Processor: Clip

Registry key: clip | Priority: 65 | Category: Spatial Processing

Subset climate data to specific geographic regions, points, or boundaries. Extract data for counties, watersheds, weather stations, or custom lat/lon coordinates with automatic nearest-gridcell location and coordinate system handling.

Algorithm

flowchart TD
    A["Input value\n(str / list / tuple / dict)"] --> B{Parse type}
    B -->|"Named string"| C["Load boundary or\nstation geometry"]
    B -->|"Point tuple"| D["Find closest\ngrid cell(s)"]
    B -->|"Bounding box"| E["Build bounding\nbox geometry"]
    B -->|"File path"| F["Load custom\nshapefile"]
    C & E & F --> G["Clip with rioxarray"]
    D --> H["Select nearest\ngrid point(s)"]
    G & H --> I{persist=True?}
    I -->|Yes| J["compute()\nDask graph"]
    I -->|No| K["Return lazy\nxarray result"]
    J --> K

Clip runs in two phases: first it parses self.value into a geometry (or routes to a point-based path), then it dispatches over the input data type and calls the appropriate clipper. See Geometry Loading and Result Dispatch in the Implementation Details section below.

Input Modes

Mode 1: Named Boundaries

Clip using predefined administrative or utility boundaries from the Cal-Adapt boundary catalog.

Examples:

"Los Angeles"              # County name
"San Francisco Bay"        # Watershed
"CA"                       # State-wide
"CA_IOU"                   # Utility (IOU = Investor-Owned Utility)

Mode 2: Weather Stations

Clip to specific weather station locations from the HadISD station network.

Examples:

"KSAC"                     # Sacramento International Airport
"KSFO"                     # San Francisco International Airport  
"KLAX"                     # Los Angeles International Airport

Mode 3: Single Point (Closest Valid Gridcell)

Extract data for a single geographic point. The processor first selects the geographically closest gridcell; if that cell is all-NaN (common in WRF data near coastlines/mountains/masked regions) it falls back to a 3×3 index-space neighborhood around the nearest cell and returns the mean of the valid (non-NaN) cells in that neighborhood.

Algorithm (_clip_data_to_point, line 665):

Call get_closest_gridcell(dataset, lat, lon) and check if it has valid data on a sample slice.
If valid, return that cell.
If NaN, find the nearest grid index ((idx1, idx2)) along the spatial dims (x/y for WRF, lat/lon for LOCA2). For projected grids, lat/lon are first transformed to the dataset's CRS via pyproj.Transformer.
Iterate the 3×3 neighborhood (di, dj) ∈ {-1, 0, 1}², skipping out-of-bounds and all-NaN cells.
If any valid neighbors exist, xr.concat(neighbors, dim="nearest_cell").mean(dim="nearest_cell") and reassign mean lat/lon coords; otherwise return None.
Fallback paths handle averaging failures by returning the center cell directly.

The earlier docs described an expanding-radius search (0.01° → 0.05° → ... → 0.5°). That was never the implementation; the current behavior is the index-space 3×3 mean above.

Example:

(37.7749, -122.4194)       # San Francisco: (lat, lon)

Mode 4: Bounding Box

Clip data to a rectangular geographic region specified by latitude and longitude ranges.

Example:

((36.0, 39.0), (-122.0, -118.0))   # ((lat_min, lat_max), (lon_min, lon_max))

Mode 5: Custom File

Clip using geometry from shapefile or GeoJSON file.

Example:

"/path/to/custom_region.shp"       # Shapefile with geometry

Multi-Input Handling

Multiple Boundaries

When providing a list of boundary names, they are combined using union (OR logic):

["Alameda", "Contra Costa", "Santa Clara"]  # All three counties combined

With separation:

{"boundaries": ["Alameda", "Contra Costa"], "separated": True}
# Returns: Dict with separate Dataset for each county

Multiple Points

When providing multiple point coordinates:

Each point gets independent smart-gridcell search
Results include closest_cell dimension with length = number of points
Automatic duplicate filtering if multiple points map to same gridcell

Example:

[(37.7749, -122.4194), (34.0522, -118.2437), (32.7157, -117.1611)]
# Returns: Dataset with closest_cell dimension = 3 (SF, LA, San Diego)

Spatial Processing Details

Coordinate System Handling

Data CRS detection (_clip_data_with_geom, line 595): if data.rio.crs is unset, the processor detects WRF data by the presence of a Lambert_Conformal coordinate and writes the CRS from spatial_ref or CF-convention attributes; otherwise it assumes EPSG:4326 (LOCA2 lat/lon).
Boundary CRS: boundaries are assumed to be EPSG:4326 (a warning is logged if no CRS is set).
Reprojection direction: when data.rio.crs != gdf.crs, the GeoDataFrame is reprojected to the data's CRS — the data stays in its native projection (i.e., WRF output remains in Lambert Conformal).

