Feature/add spatial disaggregation recipe

docs/api.rst

@@ -59,3 +59,14 @@ Groupers
    DAY_GROUPER
    MONTH_GROUPER
    PaddedDOYGrouper
+
+Spatial Models
+=================
+
+Xarray Wrappers
+~~~~~~~~~~~~~~~
+
+.. autosummary::
+   :toctree: generated/
+
+   SpatialDisaggregator

docs/roadmap.rst

@@ -59,6 +59,8 @@ Components
 1. `pointwise_models`: a collection of linear models that are intended to be
    applied point-by-point. These may be sklearn Pipelines or custom sklearn-like
    models (e.g. BCSDTemperature).
+2. `spatial_models`: a collection of spatial models where the models have spatial
+   dependence.
 2. `global_models`: (not implemented) concept space for deep learning-based
    models.
 3. `metrics`: (not implemented) concept space for a benchmarking suite
@@ -82,6 +84,50 @@ Pointwise models
 
 Other methods, like LOCA, MACA, BCCA, etc, should also be possible.
 
+Spatial models
+~~~~~~~~~~~~~~~~
+
+This category of methods includes models that have spatial dependence.
+Thus far only the BCSD method (see Thrasher et al., 2012) for spatial disaggregation has
+been implemented (for temperature and precipitation). Because the method involves regridding,
+which depends on the xESMF package, a recipe is provided herein versus a full implementation
+in scikit-downscale.
+
+.. code-block:: Python
+
+    from skdownscale.spatial_models import SpatialDisaggregator, apply_weights
+    import xesmf as xe
+    import xarray as xr
+
+    # da_climo_fine: xarray.DataArray (daily multi-decade climatology)
+    # da_obs : xarray.DataArray (daily obs data)
+    # da_model_train: xarray.DataArray (daily training data)
+    # ds_sd: xarray.Dataset (daily spatially disaggregated data)
+
+    # regrid climo to model resolution
+    obs_to_mod_weights = "filename"
+    regridder_obs_to_mod = xe.Regridder(da_obs.isel(time=0), da_model_train.isel(time=0),
+                                        'bilinear', filename=obs_to_mod_weights, reuse_weights=True)
+    climo_regrid = xr.map_blocks(apply_weights, regridder_obs_to_mod,
+                                          args=[da_climo_fine])
+    climo_coarse = climo_regrid.compute()
+
+    # fit the scaling factor model
+    sfc = SpatialDisaggregator.fit(da_model_train, climo_coarse, var_name='tasmax')
+
+    # regrid scale factor
+    mod_to_obs_weights = "filename"
+    regridder_mod_to_obs = xe.Regridder(da_model_train.isel(time=0),
+                                        da_obs.isel(time=0), 'bilinear',
+                                        filename=mod_to_obs_weights, reuse_weights=True)
+    sff_regrid = xr.map_blocks(apply_weights, regridder_mod_to_obs, args=[sfc])
+    sff = sff_regrid.compute()
+
+    # predict using scale factor
+    ds_varname = 'scale_factor_fine'
+    sff_ds = sff.to_dataset(name=ds_varname)
+    ds_sd = SpatialDisaggregator.predict(sff_ds, da_climo_fine, var_name=ds_varname)
+
 Global models
 ~~~~~~~~~~~~~
 
@@ -110,7 +156,7 @@ Dependencies
 ------------
 
 - Core: Xarray, Pandas, Dask, Scikit-learn, Numpy, Scipy
-- Optional: Statsmodels, Keras, PyTorch, Tensorflow, etc.
+- Optional: Statsmodels, Keras, PyTorch, Tensorflow, xESMF, etc.
 
 Related projects
 ----------------

setup.cfg

@@ -11,7 +11,7 @@ select=B,C,E,F,W,T4,B9
 
 [isort]
 known_first_party=skdownscale
-known_third_party = click,matplotlib,numpy,pandas,pkg_resources,probscale,pytest,scipy,seaborn,setuptools,sklearn,xarray,xsd
+known_third_party = click,matplotlib,numpy,pandas,pkg_resources,probscale,pytest,scipy,seaborn,setuptools,sklearn,xarray,xesmf,xsd
 multi_line_output=3
 include_trailing_comma=True
 line_length=100

skdownscale/pointwise_models/arrm.py

@@ -12,7 +12,7 @@
 
 
 def arrm_breakpoints(X, y, window_width, max_breakpoints):
-    '''Calculate breakpoints in x and y
+    """Calculate breakpoints in x and y
 
     Parameters
     ----------
@@ -26,7 +26,7 @@ def arrm_breakpoints(X, y, window_width, max_breakpoints):
     Returns
     -------
     breakpoints : ndarray
-    '''
+    """
     min_width = 10
 
     npoints = len(X)
@@ -97,7 +97,7 @@ def arrm_breakpoints(X, y, window_width, max_breakpoints):
 
 
 class PiecewiseLinearRegression(RegressorMixin, BaseEstimator):
-    """ Piecewise Linear Regression
+    """Piecewise Linear Regression
 
     Parameters
     ----------

skdownscale/pointwise_models/gard.py

@@ -22,7 +22,7 @@ class AnalogBase(RegressorMixin, BaseEstimator):
     _fit_attributes = ['kdtree_', 'X_', 'y_', 'k_']
 
     def fit(self, X, y):
-        """ Fit Analog model using a KDTree
+        """Fit Analog model using a KDTree
 
         Parameters
