Source code for imod.evaluate.budget

import dask
import dask.array
import numba
import numpy as np
import scipy.ndimage
import xarray as xr

DIM_Z = 0
DIM_Y = 1
DIM_X = 2
LOWER = 3
UPPER = 4
FRONT = 5
BACK = 6
RIGHT = 7
LEFT = 8


def _outer_edge(da):
    faces = np.full(da.shape, np.nan)
    unique_ids = np.unique(da.values)
    unique_ids = unique_ids[~np.isnan(unique_ids)]
    # Number the faces by their id
    for unique_id in unique_ids:
        data = (da == unique_id).values
        from_edge = scipy.ndimage.morphology.binary_erosion(data)
        is_edge = (data == 1) & (from_edge == 0)
        faces[is_edge] = unique_id
    return xr.DataArray(faces, da.coords, da.dims)


@numba.njit
def _face_indices(face, budgetzone):
    nface = int(np.nansum(face))
    shape = (nface, 9)
    indices = np.zeros(shape, dtype=np.int32)
    if nface == 0:
        return indices
    # Loop over cells
    nlay, nrow, ncol = budgetzone.shape
    count = 0
    for k in range(nlay):
        for i in range(nrow):
            for j in range(ncol):
                face_value = face[k, i, j]
                if ~np.isnan(face_value):
                    # Store indices
                    indices[count, DIM_Z] = k
                    indices[count, DIM_Y] = i
                    indices[count, DIM_X] = j
                    # Default value: part of domain for edges
                    lower = front = right = upper = back = left = face_value
                    if k > 0:
                        upper = budgetzone[k - 1, i, j]
                    if k < (nlay - 1):
                        lower = budgetzone[k + 1, i, j]
                    if i > 0:
                        back = budgetzone[k, i - 1, j]
                    if i < (nrow - 1):
                        front = budgetzone[k, i + 1, j]
                    if j > 0:
                        left = budgetzone[k, i, j - 1]
                    if j < (ncol - 1):
                        right = budgetzone[k, i, j + 1]

                    # Test if cell is a control surface cell for the direction
                    if lower != face_value:
                        indices[count, LOWER] = 1
                    if upper != face_value:
                        indices[count, UPPER] = 1
                    if front != face_value:
                        indices[count, FRONT] = 1
                    if back != face_value:
                        indices[count, BACK] = 1
                    if right != face_value:
                        indices[count, RIGHT] = 1
                    if left != face_value:
                        indices[count, LEFT] = 1

                    # Incrementer counter
                    count += 1
    return indices


@numba.njit
def _collect_flowfront(indices, flow):
    result = np.zeros(flow.shape)
    nface = indices.shape[0]
    for count in range(nface):
        k = indices[count, DIM_Z]
        i = indices[count, DIM_Y]
        j = indices[count, DIM_X]
        if indices[count, FRONT]:
            result[k, i, j] += flow[k, i, j]
        if indices[count, BACK]:
            result[k, i, j] -= flow[k, i - 1, j]
    return result


@numba.njit
def _collect_flowlower(indices, flow):
    result = np.zeros(flow.shape)
    nface = indices.shape[0]
    for count in range(nface):
        k = indices[count, DIM_Z]
        i = indices[count, DIM_Y]
        j = indices[count, DIM_X]
        if indices[count, LOWER]:
            result[k, i, j] += flow[k, i, j]
        if indices[count, UPPER]:
            result[k, i, j] -= flow[k - 1, i, j]
    return result


@numba.njit
def _collect_flowright(indices, flow):
    result = np.zeros(flow.shape)
    nface = indices.shape[0]
    for count in range(nface):
        k = indices[count, DIM_Z]
        i = indices[count, DIM_Y]
        j = indices[count, DIM_X]
        if indices[count, RIGHT]:
            result[k, i, j] += flow[k, i, j]
        if indices[count, LEFT]:
            result[k, i, j] -= flow[k, i, j - 1]
    return result


def delayed_collect(indices, front, lower, right):
    indices = indices.compute()
    result_front = dask.delayed(_collect_flowfront, nout=1)(indices, front.values)
    result_lower = dask.delayed(_collect_flowlower, nout=1)(indices, lower.values)
    result_right = dask.delayed(_collect_flowright, nout=1)(indices, right.values)
    dask_front = dask.array.from_delayed(result_front, front.shape, dtype=np.float)
    dask_lower = dask.array.from_delayed(result_lower, lower.shape, dtype=np.float)
    dask_right = dask.array.from_delayed(result_right, right.shape, dtype=np.float)
    return dask_front, dask_lower, dask_right


