Add a function to fill NaNs in grids by interpolation

verde/tests/test_utils.py

@@ -27,9 +27,22 @@
     meshgrid_to_1d,
     parse_engine,
     partition_by_sum,
+    fill_nans
 )
 
 
+def test_fill_nans():
+    """
+    This function tests the fill_nans function.
+    """
+
+    grid = np.array([[1, np.nan, 3],
+                     [4, 5, np.nan],
+                     [np.nan, 7, 8]])
+    filled_grid = fill_nans(grid)
+    assert np.isnan(filled_grid).sum() == 0
+
+
 def test_parse_engine():
     "Check that it works for common input"
     assert parse_engine("numba") == "numba"

verde/utils.py

@@ -14,6 +14,7 @@
 import pandas as pd
 import xarray as xr
 from scipy.spatial import cKDTree
+from sklearn.impute import KNNImputer
 
 try:
     from pykdtree.kdtree import KDTree as pyKDTree
@@ -681,6 +682,36 @@ def kdtree(coordinates, use_pykdtree=True, **kwargs):
     return tree
 
 
+def fill_nans(grid, n_neighbors=1):
+    """
+    This methos is responsible for fill the NaN values in the grid using the KNN algorithm.
+
+    Parameters
+    ----------
+    grid : :class:`xarray.Dataset` or :class:`xarray.DataArray`
+        A 2D grid with one or more data variables.
+    n_neighbors : int
+        Number of nearest neighbors to use to fill the NaN values in the grid. 
+        The greater the quantity, the longer the processing time, depending on the size of the matrix
+
+    Returns
+    -------
+    grid : :class:`xarray.Dataset` or :class:`xarray.DataArray`
+        A 2D grid with the NaN values filled.
+    """
+
+    not_nan_values = np.argwhere(~np.isnan(grid)).reshape(-1, 1)
+    unknown_indices = np.argwhere(np.isnan(grid))
+
+    knn_imputer = KNNImputer(n_neighbors=n_neighbors)
+    knn_imputer.fit(not_nan_values)
+
+    predicted_values = knn_imputer.transform(not_nan_values)
+    for i, idx in enumerate(unknown_indices):
+        grid[tuple(idx)] = predicted_values[i]
+
+    return grid
+
 def partition_by_sum(array, parts):
     """
     Partition an array into parts of approximately equal sum.