create_forcings.py

# Standard library
import argparse
import io
import os
import zipfile

# Third-party
import cartopy.crs as ccrs
import cartopy.feature as cfeature
import geopandas as gpd
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import requests
import xarray as xr
from affine import Affine
from anemoi.datasets.grids import cutout_mask
from pyproj import datadir
from rasterio.features import rasterize
from scipy.ndimage import binary_dilation, center_of_mass

proj_lib = os.environ.get("PROJ_LIB")
if proj_lib:
    datadir.set_data_dir(proj_lib)
else:
    print("PROJ_LIB environment variable is not set")


def calculate_datetime_forcing(ds, args):
    """Calculate the datetime forcing for the neural LAM model."""
    time_dim = ds.time
    dt_range = pd.date_range(
        start=pd.to_datetime(time_dim[0].values).to_pydatetime(),
        end=pd.to_datetime(time_dim[-1].values).to_pydatetime(),
        freq="h",
        name="time",
    )
    dt_range_xr = xr.DataArray(dt_range, dims=["time"])
    hours_of_day = xr.DataArray(dt_range_xr.dt.hour, dims=["time"])
    seconds_into_year = xr.DataArray(
        [
            (
                pd.Timestamp(dt_obj)
                - pd.Timestamp(f"{pd.Timestamp(dt_obj).year}-01-01")
            ).total_seconds()
            for dt_obj in dt_range_xr.values
        ],
        dims=["time"],
    )
    hour_angle = (hours_of_day / 12) * np.pi
    year_angle = (seconds_into_year / args.seconds_in_year) * 2 * np.pi
    datetime_forcing = xr.Dataset(
        {
            "hour_sin": np.sin(hour_angle),
            "hour_cos": np.cos(hour_angle),
            "year_sin": np.sin(year_angle),
            "year_cos": np.cos(year_angle),
        },
        coords={"time": dt_range_xr},
    )
    datetime_forcing = (datetime_forcing + 1) / 2
    return datetime_forcing


def download_and_extract_shapefile(url, tempdir):
    """Download and extract a shapefile from a URL."""
    response = requests.get(url, timeout=10)
    with zipfile.ZipFile(io.BytesIO(response.content), "r") as zip_ref:
        zip_ref.extractall(tempdir)
    return next((f for f in zip_ref.namelist() if f.endswith(".shp")), None)


def generate_land_sea_mask(xy_state, tempdir, projection, high_res_factor=10):
    """Generate a land-sea mask for the neural LAM model."""
    url = "https://naturalearth.s3.amazonaws.com/50m_physical/ne_50m_land.zip"
    shp_path = download_and_extract_shapefile(url, tempdir)

    gdf = gpd.read_file(os.path.join(tempdir, shp_path))
    gdf = gdf.to_crs(projection)
    gdf = gdf[gdf.is_valid]
    land_geometry = gdf.unary_union

    x, y = xy_state[0], xy_state[1]
    nrows, ncols = y.shape
    xres = (x.max() - x.min()) / float(ncols)
    yres = (y.max() - y.min()) / float(nrows)

    # align center of pixels
    transform = Affine.translation(
        x.min() - xres / 2, y.min() - yres / 2
    ) * Affine.scale(xres, yres)

    high_res_out_shape = (nrows * high_res_factor, ncols * high_res_factor)
    high_res_transform = transform * transform.scale(
        1 / high_res_factor, 1 / high_res_factor
    )

    high_res_out_image = rasterize(
        shapes=[land_geometry.__geo_interface__],
        out_shape=high_res_out_shape,
        transform=high_res_transform,
        fill=0,
        dtype=np.uint8,
    )

    aggregated_out_image = high_res_out_image.reshape(
        (nrows, high_res_factor, ncols, high_res_factor)
    ).mean(axis=(1, 3))

    land_sea_mask = xr.DataArray(
        aggregated_out_image,
        dims=("y", "x"),
        coords={
            "proj_y": (("y", "x"), y),
            "proj_x": (("y", "x"), x),
        },
    )

    return land_sea_mask


def create_boundary_mask(
    lonlat_state,
    lonlat_boundary,
    boundary_thickness,
    overlap=False,
):
    """Create a boundary mask for the neural LAM model."""
    state_lon, state_lat = lonlat_state[0], lonlat_state[1]
    boundary_lon, boundary_lat = lonlat_boundary[0], lonlat_boundary[1]

    state_lat_flat, state_lon_flat = state_lat.flatten(), state_lon.flatten()
    boundary_lat_flat, boundary_lon_flat = (
        boundary_lat.flatten(),
        boundary_lon.flatten(),
    )

    mask_flat = cutout_mask(
        state_lat_flat, state_lon_flat, boundary_lat_flat, boundary_lon_flat
    )

    mask = mask_flat.reshape((boundary_lat.shape)).astype(bool)
    interior_mask = ~mask

