Leaf Area Index and Fraction of Absorbed Photosynthetically Active Radiation Thematic Climate Data Record from AVHRR

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Wednesday, 7 January 2015
Martin claverie, University of Maryland, College Park, DC; and E. Vermote and C. O. Justice

Handout (3.8 MB)

The continuous and long term global monitoring of the terrestrial biosphere requires the development of consistent long-term Climate Data Records (CDR) of Biophysical Variables (BV). In spring 2014, the AVHRR surface reflectance CDR (Land-CDR) was released by the NOAA National Climatic Data Center CDR project. It offers a unique data records to derive daily global BV maps at the Climate Modeling Grid resolution (0.05 degree) for a period of nearly 32 years. Two Thematic CDR (TCDR) corresponding to two key BV were developed from that record: Leaf Area Index (LAI) and Fraction of Absorbed Photosynthetically Active Radiation (FAPAR) and will be released in early winter 2014. In this paper, we described the algorithm used to produce the LAI/FAPAR TCDR. It is based on Artificial Neural Networks (ANN) trained with the official MODIS LAI/FAPAR products and AVHRR surface reflectance CDR. Ten ANN were built over five land cover classes (including most common biomes) and for each of the two BV. In addition, the evaluation of the uncertainties was performed, as it is required to achieve CDR level of quality. The validation procedure used in situ measurements from the DIRECT network (113 sites globally distributed and 181 measurements for a period spanning from 2001 to 2004). The results show overall uncertainties of 0.98 and 0.15 for LAI and FAPAR, respectively. Initially designed for cropland (where uncertainty on LAI is 0.65), the algorithm shows higher uncertainty over evergreen broadleaf forest (1.31 for LAI). Finally, sensor to sensor consistency was evaluated. BV derived from AVHRR sensors onboard of NOAA-16 and NOAA-18 (overlapping period of 18 months in 2005-2006) are highly correlated (r2>0.9) and show almost no systematic error (relative biases < 1%).