Skip to content

Station weights

Station weights decide how much each station contributes to the stack of a node. Qseek calculates a weight for every pair of station and node in the search volume from their distance, see stacking and migration. For every node and phase, the weights are normalized to a sum of one. Close stations get full weight: they record small events with the highest phase confidence and constrain the location best. Distant stations are tapered.

Station weights Full weight Taper
DistanceWeights (default) The 4 closest stations Gaussian, full width at half maximum twice the mean interstation distance
DensityTaperWeights The 4 closest stations As DistanceWeights; far away, each weight approaches the taper times the independent station count

station_weights takes one of them. With null, all stations get the same weight.

Station weights

Station weights of a node, with the defaults, for two examples of the playground. Left: Campi Flegrei, 18 stations within 12 km, node at 2 km depth below Solfatara. Right: Mount Spurr, Alaska, node at 4 km depth below sea level beneath the volcano; ten local stations within 32 km and ten regional stations from 79 km to 131 km. Dashed: DistanceWeights with the taper "nearest_neighbor". Bottom: the independent station count of each station.

Choose the station weights

  • DistanceWeights, the default, locate best on the examples of the playground and suit most local and regional networks.
  • DensityTaperWeights keep the full weight of the closest stations and count a distant cluster of stations as its independent stations. Try them for a network with a dense cluster next to sparse stations. On the playground they locate as consistently as DistanceWeights and find slightly more well-constrained detections where a cluster lies away from the seismicity.

The weights change from node to node, and they are normalized per node. A taper that falls off steeply between neighboring nodes therefore moves the locations as well: the stack prefers the nodes where the stations that recorded the event carry most of the weight. Keep the taper smooth. Over the distances where stations still record the events, the weights should decay as the phase confidence decays with distance. On the playground, a narrower taper and more or fewer stations with full weight both locate the detections less consistently than the defaults of DistanceWeights. The benchmark has the results.

Distance weights

The required_closest_stations closest stations of a node get full weight, 4 by default. More distant stations are tapered with a Gaussian function; distance_taper is its full width at half maximum, twice the mean interstation distance by default, "mean_interstation".

"nearest_neighbor" sets the taper to 6 times the median distance between neighboring station sites. On networks of similar station spacing it adapts the taper to the spacing: at Mount Spurr, where the regional stations widen the mean interstation distance, it moves the epicenters closer to the catalog. Do not use it for networks with a dense array: the spacing of the array sets the median.

DistanceWeights
{
  "weights": "DistanceWeights",
  "waterlevel": 0.0,
  "distance_taper": "mean_interstation",
  "required_closest_stations": 4
}

DistanceWeights pydantic-model

Bases: StationWeights

The closest stations get full weight, more distant stations a Gaussian taper.

Close stations constrain the location of an event best. The taper width is absolute, the same for all nodes. With the defaults, these are the station weights of earlier Qseek versions.

Fields:

required_closest_stations pydantic-field

required_closest_stations: PositiveInt = 4

Number of closest stations of a node that get full weight. Only more distant stations are tapered, so that the closest stations contribute equally and the most to the detection and localization.

waterlevel pydantic-field

waterlevel: float = 0.0

Minimum weight of distant stations, as a fraction of the full weight. With 0.0, distant stations do not contribute.

get_subclasses classmethod

get_subclasses() -> tuple[type[StationWeights], ...]

Get the subclasses of this class, also of its subclasses.

Returns:

Type Description
tuple[type[StationWeights], ...]

tuple[type[StationWeights], ...]: The subclasses of this class.

get_distances

get_distances(nodes: Sequence[Node]) -> ndarray

Get the distances between nodes and all stations.

Parameters:

Name Type Description Default
nodes Sequence[Node]

Nodes of the octree.

required

Returns:

Type Description
ndarray

Array of shape (n_nodes, n_stations) with distances in meters, float32.

fill_lut

fill_lut(nodes: Sequence[Node]) -> None

Calculate the node-station distances of nodes for the lookup table.

