Patch.hampel_filter replaces samples whose deviation from the local median exceeds a threshold in median absolute deviation units with the local median.
import numpy as np
import dascore as dc
plot_kwargs = dict(scale=(-200, 200), scale_type="absolute")
patch = dc.get_example_patch("example_event_2")
data = patch.data.copy()
data[50, 100] = 600
data[100, 150] = -1000
data[150, 200] = 1200.0
data[:, 75] = 1500.0
data[75, :] = -1400
spike_patch = patch.update(data=data)
ax = spike_patch.viz.waterfall(**plot_kwargs)
ax.set_title("Data with Artificial Spikes");
Lower thresholds filter more aggressively:
filtered_time = spike_patch.hampel_filter(time=0.02, threshold=3.5)
ax = filtered_time.viz.waterfall(**plot_kwargs)
ax.set_title("Filtered along Time (threshold=3.5)");
Large exact windows (approximate=False) are expensive because cost grows with window area. The default approximation is nearly insensitive to window size.
Include distance to remove channel-wide spikes:
filtered_2d = spike_patch.hampel_filter(time=0.02, distance=5.0, threshold=3.5)
ax = filtered_2d.viz.waterfall(**plot_kwargs)
ax.set_title("Filtered along Time and Distance");
Start with thresholds of 3.5–5 and odd windows of 5, 7, or 9 samples per dimension (samples=True), then inspect the result. Use exact medians only when the approximation is insufficient.