Masking Strategy

Geometry-based clipping uses rioxarray:

data.rio.clip(gdf.geometry.apply(mapping), gdf.crs, drop=True, all_touched=True)

all_touched=True includes cells that any part of the geometry touches.
drop=True trims the bounding box to the clipped extent.

For multi-point clipping, _clip_data_to_points_as_mask either applies a points-mask in place or extracts along a new points dimension when extract_points=True (set by {"points": [...], "separated": True}).

Persisting to Memory

When persist=True (constructor arg or persist key in dict input), the processor calls .compute() on the result before returning. This collapses the Dask task graph and is recommended for large multi-point clipping followed by 1-in-X analysis or other graph-heavy operations.

Boundary Catalog Access

Boundaries loaded lazily via _get_boundary_geometry (line 1443) / _get_multi_boundary_geometry (line 1683).
Sourced from S3 intake-esm catalog and cached for the process lifetime.

Parameters

Parameter	Type	Required	Default	Description
`value`	str / list / tuple / dict	✓	—	Geometry specification (modes 1–5 above). For dict input, must contain `boundaries` or `points`.
`separated`	bool (in dict input)		False	For boundaries: keep each as its own dataset entry. For points: extract along a new `points` dimension.
`persist`	bool (in dict input or constructor)		False	Call `.compute()` after clipping to collapse the Dask task graph. Recommended for large multi-point workflows.

Code References

Method	Lines	Purpose
`__init__`	186–291	Parse `value`, set mode flags (`is_single_point`, `is_multi_point`, `separated`, `extract_points`, `persist`)
`execute`	294–477	Two-phase: parse value → geom (line 303), then dispatch over result type (line 386); optional `.compute()` if `persist`
`update_context`	479–535	Record clipped-region metadata under `new_attrs["clip"]`
`set_data_accessor`	537–539	Receive `DataCatalog` reference for boundary/station lookups
`_get_station_coordinates`	541–568	HadISD station code → (lat, lon, metadata)
`_convert_stations_to_points`	570–593	Multi-station list → point_list + metadata list
`_clip_data_with_geom`	595–663	CRS detection + `rio.clip(all_touched=True, drop=True)`
`_clip_data_to_point`	665–878	Closest cell, fallback to 3×3 index-space neighborhood mean
`_clip_data_to_multiple_points`	880–961	Vectorized multi-point selection
`_clip_data_to_multiple_points_fallback`	963–1059	Fallback path used when vectorized lookup fails
`_clip_data_to_points_as_mask`	1061–1346	Mask-based multi-point clip; supports `extract_points` along new `points` dim
`_clip_data_separated`	1348–1441	Per-boundary clipping returning dict/list keyed by boundary
`_get_boundary_geometry`	1443–1503	Single boundary key → GeoDataFrame
`_get_multi_boundary_geometry`	1683–1749	Multi-boundary union via `_combine_geometries`
`_combine_geometries`	1751–	Geometry union helper

Examples

Single County

from climakitae.new_core.user_interface import ClimateData

data = (ClimateData()
    .catalog("cadcat")
    .activity_id("WRF")
    .variable("t2max")
    .table_id("day")
    .grid_label("d03")
    .processes({
        "clip": "Alameda"
    })
    .get())

Multiple Counties (Separated)

# Get each county in separate dataset
data = (ClimateData()
    .catalog("cadcat")
    .activity_id("WRF")
    .variable("pr")
    .table_id("mon")
    .grid_label("d02")
    .processes({
        "clip": {
            "boundaries": ["Alameda", "Contra Costa", "Santa Clara"],
            "separated": True
        }
    })
    .get())

# data is dict: {"Alameda": ds1, "Contra Costa": ds2, "Santa Clara": ds3}

Single Lat/Lon Point

# Closest grid cell to San Francisco
data = (ClimateData()
    .catalog("cadcat")
    .activity_id("WRF")
    .variable("t2max")
    .table_id("day")
    .grid_label("d03")
    .processes({
        "clip": (37.7749, -122.4194)
    })
    .get())

# Scalar lat/lon coordinates (size 1)

Multiple Points (Separated)