Parameters:

Name Type Description Default
nodes Sequence[Node]

Nodes of the octree.

required

get_interstation_distances

get_interstation_distances(station_indices: ndarray) -> ndarray

Get the distances between a set of stations.

Parameters:

Name Type Description Default
station_indices ndarray

Indices of the stations.

required

Returns:

Type Description
ndarray

Array of shape (n_stations, n_stations) in meters, NaN on the diagonal.

get_station_counts

get_station_counts(station_indices: ndarray) -> ndarray

Get the independent station counts of a set of stations.

The counts depend on the available stations, they are cached for each set of stations.

Parameters:

Name Type Description Default
station_indices ndarray

Indices of the available stations.

required

Returns:

Type Description
ndarray

Array of shape (n_stations,) with values between 0 and 1.

resolve_distance_taper

resolve_distance_taper(distance_taper: float | Literal['mean_interstation', 'nearest_neighbor'], stations: StationInventory) -> float

Resolve the width of a Gaussian distance taper for the stations.

Parameters:

Name Type Description Default
distance_taper float | Literal['mean_interstation', 'nearest_neighbor']

Full width at half maximum in meters, "mean_interstation" for twice the mean interstation distance or "nearest_neighbor" for NEAREST_NEIGHBOR_TAPER times the median distance between neighboring sites.

required
stations StationInventory

Stations of the search.

required

Returns:

Type Description
float

Full width at half maximum in meters, infinite with fewer than two station sites.

get_weights async

get_weights(nodes: Sequence[Node], stations: Sequence[Station]) -> ndarray

Get the weights of the stations for nodes.

Parameters:

Name Type Description Default
nodes Sequence[Node]

Nodes of the octree.

required
stations Sequence[Station]

Available stations.

required

Returns:

Type Description
ndarray

Array of shape (n_nodes, n_stations) with weights between 0 and 1.

Independent stations

Stations a few hundred meters apart record nearly the same waveforms and share the errors of the velocity model along their paths. DensityTaperWeights count how many independent stations each station is worth. An isolated station counts as one independent station, a station in a dense cluster as a fraction of one:

\[ c = 1 - \frac{k - k_\mathrm{min}}{k_\mathrm{max}} \]

with the density \(k\) of the neighboring stations, a Gaussian kernel as wide as the median distance between neighboring sites, and the lowest and highest density of the network, \(k_\mathrm{min}\) and \(k_\mathrm{max}\). The densest station keeps \(k_\mathrm{min} / k_\mathrm{max}\), so every station counts. Sensors closer than 50 m, e.g. a broadband and a strong-motion sensor of one site, share a site and together count as the site. The counts are relative: they change when stations have no data in a window.

On Campi Flegrei, the 18 INGV stations count as 11.1 independent stations: the six stations in the center of the caldera, 200 m to 500 m apart, count as 0.2 each. At Mount Spurr, the 20 stations count as 13.8 independent stations, the ten local stations as 4.6. Qseek logs the count when the search starts.

Density taper weights

DensityTaperWeights start from the weight \(g\) of DistanceWeights and the independent station count \(c\) of each station:

\[ w = g \left( c^\beta + (1 - c^\beta)\, g \right) \]

The closest stations keep their full weight (\(g = 1\)), also within a cluster. Seen from a distance, the stations of a cluster record an event from nearly the same direction; with \(g\) falling off, their weight approaches \(g\, c^\beta\), so the cluster weighs as much as its independent stations. station_density is the exponent \(\beta\), 1 by default; with 0 these are the DistanceWeights. The weights change between nodes about as smoothly as the DistanceWeights.

On Campi Flegrei, the cluster of six stations lies above the seismicity and stays on the plateau for most detections: DensityTaperWeights locate as consistently as DistanceWeights. At Norcia, the cluster of 12 stations lies at about 28 km from the seismicity, beyond the plateau, where DistanceWeights give it nearly full weight; DensityTaperWeights count it as its independent stations and find more well-constrained detections. The benchmark has the numbers.