    # Find the center of mass of the interior mask
    center_lat, center_lon = center_of_mass(interior_mask)

    # Calculate the shift needed to center the mask
    shift_lat = interior_mask.shape[0] // 2 - int(center_lat)
    shift_lon = interior_mask.shape[1] // 2 - int(center_lon)

    centered_interior_mask = np.roll(
        interior_mask, shift=(shift_lat, shift_lon), axis=(0, 1)
    )

    # Create the boundary mask with appropriate thickness
    structure = np.ones((3, 3), dtype=bool)
    dilated_mask = binary_dilation(
        centered_interior_mask, structure, boundary_thickness
    )

    dilated_mask = np.roll(dilated_mask, -shift_lat, axis=0)
    dilated_mask = np.roll(dilated_mask, -shift_lon, axis=1)

    if overlap:
        boundary_mask = dilated_mask
    else:
        boundary_mask = dilated_mask & ~interior_mask

    interior_mask = xr.DataArray(
        interior_mask.astype(int),
        dims=("y", "x"),
        coords={
            "lat": (("y", "x"), boundary_lat),
            "lon": (("y", "x"), boundary_lon),
        },
    )
    boundary_mask = xr.DataArray(
        boundary_mask.astype(int),
        dims=("y", "x"),
        coords={
            "lat": (("y", "x"), boundary_lat),
            "lon": (("y", "x"), boundary_lon),
        },
    )
    return interior_mask, boundary_mask


def create_boundary_mask_alt(
    lonlat_state,
    lonlat_boundary,
    crs_state,
    boundary_min_distance,
    boundary_max_distance,
    overlap=False,
):
    """
    Returns a mask such that lonlat_boundary[msk] gives the boundary points
    satisfying the requirements to be within a distance of
    boundary_max_distance from the state area while also not being closer to
    it than boundary_min_distance.
    If overap = True then also the points within the state area are included.
    In this case boundary_min_distance is set to zero.

    Args:
        lonlat_state: A NumPy array of shape (2, sy, sx)
        lonlat_boundary: A NumPy array of shape (2, by, bx)
        crs_state: Instance of cartopy class CRS
        boundary_min_distance: Minimum distance [m]
        boundary_max_distance: Maximum distance [m]
        overlap: Boolean

    Returns:
        msk_interior: A boolean (integer) xarray of shape (y, x) indicating
        the boundary points within the interior.
        msk_boundary: A boolean (integer) xarray of shape (y, x) indicating
        the queried boundary points.
    """

    def distance_points_to_line(points, line_start, line_end):
        """
        Calculates distances from points in an array to a line in 2D.

        Args:
            points: A NumPy array of shape (N, 2) representing N points (x, y).
            line_start: Array with start point (x1, y1) of the line.
            line_end: An array with the end point (x2, y2) of the line.

        Returns:
            A NumPy array of shape (N,) containing the distances for each point.
        """

        # Line direction vector
        line_direction = line_end - line_start

        # Vector from each point to line start
        point_to_start = points - line_start

        # Calculate projection scalars for all points
        t = np.dot(point_to_start, line_direction) / np.sum(line_direction**2)

        # constrain to line segment
        t = np.fmax(0, np.fmin(1, t))

        # Project all points onto the line
        projections = line_start + t[:, np.newaxis] * line_direction

        # Distances between points and their projections
        distances = np.sqrt(np.sum((points - projections) ** 2, axis=1))

        return distances

    by, bx = lonlat_boundary[0].shape
    lon_state, lat_state = lonlat_state[0].ravel(), lonlat_state[1].ravel()
    lon_boundary, lat_boundary = (
        lonlat_boundary[0].ravel(),
        lonlat_boundary[1].ravel(),
    )

