China - Shandong

Specific Project Objectives & Deliverables

Fraction of Absorbed Photosynthetically Active Radiation (FAPAR) Retrieval by Chlorophyll-related Vegetation Indices

Over the China Shandong JECAM site, chlorophyll-related vegetation indices (VIs) were selected and tested for their capability in crop FAPAR estimation using simulated Sentinel-2 data. These indices can be categorized into four classes:

  • ratio indices,
  • normalized difference indices,
  • triangular area based indices, and
  • integrated indices.

Two crops with distinctive canopy and leaf structure, wheat and corn, were studied. Regression analysis was conducted between measured FAPAR and different vegetation indices derived from Sentinel-2 reflectance simulated from field spectral measurements. At the same time, the effects of the red-edge reflectance on crop FAPAR estimation and the impact of different crop types on FAPAR estimation were explored. It is found that VIs using the near-infrared and red-edge reflectance, including the modified Simple Ratio2 (mSR2), the red-edge Simple Ratio (SR705), the Red-edge Normalized Difference Vegetation Index (ND705), MERIS terrestrial chlorophyll index (MTCI), and the Revised Optimized Soil-Adjusted Vegetation Index (OSAVI[705, 750]), were strongly correlated with FAPAR, especially in the high biomass range. Among all the indices, RDVI705 and mSR2 were more linearly correlated with FAPAR, whereas the other indices deviated slightly from a linear correlation.

When the red-edge reflectance was used, the ratio indices (e.g., mSR2 and SR705) had a stronger correlation with crop FAPAR than the normalized difference indices (e.g., ND705). Sensitivity analysis showed that mSR2 had the strongest linear correlation with FAPAR for the two crops across the growing season. Further analysis indicated that indices using the red-edge reflectance might be useful for FAPAR retrieval. Indices using the red-edge reflectance are independent of crop types. This suggests the potential for high resolution and high quality mapping of FPAR for precision farming using Sentinel-2 data.

The same approach was also applied to investigate the relationship between above ground biomass and various vegetation indices.

Figure 1   Relationships between FAPAR and Vegetation Indices with a Linear Coefficient of Determination greater than 0.83

Annual Cropland Mapping

Together with other four JECAM sites, we applied five different existing methodologies over five JECAM sites using same dataset (7-day 250m resolution MODIS NDVI time series). Confusion matrices and derived accuracy indicators were produced with and without equalizing class proportions of validation samples and correcting for the spatial resolution bias. A decision tree was used as the general method from China JECAM site.

Figure 2   Features Extracted from Smoothed Time Series NDVI Data

The MODIS time-series were not directly inputs for the classification as it was reported that when a high number of input data are used for classification, some make negligible or even negative contributions in terms of classification accuracy (Zhao et al., 2008; Zhang et al., 2012). In order to avoid such issues and to reduce the classification computing time, four temporal features were extracted from smoothed Normalized Difference Vegetation Index (NDVI) temporal profiles: the maximum vegetation index values observed at the date of the peak, the average vegetation index during the growing season as well as the green-up ratio and withering ratio.  Cropping intensity derived from time-series NDVI data is also considered to identify cropland and non-cropland. For pixels with two growing seasons, four temporal features were only extracted from the first growing season. The smoothing was achieved by applying a Savitzky-Golay filter (Savitzky and Golay,1964;Tsai and Philpot,1998). Based on the extracted parameters and the training samples, a decision tree was generated using the Classification and Regression Tree (CART) algorithm and applied to the whole study area to produce a land cover map.

Figure 3   Cropland Agreement Map using 5 Different Methods


Results using five different methods were overlapped to evaluate the agreement of the five cropland maps. Overall accuracy (OA) using five different methods over China JECAM site all exceeded 0.9. OA using the decision tree method over five different JECAM sites also presented high accuracy (higher than 0.9) except for the Sao Paulo, Brazil site.

©2015 Joint Experiment for Crop Assessment and Monitoring