# Time series for 3 cities
locations = [
    (34.05, -118.25),    # Los Angeles
    (37.77, -122.42),    # San Francisco
    (32.72, -117.16)     # San Diego
]

data = (ClimateData()
    .catalog("cadcat")
    .activity_id("WRF")
    .variable("t2max")
    .table_id("day")
    .grid_label("d03")
    .processes({
        "clip": {
            "boundaries": locations,
            "separated": True,
            "location_based_naming": True
        }
    })
    .get())

# data is dict with lat/lon in keys

Bounding Box

# Bay Area region (rough bbox)
data = (ClimateData()
    .catalog("cadcat")
    .activity_id("WRF")
    .variable("pr")
    .table_id("mon")
    .grid_label("d03")
    .processes({
        "clip": ((37.5, 38.5), (-123.0, -121.5))
    })
    .get())

Weather Station

# Sacramento airport observations reference point
data = (ClimateData()
    .catalog("cadcat")
    .activity_id("WRF")
    .variable("t2max")
    .table_id("day")
    .grid_label("d03")
    .processes({
        "clip": "KSAC"
    })
    .get())

Chained: Clip → Warming Level → Export

data = (ClimateData()
    .catalog("cadcat")
    .activity_id("WRF")
    .experiment_id("ssp245")
    .variable("t2max")
    .table_id("day")
    .grid_label("d03")
    .processes({
        "clip": "Los Angeles",
        "warming_level": {"warming_levels": [1.5, 2.0, 3.0]},
        "export": {
            "filename": "la_warming",
            "file_format": "NetCDF"
        }
    })
    .get())

Implementation Details

Geometry Loading

Clip resolves self.value to a geometry (or to a point-mode flag) in this order:

String → station id (is_station_identifier, line 305) → file path (os.path.exists, line 318) → boundary key (_get_boundary_geometry, line 320).
List → all-station list (line 324) → lat/lon tuples (is_multi_point, line 340) → separated boundaries (line 343) → union of boundaries (_get_multi_boundary_geometry, line 345).
Tuple → single (lat, lon) (line 347) or bbox ((lat_min, lat_max), (lon_min, lon_max)) (line 355, builds shapely.box in EPSG:4326).

Result Dispatch

The second match (line 386) routes the input data:

dict: per-key clip; mode flag selects the clipper.
Dataset / DataArray: single clip via the matching mode’s clipper.
list / tuple: per-item clip, container type preserved; None results filtered out for point modes.

Persist (`persist=True`)

After clipping, if self.persist is true the processor calls .compute() on the result (per-value for dicts, on the whole object otherwise). This collapses very large Dask task graphs that arise from multi-point clipping followed by quantile or block operations and prevents OOMs in downstream steps.

Error Handling

Invalid station code, missing boundary key, or unsupported value type → ValueError.
All-NaN single point with no valid 3×3 neighbors → _clip_data_to_point returns None; the per-item paths filter these out.
Failed CRS detection on WRF data missing required CF attributes → ValueError with the missing key.

Common Patterns

County Loop

import climakitae

counties = ["Alameda", "Contra Costa", "Santa Clara", "San Mateo"]
data_by_county = {}

for county in counties:
    data_by_county[county] = (ClimateData()
        .catalog("cadcat")
        .activity_id("WRF")
        .variable("t2max")
        .table_id("day")
        .grid_label("d03")
        .processes({"clip": county})
        .get())

Urban Heat Island Study

# Urban and rural points for comparison
urban_point = (37.7749, -122.4194)      # San Francisco downtown
rural_point = (37.5, -122.0)            # Sierra foothills

data = (ClimateData()
    .catalog("cadcat")
    .activity_id("WRF")
    .variable("t2max")
    .table_id("day")
    .grid_label("d03")
    .processes({
        "clip": {
            "boundaries": [urban_point, rural_point],
            "separated": True
        }
    })
    .get())

# data["urban_point"] vs data["rural_point"] comparison

Processor: Clip

Algorithm

Input Modes

Mode 1: Named Boundaries

Mode 2: Weather Stations

Mode 3: Single Point (Closest Valid Gridcell)

Mode 4: Bounding Box

Mode 5: Custom File

Multi-Input Handling

Multiple Boundaries

Multiple Points

Spatial Processing Details

Coordinate System Handling

Masking Strategy

Persisting to Memory

Boundary Catalog Access

Parameters

Code References

Examples

Single County

Multiple Counties (Separated)

Single Lat/Lon Point

Multiple Points (Separated)

Bounding Box

Weather Station

Chained: Clip → Warming Level → Export

Implementation Details

Geometry Loading

Result Dispatch

Persist (persist=True)

Error Handling

Common Patterns

County Loop

Urban Heat Island Study

See Also

Persist (`persist=True`)