A spatially adaptive filter for error reduction in satellite-based change detection algorithms

Abstract:

The last decade has seen a tremendous increase in use of satellite data for earth’s land surface monitoring due to increased data availability and decreased computational cost. There are now various operational programs that assess land cover over time, such as the Land Change Monitoring, Assessment, and Projection (LCMAP), National Land Cover Database (NLCD) and LANDFIRE (LF) program. The LF program has repeatedly and consistently produced spatially explicit vegetation change detection maps (i.e., the LF Disturbance products) over the United States for years 1999-2016. LF Disturbance products are produced using remote sensing-based change detection algorithms applied to each pixel (30 m) in Landsat data based annual composites. Enhancements to the current disturbance mapping procedures are needed to reduce errors of omission and commission. Omission errors may be reduced by making the algorithms more sensitive. However, making the algorithms more sensitive can increase errors of commission. Spatial contextual filters that compare undisturbed neighboring regions with disturbed regions to assign change probabilities have been shown to be effective in reducing commission errors especially in mapping fires and its effects over vegetation. In this work we present the results of application of a spatially adaptive filter for error reduction (SAFER) post change detection. SAFER uses established contextual tests and procedures to quantify per pixel change probabilities. We discuss results of application of SAFER over representative regions within conterminous United States in LF change detection schema.