Robust nodata detection in raster builders

libpysal/weights/raster.py

@@ -353,11 +353,22 @@ def _da2wsp(
     ser = da.to_series()
     dtype = np.int32 if (shape[0] * shape[1]) < 46340**2 else np.int64
     nodata = get_nodata(da)
-    if nodata is not None:
-        mask = (ser != nodata).to_numpy()
+
+    values = ser.to_numpy()
+    mask = None
+
+    # Handle NaN masking for float rasters (explicit or implicit nodata)
+    if np.issubdtype(values.dtype, np.floating) and (
+        nodata is None or (isinstance(nodata, float) and np.isnan(nodata))
+    ):
+        mask = ~np.isnan(values)
+    else:
+        mask = values != nodata
+
+    if mask is not None:
         ids = np.where(mask)[0]
         id_map = _idmap(ids, mask, dtype)
-        ser = ser[ser != nodata]
+        ser = ser[mask]
     else:
         ids = np.arange(len(ser), dtype=dtype)
         id_map = ids.copy()
@@ -393,7 +404,7 @@ def _da2wsp(
     else:
         # Fallback method to build sparse matrix
         sw = lat2SW(*shape, criterion)
-        if "nodatavals" in da.attrs and da.attrs["nodatavals"]:
+        if mask is not None:
             sw = sw[mask]
             sw = sw[:, mask]
         return sw, ser

libpysal/weights/tests/test_raster.py

@@ -105,3 +105,60 @@ def test_dataarray(self):
         )
         w = Queen.from_xarray(da)
         assert w.n == 97232
+
+
+class TestNodata:
+    def setup_method(self):
+        self.xr = pytest.importorskip("xarray")
+
+    def test_nan_nodata_float_raster_not_masked(self):
+        """
+        Float rasters with NaN nodata should exclude NaN cells from contiguity,
+        but current logic does not mask them since `ser != np.nan` is always True.
+        """
+        data = np.array([[1.0, 1.0, 1.0], [1.0, np.nan, 1.0], [1.0, 1.0, 1.0]])
+        da = self.xr.DataArray(data, dims=("y", "x"))
+
+        # Expected: center NaN pixel excluded -> 8 valid cells
+        # Actual (current bug): NaN is not masked -> 9 cells
+        wsp = raster.da2WSP(da)
+        assert wsp.n == 8
+
+    def test_rio_nodata(self):
+        """
+        Raster with da.rio.nodata should be correctly masked.
+        This guards existing behavior for rioxarray.
+        """
+        pytest.importorskip("rioxarray")
+        import rioxarray  # noqa: F401
+
+        data = np.array([[1, 1, 1], [1, -9999, 1], [1, 1, 1]])
+        # xarray with rioxarray needs coords to be valid usually
+        # for some ops, but write_nodata might be fine
+        da = self.xr.DataArray(data, dims=("y", "x"))
+        da = da.rio.write_nodata(-9999)
+
+        wsp = raster.da2WSP(da)
+        assert wsp.n == 8
+
+    def test_attrs_nodata(self):
+        """
+        Raster with da.attrs['nodatavals'] should be correctly masked.
+        This guards backward compatibility.
+        """
+        data = np.array([[1, 1, 1], [1, -1, 1], [1, 1, 1]], dtype=int)
+        da = self.xr.DataArray(data, dims=("y", "x"))
+        da.attrs["nodatavals"] = (-1,)
+
+        wsp = raster.da2WSP(da)
+        assert wsp.n == 8
+
+    def test_no_nodata(self):
+        """
+        Raster without nodata should include all pixels.
+        """
+        data = np.array([[1.0, 1.0, 1.0], [1.0, 2.0, 1.0], [1.0, 1.0, 1.0]])
+        da = self.xr.DataArray(data, dims=("y", "x"))
+
+        wsp = raster.da2WSP(da)
+        assert wsp.n == 9