat, and the aggregation/unmixing processes, can introduce scale effects that influence retrieval accuracy.
  • Geometrical Registration: The study did not perform precise geometrical registration between Landsat and MODIS data, which could introduce bias in the CV calculation for homogeneous pixels.
  • CV Threshold Determination: The CV threshold for pure pixels was determined based on subjective experience and may need further quantification and validation across different study sites.
  • Spectral Response Function Mismatch: Differences in spectral response functions between Landsat and MODIS bands, particularly in the near-infrared, were noted as a potential source of error. Converting between these bands using spectral response functions is suggested for future work.
  • Generalizability: While tested on two sites with different crops, further validation in other regions, especially tropical areas with frequent cloud cover, is needed to assess the broader applicability of the approach.
  • Biophysical Mechanisms: For large-scale applications, incorporating biophysical mechanisms of vegetation into the models might be beneficial.

Summary updated

Abstract

Leaf area index (LAI) is a critical vegetation structural parameter in biogeochemical and biophysical ecosystems. High-resolution LAI products play an essential role in regional studies. Empirical methods, which normally use field measurements as their training samples and have been identified as the most commonly used approaches to retrieve structural parameters of vegetation from high-resolution remote-sensing data, are limited by the quality of training samples. Few efforts have been made to generate training samples from existing global LAI products. In this study, two methods (a homogeneous and pure pixel filter method (method A) and a pixel unmixing method (method B)) were developed to extract training samples from moderate-resolution imaging spectroradiometer (MODIS) surface reflectance and LAI products, and a support vector regression (SVR) algorithm trained by the samples was used to retrieve the high-resolution LAI from Landsat data at Baoding, situated in the Hebei Province in China, and Des Moines, situated in Iowa, United States. For the homogeneous and pure pixel filter method, two different sets of training samples were designed. One was composed of upscaled Landsat reflectance at the 500-m resolution and MODIS LAI products (dataset A1); the other was composed of MODIS reflectance and LAI products (dataset A2). With them, two inversion models were developed using SVR. For the pixel unmixing method, the training samples (dataset B) were extracted from unmixed MODIS surface reflectance and LAI products at 30-m resolution, and the third inversion model was obtained with them. LAI inversion results showed that good agreement with field measurements was achieved using these three inversion models. The R2 (coefficient of determination) value and the root mean square error (RMSE) value were computed to assess the results. For all tests, the R2 values are higher than 0.74 and RMSE values are less than 0.73. These tests showed that three models for the two methods combined with MODIS products can retrieve 30-m resolution LAI from Landsat data. The results of the pixel unmixing method was slightly better than that of the homogeneous and pure pixel filter method.
BibTeX
@inproceedings{zhou2018retrieval,
  author = {Jianmin Zhou and Shan Zhang and Hua Yang and Zhiqiang Xiao and Feng Gao},
  title = {{The Retrieval of 30-m Resolution LAI from Landsat Data by Combining MODIS Products}},
  booktitle = {Remote},
  year = {2018},
  doi = {10.3390/rs10081187},
}
Figure