    # regular state grid coordinates
    crs_boundary = ccrs.Geodetic()
    coords_state = crs_state.transform_points(
        crs_boundary, lon_state, lat_state
    )

    # state corner points
    xm_state, xM_state = np.amin(coords_state[:, 0]), np.amax(
        coords_state[:, 0]
    )
    ym_state, yM_state = np.amin(coords_state[:, 1]), np.amax(
        coords_state[:, 1]
    )
    x_ll_state, y_ll_state = xm_state, ym_state
    x_ur_state, y_ur_state = xM_state, yM_state
    # ll ul ur lr
    pts_corners = np.array(
        [
            [x_ll_state, y_ll_state],
            [x_ll_state, y_ur_state],
            [x_ur_state, y_ur_state],
            [x_ur_state, y_ll_state],
        ]
    )

    # boundary points in state area projection
    coords_boundary = crs_state.transform_points(
        crs_boundary, lon_boundary, lat_boundary
    )
    # find boundary points outside the state area
    msk_outside = ~(
        (coords_boundary[:, 0] > xm_state)
        & (coords_boundary[:, 0] < xM_state)
        & (coords_boundary[:, 1] > ym_state)
        & (coords_boundary[:, 1] < yM_state)
    )

    # boundary points outside
    pts_outside = np.vstack(
        (coords_boundary[msk_outside, 0], coords_boundary[msk_outside, 1])
    ).T

    # distances from points outside to left, top, right and bottom borders
    d_l = distance_points_to_line(pts_outside, pts_corners[0], pts_corners[1])
    d_t = distance_points_to_line(pts_outside, pts_corners[1], pts_corners[2])
    d_r = distance_points_to_line(pts_outside, pts_corners[2], pts_corners[3])
    d_b = distance_points_to_line(pts_outside, pts_corners[3], pts_corners[0])

    if overlap:
        boundary_min_distance = 0

    min_d, max_d = boundary_min_distance, boundary_max_distance
    msk_min = (d_l > min_d) & (d_t > min_d) & (d_r > min_d) & (d_b > min_d)
    msk_max = (d_l < max_d) | (d_t < max_d) | (d_r < max_d) | (d_b < max_d)

    msk_border = msk_min & msk_max

    # construct mask to apply to input boundary data
    msk_boundary = np.zeros(lon_boundary.shape, dtype="bool")
    msk_boundary[np.where(msk_outside)[0][msk_border]] = True

    msk_interior = np.zeros(lon_boundary.shape, dtype="bool")
    if overlap:
        # add boundary points inside
        msk_interior = (
            (coords_boundary[:, 0] >= xm_state)
            & (coords_boundary[:, 0] <= xM_state)
            & (coords_boundary[:, 1] >= ym_state)
            & (coords_boundary[:, 1] <= yM_state)
        )
        msk_boundary[msk_interior] = True

    msk_interior = xr.DataArray(
        msk_interior.reshape((by, bx)).astype(int),
        dims=("y", "x"),
        coords={
            "lat": (("y", "x"), lat_boundary.reshape((by, bx))),
            "lon": (("y", "x"), lon_boundary.reshape((by, bx))),
        },
    )
    msk_boundary = xr.DataArray(
        msk_boundary.reshape((by, bx)).astype(int),
        dims=("y", "x"),
        coords={
            "lat": (("y", "x"), lat_boundary.reshape((by, bx))),
            "lon": (("y", "x"), lon_boundary.reshape((by, bx))),
        },
    )
    return msk_interior, msk_boundary


def generate_toa_radiation_forcing(ds, lonlat):
    """
    Calculates approximate TOA irradiance (instantaneous values [W*m**-2])
    """

