Hampel filter a stack over its time axis.

The classic despiking filter: for each time step, a centred window of k steps either side supplies a median and a MAD (median absolute deviation, scaled by 1.4826 to estimate a standard deviation); an observation further than t0 MADs from the window median is replaced by that median. Missing steps (NaN) are skipped by the window statistics and stay missing in the output: this filter removes spikes, it does not fill gaps (compose with fill_gaps() or kalman_smooth() for that).

Usage

hampel_smooth(x, k = 3L, t0 = 3, bands = NULL)

Arguments

x

A (t, y, x) LazyRaster, or a LazyDataset (each band filtered independently).

k

Window half-width in time steps (window size 2k + 1).

t0

Threshold in scaled MADs (default 3; 0 = rolling median).

bands

LazyDataset only: bands to filter (default: all).

Value

The filtered object, same class and grid as x.

Details

The window is positional (k slices, not calendar days), and shrinks at the series ends. t0 = 0 degenerates to a rolling median. With t0 > 0, note the MAD of a window whose valid values are mostly identical is 0, so any deviating centre is replaced; this is inherent to the Hampel construction.

Like every garry verb it is lazy (nothing computes until collect()), and it applies per band over a LazyDataset.

See also

kalman_smooth(), fill_gaps(), scan_over()