Reconstruction of Satellite Time Series With a Dynamic Smoother

Abstract

Time series reconstruction methods are widely used to generate smooth and gap-free time series using imagery acquired at coarse spatial resolution and high frequency return intervals. However, as interest has grown in leveraging the nearly 40-a record of Landsat to study long-term changes in terrestrial ecosystems at 30-m spatial resolution, new methods are required to reconstruct time series of Landsat imagery, which have lower temporal density than coarse resolution sensors such as AVHRR or MODIS. To address this need, we introduce a dynamic temporal smoothing (DTS) method that reconstructs sparse and noisy signals into dense time series at regular intervals. The DTS is a weighted smoother with parameters that adjust dynamically to variation in time series and can be applied to both dense and sparse time series measurements. Because the DTS smoother we describe is specifically designed to reconstruct high-quality time series of optical imagery, it has utility for applications focused on land cover and vegetation remote sensing over long time periods at moderate spatial resolution. We present the DTS algorithm that we have implemented and illustrate the ability of the DTS to reconstruct time series of Landsat imagery across multiple sensors (TM, ETM+, and OLI). To demonstrate the effectiveness of the DTS algorithm we apply it and evaluate results across a diverse range of land cover and vegetation types in the South American Southern Cone region.