    # solar constant
    E0 = 1366

    dt = ds.time.dt
    lon_m = lonlat[0][np.newaxis, :]
    lat_m = lonlat[1][np.newaxis, :]
    day = dt.dayofyear.values[:, np.newaxis, np.newaxis]
    hr_utc = dt.hour.values[:, np.newaxis, np.newaxis]
    # Eq. 1.6.1a in Solar Engineering of Thermal Processes 4th ed.
    dec = np.pi / 180 * 23.45 * np.sin(2 * np.pi * (284 + day) / 365)
    hr_lst = hr_utc + lon_m / 15
    hr_angle = 15 * (hr_lst - 12)
    # Eq. 1.6.2 with beta=0 in Solar Engineering of Thermal Processes 4th ed.
    cos_sza = np.sin(lat_m * np.pi / 180) * np.sin(dec) + np.cos(
        lat_m * np.pi / 180
    ) * np.cos(dec) * np.cos(hr_angle * np.pi / 180)
    toa_radiation = np.fmax(0, E0 * cos_sza)
    toa_radiation = toa_radiation.transpose(0, 2, 1)

    toa_radiation = xr.DataArray(
        toa_radiation,
        dims=("time", "x", "y"),
        coords={
            "time": ds.time,
            "lat": (("y", "x"), lonlat[1]),
            "lon": (("y", "x"), lonlat[0]),
        },
    )
    return toa_radiation


def main():
    """Main function for creating the datetime forcing and boundary mask."""
    parser = argparse.ArgumentParser()
    parser.add_argument(
        "--data_config", type=str, default="neural_lam/data_config.yaml"
    )
    parser.add_argument("--zarr_path", type=str, default="forcings.zarr")
    parser.add_argument("--tempdir", type=str, default="./shps")
    parser.add_argument("--seconds_in_year", type=int, default=31536000)
    parser.add_argument("--boundary_thickness", type=int, default=30)
    parser.add_argument("--boundary_min_distance", type=int, default=15000)
    parser.add_argument("--boundary_max_distance", type=int, default=150000)
    parser.add_argument("--overlap", type=bool, default=False)
    parser.add_argument("--plot", type=bool, default=True)
    args = parser.parse_args()

    lambert_proj_params = {
        "a": 6367470,
        "b": 6367470,
        "lat_0": 63.3,
        "lat_1": 63.3,
        "lat_2": 63.3,
        "lon_0": 15.0,
        "proj": "lcc",
    }

    globe = ccrs.Globe(
        semimajor_axis=lambert_proj_params["a"],
        semiminor_axis=lambert_proj_params["b"],
    )
    lambert_proj = ccrs.LambertConformal(
        central_longitude=lambert_proj_params["lon_0"],
        central_latitude=lambert_proj_params["lat_0"],
        standard_parallels=(
            lambert_proj_params["lat_1"],
            lambert_proj_params["lat_2"],
        ),
        globe=globe,
    )

    # Make sure the example data is available
    xy_state = np.load("data/meps_example/static/nwp_xy.npy")
    latlong_proj = ccrs.Geodetic()
    lonlatz_state = latlong_proj.transform_points(
        lambert_proj, xy_state[0], xy_state[1]
    )
    lonlat_state = np.stack(
        (lonlatz_state[:, :, 0], lonlatz_state[:, :, 1]), axis=0
    )

    ds_meps = np.load(
        "data/meps_example/samples/train/nwp_2022040100_mbr000.npy"
    )
    data_vars = {
        f"var_{i}": (("time", "y", "x"), ds_meps[:, :, :, i])
        for i in range(ds_meps.shape[3])
    }
    # This is certainly wrong, but its only for demonstration purposes
    reference_time = np.datetime64("1990-01-01T00:00:00", "ns")
    normalized_hours = np.arange(len(ds_meps[:, 0, 0, 0]))
    hourly_timesteps = reference_time + normalized_hours.astype(
        "timedelta64[h]"
    )
    ds_state = xr.Dataset(
        data_vars,
        coords={
            "time": hourly_timesteps,
        },
    )

    # open era5 from weatherbench
    ds_boundary = xr.open_zarr(
        "gs://weatherbench2/datasets/era5/1959-2023_01_"
        "10-wb13-6h-1440x721_with_derived_variables.zarr"
    )
    ds_boundary = ds_boundary.sel(time=slice("1990-01-01", "1990-01-02"))
    lon_boundary, lat_boundary = ds_boundary.longitude, ds_boundary.latitude
    (
        llon,
        llat,
    ) = np.meshgrid(lon_boundary, lat_boundary)
    lonlat_boundary = np.stack((llon, llat), axis=0)