[docs]def facebudget(budgetzone, front=None, lower=None, right=None, netflow=True): """ Computes net face flow into a control volume, as defined by ``budgetzone``. Returns a three dimensional DataArray with in- and outgoing flows for every cell that is located on the edge of the control volume, thereby calculating the flow through the control surface of the control volume. Front, lower, and right arguments refer to iMOD face flow budgets, in cubic meters per day. In terms of flow direction these are defined as: * ``front``: positive with ``y`` (negative with row index) * ``lower``: positive with ``layer`` (positive with layer index) * ``right``: negative with ``x`` (negative with column index) Only a single face flow has to be defined, for example, if only vertical fluxes (``lower``) are to be considered. Note that you generally should not define a budgetzone that is only one cell wide. In that case, flow will both enter and leave the cell, resulting in a net flow of zero (given there are no boundary conditions present). The output of this function defines ingoing flow as positive, and outgoing flow as negative. The output is a 3D array with net flows for every control surface cell. You can select specific parts of the control surface, for example only the east-facing side of the control volume. Please refer to the examples. Parameters ---------- budgetzone: xr.DataAray of floats Array defining zones. Non-zones should be with a ``np.nan`` value. Dimensions must be exactly ``("layer", "y", "x")``. front: xr.DataArray of floats, optional Dimensions must be exactly ``("layer", "y", "x")`` or ``("time", "layer", "y", "x")``. lower: xr.DataArray of floats, optional Dimensions must be exactly ``("layer", "y", "x")`` or ``("time", "layer", "y", "x")``. right: xr.DataArray of floats, optional Dimensions must be exactly ``("layer", "y", "x")`` or ``("time", "layer", "y", "x")``. netflow : bool, optional Whether to split flows by direction (front, lower, right). True: sum all flows. False: return individual directions. Returns ------- facebudget_front, facebudget_lower, face_budget_right : xr.DataArray of floats Only returned if `netflow` is False. facebudget_net : xr.DataArray of floats Only returned if `netflow` is True. Examples -------- Load the face flows, and select the last time using index selection: >>> import imod >>> lower = imod.idf.open("bdgflf*.idf").isel(time=-1) >>> right = imod.idf.open("bdgfrf*.idf").isel(time=-1) >>> front = imod.idf.open("bdgfff*.idf").isel(time=-1) Define the zone of interest, e.g. via rasterizing a shapefile: >>> import geopandas as gpd >>> gdf = gpd.read_file("zone_of_interest.shp") >>> zone2D = imod.prepare.rasterize(gdf, like=lower.isel(layer=0)) Broadcast it to three dimensions: >>> zone = xr.full_like(flow, zone2D) Compute net flow through the (control) surface of the budget zone: >>> flow = imod.evaluate.facebudget( >>> budgetzone=zone, front=front, lower=lower, right=right >>> ) Or evaluate only horizontal fluxes: >>> flow = imod.evaluate.facebudget( >>> budgetzone=zone, front=front, right=right >>> ) Extract the net flow, only on the right side of the zone, for example as defined by x > 10000: >>> netflow_right = flow.where(flow["x"] > 10_000.0).sum() Extract the net flow for only the first four layers: >>> netflow_layers = netflow_right.sel(layer=slice(1, 4)).sum() Extract the net flow to the right of an arbitrary diagonal in ``x`` and ``y`` is simple using the equation for a straight line: >>> netflow_right_of_diagonal = flow.where( >>> flow["y"] < (line_slope * flow["x"] + line_intercept) >>> ) There are many ways to extract flows for a certain part of the zone of interest. The most flexible one with regards to the ``x`` and ``y`` dimensions is by drawing a vector file, rasterizing it, and using it to select with ``xarray.where()``. To get the flows per direction, pass ``netflow=False``. >>> flowfront, flowlower, flowright = imod.evaluate.facebudget( >>> budgetzone=zone, front=front, lower=lower, right=right, netflow=False >>> ) """ # Error handling if front is None and lower is None and right is None: raise ValueError("Atleast a single flow budget DataArray must be given") if tuple(budgetzone.dims) != ("layer", "y", "x"): raise ValueError('Dimensions of budgetzone must be exactly ("layer", "y", "x")') shape = budgetzone.shape for da, daname in zip((front, lower, right), ("front", "lower", "right")): if da is not None: dims = da.dims coords = da.coords if "time" in dims: da_shape = da.shape[1:] else: da_shape = da.shape if da_shape != shape: raise ValueError(f"Shape of {daname} does not match budgetzone") # Determine control surface face = _outer_edge(budgetzone) # Convert indices array to dask array, otherwise garbage collection gets # rid of the array and we get a segfault. indices = dask.array.from_array(_face_indices(face.values, budgetzone.values)) # Make sure it returns NaNs if no zones are defined. print(indices.size) print(indices) if indices.size > 0: # Create dummy arrays for skipped values, allocate just once if front is None: f = xr.full_like(budgetzone, 0.0, dtype=np.float) if lower is None: l = xr.full_like(budgetzone, 0.0, dtype=np.float) if right is None: r = xr.full_like(budgetzone, 0.0, dtype=np.float) results_front = [] results_lower = [] results_right = [] if "time" in dims: for itime in range(front.coords["time"].size): if front is not None: f = front.isel(time=itime) if lower is not None: l = lower.isel(time=itime) if right is not None: r = right.isel(time=itime) # collect dask arrays df, dl, dr = delayed_collect(indices, f, l, r) # append results_front.append(df) results_lower.append(dl) results_right.append(dr) # Concatenate over time dimension dask_front = dask.array.stack(results_front, axis=0) dask_lower = dask.array.stack(results_lower, axis=0) dask_right = dask.array.stack(results_right, axis=0) else: if front is not None: f = front if lower is not None: l = lower if right is not None: r = right dask_front, dask_lower, dask_right = delayed_collect(indices, f, l, r) else: chunks = (1, *da_shape) dask_front = dask.array.full(front.shape, np.nan, chunks=chunks) dask_lower = dask.array.full(lower.shape, np.nan, chunks=chunks) dask_right = dask.array.full(right.shape, np.nan, chunks=chunks) if netflow: return xr.DataArray(dask_front + dask_lower + dask_right, coords, dims) else: return ( xr.DataArray(dask_front, coords, dims), xr.DataArray(dask_lower, coords, dims), xr.DataArray(dask_right, coords, dims), )