Bayesian MODIS NDVI back-prediction by intersensor calibration with AVHRR

•Bayesian modeling of variation in inter-sensor relation from snow and land cover•Two year out-sample validation of the method on four representative U.S. states•A validated and consistent 16-day 8 km 30 year NDVI record

This article describes a Bayesian approach for ensuring the continuity of remote sensing records (1982–2014) by back-prediction of Moderate Resolution Imaging Spectroradiometer (MODIS) using historical Advanced Very High Resolution Radiometer (AVHRR) data. First, a historical 8 km snow data was generated using MODIS snow Quality Assurance field and North American Regional Reanalysis snow depth and air temperature outputs. Second, the relationships between coarsened 8 km MODIS and AVHRR Normalized Difference Vegetation Index (NDVI) were modeled pixel by pixel using spatial autocorrelation. Then the NDVI was back-predicted to 1982 from historical AVHRR observations and snow data. The back-predicted NDVI was validated against those data held-out from MODIS in 2001 and 2002 from four distinct biogeographic state regions (California, Kansas, Minnesota and Mississippi) in the continental United States. The results show consistency of the derived NDVI from two sensors, with a Root Mean Square Error within 0.05 over most land cover classes. This validation results suggest potential application of this approach for generating consistent long term multi-sensor NDVI data records for ecology and global climate change studies.