    datetime_forcing = calculate_datetime_forcing(ds_state, args)
    datetime_forcing.to_zarr(args.zarr_path, mode="w")
    print(f"Datetime forcing saved to {args.zarr_path}")

    land_sea_mask = generate_land_sea_mask(xy_state, args.tempdir, lambert_proj)
    land_sea_mask.to_zarr(args.zarr_path, mode="w")
    print(f"Land-sea mask saved to {args.zarr_path}")

    interior_mask, boundary_mask = create_boundary_mask(
        lonlat_state, lonlat_boundary, args.boundary_thickness, args.overlap
    )
    boundary_mask.to_zarr(f"boundary_{args.zarr_path}", mode="w")
    print(f"Boundary mask saved to boundary_{args.zarr_path}")

    toa_radiation_state = generate_toa_radiation_forcing(ds_state, lonlat_state)
    toa_radiation_state.to_zarr(args.zarr_path, mode="w")
    toa_radtiation_boundary = generate_toa_radiation_forcing(
        ds_boundary, lonlat_boundary
    )
    toa_radtiation_boundary.to_zarr(f"boundary_{args.zarr_path}", mode="w")
    print(
        f"TOA radiation saved to {args.zarr_path} "
        f"and boundary_{args.zarr_path}"
    )

    # alternative boundary masks
    interior_mask, boundary_mask = create_boundary_mask_alt(
        lonlat_state,
        lonlat_boundary,
        lambert_proj,
        args.boundary_min_distance,
        args.boundary_max_distance,
        args.overlap,
    )

    if args.plot:

        _, ax = plt.subplots(2, 2, figsize=(10, 8))

        datetime_forcing.hour_sin.plot(ax=ax[0, 0])
        datetime_forcing.hour_cos.plot(ax=ax[0, 1])
        datetime_forcing.year_sin.plot(ax=ax[1, 0])
        datetime_forcing.year_cos.plot(ax=ax[1, 1])

        plt.tight_layout()
        plt.savefig("datetime_forcing_state.png")

        _, ax = plt.subplots(
            subplot_kw={"projection": lambert_proj}, figsize=(10, 6)
        )
        land_sea_mask.plot(
            x="proj_x",
            y="proj_y",
            ax=ax,
            cmap="viridis",
        )
        ax.add_feature(cfeature.BORDERS, linestyle="-", alpha=0.5)
        ax.coastlines()
        ax.gridlines()
        plt.savefig("land_sea_mask.png")

        lons_boundary = boundary_mask.lon.values[boundary_mask == 1].flatten()
        lats_boundary = boundary_mask.lat.values[boundary_mask == 1].flatten()
        lons_interior = interior_mask.lon.values[interior_mask == 1].flatten()
        lats_interior = interior_mask.lat.values[interior_mask == 1].flatten()

        # Transforming points to the projected coordinate system
        transformed_coords_interior = lambert_proj.transform_points(
            ccrs.PlateCarree(), lons_interior, lats_interior
        )
        transformed_coords_boundary = lambert_proj.transform_points(
            ccrs.PlateCarree(), lons_boundary, lats_boundary
        )

        _, ax = plt.subplots(
            subplot_kw={"projection": lambert_proj}, figsize=(10, 6)
        )
        ax.scatter(
            transformed_coords_interior[:, 0],
            transformed_coords_interior[:, 1],
            color="red",
            marker="o",
            s=8,
            transform=lambert_proj,
        )
        ax.scatter(
            transformed_coords_boundary[:, 0],
            transformed_coords_boundary[:, 1],
            color="blue",
            marker="+",
            s=6,
            transform=lambert_proj,
        )
        ax.add_feature(cfeature.BORDERS, linestyle="-", alpha=0.5)
        ax.coastlines()
        ax.gridlines()

        plt.savefig("boundary_mask.png")

        _, ax = plt.subplots(figsize=(10, 6))
        toa_radiation_state_mean = toa_radiation_state.mean(dim="y")
        toa_radiation_state_mean.plot(x="x", y="time", ax=ax, cmap="viridis")
        ax.set_xlabel("X")
        ax.set_ylabel("Time")
        ax.set_title("Hovmöller Diagram - TOA Radiation (State)")
        plt.tight_layout()
        plt.savefig("toa_radiation_state_hovmoller.png")


if __name__ == "__main__":
    main()