         ----------
@@ -54,7 +54,7 @@ def fit(self, X, y):
 
 
 class AnalogRegression(AnalogBase):
-    """ AnalogRegression
+    """AnalogRegression
 
     Parameters
     ----------
@@ -82,7 +82,7 @@ def __init__(self, n_analogs=200, kdtree_kwargs=None, query_kwargs=None, lr_kwar
         self.lr_kwargs = lr_kwargs
 
     def predict(self, X):
-        """ Predict using the AnalogRegression model
+        """Predict using the AnalogRegression model
 
         Parameters
         ----------
@@ -130,7 +130,7 @@ def _predict_one_step(self, lr_model, X):
 
 
 class PureAnalog(AnalogBase):
-    """ PureAnalog
+    """PureAnalog
 
     Attributes
     ----------

skdownscale/pointwise_models/grouping.py

@@ -5,7 +5,7 @@
 
 
 class GroupedRegressor:
-    """ Grouped Regressor
+    """Grouped Regressor
 
     Wrapper supporting fitting seperate estimators distinct groups
 
@@ -45,7 +45,7 @@ def __init__(
         self.predict_grouper_kwargs = predict_grouper_kwargs
 
     def fit(self, X, y, **fit_kwargs):
-        """ Fit the grouped regressor
+        """Fit the grouped regressor
 
         Parameters
         ----------
@@ -76,7 +76,7 @@ def fit(self, X, y, **fit_kwargs):
         return self
 
     def predict(self, X):
-        """ Predict estimator target for X
+        """Predict estimator target for X
 
         Parameters
         ----------
@@ -100,7 +100,7 @@ def predict(self, X):
 
 
 class PaddedDOYGrouper:
-    """ Grouper to group an Index by day-of-year +/ pad
+    """Grouper to group an Index by day-of-year +/ pad
 
     Parameters
     ----------

skdownscale/pointwise_models/quantile.py

@@ -17,7 +17,7 @@
 
 
 def plotting_positions(n, alpha=0.4, beta=0.4):
-    '''Returns a monotonic array of plotting positions.
+    """Returns a monotonic array of plotting positions.
 
     Parameters
     ----------
@@ -35,7 +35,7 @@ def plotting_positions(n, alpha=0.4, beta=0.4):
     See Also
     --------
     scipy.stats.mstats.plotting_positions
-    '''
+    """
     return (np.arange(1, n + 1) - alpha) / (n + 1.0 - alpha - beta)
 
 
@@ -182,10 +182,10 @@ def fit(self, X, y, **kwargs):
         self : object
         """
         X = check_array(
-            X, dtype='numeric', ensure_min_samples=2 * self.n_endpoints + 1, ensure_2d=True
+            X, dtype='numeric', ensure_min_samples=2 * self.n_endpoints + 1, ensure_2d=True,
         )
         y = check_array(
-            y, dtype='numeric', ensure_min_samples=2 * self.n_endpoints + 1, ensure_2d=False
+            y, dtype='numeric', ensure_min_samples=2 * self.n_endpoints + 1, ensure_2d=False,
         )
 
         X = check_max_features(X, n=1)
@@ -287,9 +287,9 @@ def predict(self, X, **kwargs):
         return y_hat
 
     def _calc_extrapolated_cdf(
-        self, data, sort=True, extrapolate=None, pp_min=SYNTHETIC_MIN, pp_max=SYNTHETIC_MAX
+        self, data, sort=True, extrapolate=None, pp_min=SYNTHETIC_MIN, pp_max=SYNTHETIC_MAX,
     ):
-        """ Calculate a new extrapolated cdf
+        """Calculate a new extrapolated cdf
 
         The goal of this function is to create a CDF with bounds outside the [0, 1] range.
         This allows for quantile mapping beyond observed data points.

skdownscale/pointwise_models/zscore.py

@@ -7,7 +7,7 @@
 
 
 class ZScoreRegressor(TimeSynchronousDownscaler):
-    """ Z Score Regressor bias correction model wrapper
+    """Z Score Regressor bias correction model wrapper
 
     Apply a scikit-learn model (e.g. Pipeline) point-by-point. The pipeline
     must implement the fit and predict methods.
@@ -28,7 +28,7 @@ def __init__(self, window_width=31):
         self.window_width = window_width
 
     def fit(self, X, y):
-        """ Fit Z-Score Model finds the shift and scale parameters
+        """Fit Z-Score Model finds the shift and scale parameters
         to inform bias correction.
 
         Parameters
@@ -65,7 +65,7 @@ def fit(self, X, y):
         return self
 
     def predict(self, X):
-        """ Predict performs the z-score bias correction
+        """Predict performs the z-score bias correction
         on the future model dataset, X.
 
         Parameters

skdownscale/spatial_models/__init__.py

@@ -0,0 +1,2 @@
+from .regridding import apply_weights
+from .sd import SpatialDisaggregator

skdownscale/spatial_models/regridding.py

@@ -0,0 +1,8 @@
+import xesmf as xe
+
+
+def apply_weights(regridder, input_data):
+    regridder._grid_in = None
+    regridder._grid_out = None
+    result = regridder(input_data)
+    return result

skdownscale/spatial_models/sd.py

@@ -0,0 +1,124 @@
+import numpy as np
