Derivation of 16-day time-series NDVI data for environmental studies using a data assimilation approach

Abstract

Many ecological and environmental applications require time-series data, which are collected with spatial extent from local, regional or continental levels and granularity ranging from fine to coarse spatial resolution. These types of data can be too difficult to collect using typical field surveys, but they can be derived using remote-sensing images and image-processing technologies. The common time-series normalized difference vegetation index (NDVI) data are AVHRR-derived at spatial resolutions of 1 or 8 km, MODIS-derived at 250 m, 500 m, 1 km or 8 km resolutions, and SPOT-VGT imagery at 1 km resolution. Landsat imagery-derived time-series NDVI data are often unavailable for many areas due to the difficulty of acquiring cloud-free images given the temporally infrequent coverage of this sensor. For this study, we used a closest-spectral-fit concept as the basis for a data mining model and developed a data assimilation model in order to derive 16-day time-series NDVI data. In a multidimensional spectral space, if pixel i has the closest reflectance values to pixel j, pixel i is called the closest-spectral-fit of pixel j, and pixels i and j are called closest-spectral-fit pixels. Fifteen Landsat TM images covering northern Michigan were acquired from 22 March 2010 to 1 November 2010, in a 16-day cycle. A cloud-free image is used as the reference image to predict NDVI images for the other 14 dates. The forecasted NDVI data explain 80% variation in the observed NDVI. TM band 4 is forecasted at the same performance level as NDVI, and forecasts of bands 3 and 2 are relatively highly correlated with the forecasts of observed bands. The closest-spectral-fit data assimilation method has the capability to produce historical NDVI data at finer spatial resolution as far back as 1972, which broaden and enhance potential applications to the modeling of environmental and ecological patterns and processes.