Publications

Pouliot, D; Latifovic, R (2018). Reconstruction of Landsat time series in the presence of irregular and sparse observations: Development and assessment in north-eastern Alberta, Canada. REMOTE SENSING OF ENVIRONMENT, 204, 979-996.

Abstract
Time series analysis of Landsat is limited by sparse and irregular sampling of clear-sky observations due to acquisition limitations, clouds, shadows, atmosphere, and sensor artifacts. Many remote sensing applications utilizing coarse spatial resolution time series methods are not suitable for Landsat due to observation sparsity. In this research we develop an imputation based approach to constrain the harmonic modeling method of Zhu et al. (2012, 2015) and Zhu and Woodcock (2014b) to reconstruct Landsat time series at a regular temporal interval. The approach was assessed for a boreal forest region in central Canada for different sparsity conditions. The imputed Landsat estimates for a specific pixel were predicted from climate or AVHRR data. These estimates were given a small weight relative to available Landsat observations in fitting the final harmonic model essentially constraining it to a more plausible range. In addition we implemented the model in a piecewise manner to handle non-linear temporal drift related to factors such as climate change, drought, or the allometric nature of vegetation regrowth. Results show that the inclusion of imputed estimates improved model predictions in the presence of observation sparsity. Where there were less than 3 observations within +/- 20 days the imputation approach performed better, with a reduction in average reflectance error of 0.001 to 2.5. Error assessment with hold out observations, comparison to MODIS time series, and example predicted images are presented.

DOI:
10.1016/j.rse.2017.07.036

ISSN:
0034-4257