+import pandas as pd
+import xarray as xr
+
+
+class SpatialDisaggregator:
+    """
+    Spatial disaggregation model class
+    Apply spatial disaggregation algorithm to an xarray Dataset with fit
+    and predict methods using NASA-NEX method for spatial disaggregation
+    (see Thrasher et al, 2012).
+
+    Parameters
+    ----------
+    var : str
+          specifies the variable being downscaled. Default is
+          temperature and other option is precipitation.
+    """
+
+    def __init__(self, var='temperature'):
+        self._var = var
+
+        if var == 'temperature':
+            pass
+        elif var == 'precipitation':
+            pass
+        else:
+            raise NotImplementedError(
+                'functionality for spatial disaggregation' ' of %s has not yet been added' % var
+            )
+
+    def fit(self, ds_bc, climo_coarse, var_name, lat_name='lat', lon_name='lon'):
+        """
+        Fit the scaling factor used in spatial disaggregation
+
+        Parameters
+        -----------
+        ds_bc : xarray.Dataset
+                Daily bias corrected data at the model resolution
+        climo_coarse : xarray.DataArray or xarray.Dataset
+                Observed climatology that has been regridded (coarsened)
+                to the model resolution
+        var_name : str
+                  Specifies the data variable name within ds_bc of the
+                  daily bias corrected data
+        lat_name : str
+                  Name of the latitude dimension of ds_bc and climo_coarse,
+                  default is 'lat'.
+        lon_name : str
+                  Name of the longitude dimension of ds_bc and climo_coarse,
+                  default is 'lon'.
+        """
+
+        # check that climo has been regridded to model res
+        if not np.array_equal(ds_bc[lat_name], climo_coarse[lat_name]):
+            raise ValueError('climo latitude dimension does not match model res')
+        if not np.array_equal(ds_bc[lon_name], climo_coarse[lon_name]):
+            raise ValueError('climo longitude dimension does not match model res')
+
+        scf = self._calculate_scaling_factor(ds_bc, climo_coarse, var_name, self._var)
+
+        return scf
+
+    def predict(self, scf, climo_fine, var_name, lat_name='lat', lon_name='lon'):
+        """
+        Predict (apply) the scaling factor to the observed climatology.
+
+        Parameters
+        -----------
+        scf : xarray.Dataset
+              Scale factor that has been regridded to the resolution of
+              the observed data.
+        climo_fine : xarray.DataArray or xarray.Dataset
+                     Observed climatology at its native resolution
+        var_name : str
+                   Specifies the data variable name within ds_bc of the
+                   daily bias corrected data
+        lat_name : str
+                   Name of the latitude dimension of ds_bc and climo_coarse,
+                   default is 'lat'.
+        lon_name : str
+                   Name of the longitude dimension of ds_bc and climo_coarse,
+                   default is 'lon'
+        """
+
+        # check that scale factor has been regridded to obs res
+        if not np.array_equal(scf[lat_name], climo_fine[lat_name]):
+            raise ValueError('scale factor latitude dimension does not match obs res')
+        if not np.array_equal(scf[lon_name], climo_fine[lon_name]):
+            raise ValueError('scale factor longitude dimension does not match obs res')
+
+        downscaled = self._apply_scaling_factor(scf, climo_fine, var_name, self._var)
+
+        return downscaled
+
+    def _calculate_scaling_factor(self, ds_bc, climo, var_name, var):
+        """
+        compute scaling factor
+        """
+        # Necessary workaround to xarray's check with zero dimensions
+        # https://github.com/pydata/xarray/issues/3575
+        da = ds_bc[var_name]
+        if sum(da.shape) == 0:
+            return da
+        groupby_type = ds_bc.time.dt.dayofyear
+        gb = da.groupby(groupby_type)
+
+        if var == 'temperature':
+            return gb - climo
+        elif var == 'precipitation':
+            return gb / climo
+
+    def _apply_scaling_factor(self, scf, climo, var_name, var):
+        """
+        apply scaling factor
+        """
+        groupby_type = scf.time.dt.dayofyear
+        da = scf[var_name]
+        sff_daily = da.groupby(groupby_type)
+
+        if var == 'temperature':
+            return sff_daily + climo
+        elif var == 'precipitation':
+            return sff_daily * climo