DensityTaperWeights
{
  "weights": "DensityTaperWeights",
  "waterlevel": 0.0,
  "distance_taper": "mean_interstation",
  "required_closest_stations": 4,
  "station_density": 1.0
}

DensityTaperWeights pydantic-model

Bases: DistanceWeights

Distance weights that count distant clusters of stations as fewer stations.

The closest stations get full weight and more distant stations the Gaussian taper of DistanceWeights. Seen from a distance, the stations of a dense cluster record the event in nearly the same direction: in the taper, each station's weight approaches its independent station count. The closest stations keep their full weight, also within a cluster.

Fields:

station_density pydantic-field

station_density: float = 1.0

Exponent of the independent station count in the taper. A station in a dense cluster counts as a fraction of an independent station; with 1.0 a distant cluster weighs as much as its independent stations, with 0.0 these are the DistanceWeights.

waterlevel pydantic-field

waterlevel: float = 0.0

Minimum weight of distant stations, as a fraction of the full weight. With 0.0, distant stations do not contribute.

required_closest_stations pydantic-field

required_closest_stations: PositiveInt = 4

Number of closest stations of a node that get full weight. Only more distant stations are tapered, so that the closest stations contribute equally and the most to the detection and localization.

get_subclasses classmethod

get_subclasses() -> tuple[type[StationWeights], ...]

Get the subclasses of this class, also of its subclasses.

Returns:

Type Description
tuple[type[StationWeights], ...]

tuple[type[StationWeights], ...]: The subclasses of this class.

get_distances

get_distances(nodes: Sequence[Node]) -> ndarray

Get the distances between nodes and all stations.

Parameters:

Name Type Description Default
nodes Sequence[Node]

Nodes of the octree.

required

Returns:

Type Description
ndarray

Array of shape (n_nodes, n_stations) with distances in meters, float32.

fill_lut

fill_lut(nodes: Sequence[Node]) -> None

Calculate the node-station distances of nodes for the lookup table.

Parameters:

Name Type Description Default
nodes Sequence[Node]

Nodes of the octree.

required

get_interstation_distances

get_interstation_distances(station_indices: ndarray) -> ndarray

Get the distances between a set of stations.

Parameters:

Name Type Description Default
station_indices ndarray

Indices of the stations.

required

Returns:

Type Description
ndarray

Array of shape (n_stations, n_stations) in meters, NaN on the diagonal.

get_station_counts

get_station_counts(station_indices: ndarray) -> ndarray

Get the independent station counts of a set of stations.

The counts depend on the available stations, they are cached for each set of stations.

Parameters:

Name Type Description Default
station_indices ndarray

Indices of the available stations.

required

Returns:

Type Description
ndarray

Array of shape (n_stations,) with values between 0 and 1.

resolve_distance_taper

resolve_distance_taper(distance_taper: float | Literal['mean_interstation', 'nearest_neighbor'], stations: StationInventory) -> float

Resolve the width of a Gaussian distance taper for the stations.

Parameters:

Name Type Description Default
distance_taper float | Literal['mean_interstation', 'nearest_neighbor']

Full width at half maximum in meters, "mean_interstation" for twice the mean interstation distance or "nearest_neighbor" for NEAREST_NEIGHBOR_TAPER times the median distance between neighboring sites.

required
stations StationInventory

Stations of the search.

required

Returns:

Type Description
float

Full width at half maximum in meters, infinite with fewer than two station sites.

get_weights async

get_weights(nodes: Sequence[Node], stations: Sequence[Station]) -> ndarray

Get the weights of the stations for nodes.

Parameters:

Name Type Description Default
nodes Sequence[Node]

Nodes of the octree.

required
stations Sequence[Station]

Available stations.

required

Returns:

Type Description
ndarray

Array of shape (n_nodes, n_stations) with weights between 0 and 1.