From aabb2e2dfe25c13bde27f8f54336fb472b6875e7 Mon Sep 17 00:00:00 2001
From: Jia Yu <jiayu@apache.org>
Date: Thu, 10 Sep 2026 23:43:30 -0700
Subject: [PATCH] [GH-3333] Handle leading nulls in local GeoSeries
 construction

---
 python/sedona/spark/geopandas/geoseries.py | 18 ++++-
 python/tests/geopandas/test_geoseries.py   | 86 ++++++++++++++++++++++
 2 files changed, 102 insertions(+), 2 deletions(-)

diff --git a/python/sedona/spark/geopandas/geoseries.py b/python/sedona/spark/geopandas/geoseries.py
index 9af03273ba8..4cd4a5289df 100644
--- a/python/sedona/spark/geopandas/geoseries.py
+++ b/python/sedona/spark/geopandas/geoseries.py
@@ -737,8 +737,22 @@ def __init__(
 
             pd_series = pd_series.astype(object)
 
-            # Initialize the parent class PySpark Series with the pandas Series.
-            super().__init__(data=pd_series)
+            if (
+                not pd_series.empty
+                and pd_series.isna().iloc[0]
+                and any(isinstance(value, BaseGeometry) for value in pd_series)
+            ):
+                # Spark 3.5 infers object UDTs from the first value. WKB lets
+                # leading missing values retain geometry type without reordering.
+                wkb_series = (
+                    gpd.GeoSeries(pd_series.where(pd_series.notna(), None))
+                    .to_wkb(include_srid=True)
+                    .rename(pd_series.name)
+                )
+                ps_series = pspd.Series(wkb_series).spark.transform(stc.ST_GeomFromWKB)
+                super().__init__(data=ps_series._anchor, index=ps_series._col_label)
+            else:
+                super().__init__(data=pd_series)
 
         # Ensure we're storing geometry types.
         if (
diff --git a/python/tests/geopandas/test_geoseries.py b/python/tests/geopandas/test_geoseries.py
index 246b72e3226..50197c3135d 100644
--- a/python/tests/geopandas/test_geoseries.py
+++ b/python/tests/geopandas/test_geoseries.py
@@ -677,6 +677,92 @@ def test_constructor(self, obj):
         sgpd_series = sgpd.GeoSeries(obj)
         assert isinstance(sgpd_series, sgpd.GeoSeries)
 
+    @pytest.mark.parametrize(
+        "wrap",
+        [list, tuple, np.asarray, pd.Series, gpd.GeoSeries, gpd.array.from_shapely],
+        ids=["list", "tuple", "numpy", "pandas", "geopandas", "geometry_array"],
+    )
+    def test_constructor_leading_null_local_inputs(self, wrap):
+        from geopandas.testing import assert_geoseries_equal
+
+        _ = self.spark
+        values = [None, Point(1, 0), None, Point()]
+        result = GeoSeries(wrap(values))
+
+        assert_geoseries_equal(
+            result.to_geopandas(), gpd.GeoSeries(values), check_index_type=False
+        )
+        assert result.crs is None
+
+    @pytest.mark.parametrize(
+        "missing", [None, np.nan, pd.NA, pd.NaT, np.datetime64("NaT")]
+    )
+    def test_constructor_leading_missing_values(self, missing):
+        from geopandas.testing import assert_geoseries_equal
+
+        _ = self.spark
+        result = GeoSeries([missing, Point(1, 0), missing])
+
+        assert_geoseries_equal(
+            result.to_geopandas(),
+            gpd.GeoSeries([None, Point(1, 0), None]),
+            check_index_type=False,
+        )
+
+    @pytest.mark.parametrize(
+        "name, inherited_crs, crs",
+        [
+            (None, None, None),
+            ("geometry", "EPSG:4326", None),
+            (("geometry", "shape"), "EPSG:4326", "EPSG:3857"),
+        ],
+    )
+    def test_constructor_leading_null_metadata(self, name, inherited_crs, crs):
+        from geopandas.testing import assert_geoseries_equal
+
+        _ = self.spark
+        index = pd.MultiIndex.from_tuples(
+            [("b", 2), ("a", 1), ("b", 2)], names=["letter", "number"]
+        )
+        values = [None, Point(1, 0, 2), Point(3, 4)]
+        local = gpd.GeoSeries(values, index=index, name=name, crs=inherited_crs)
+        result = GeoSeries(local, crs=crs)
+        expected = gpd.GeoSeries(
+            values, index=index, name=name, crs=crs or inherited_crs
+        )
+
+        assert result.name == name
+        assert_geoseries_equal(result.to_geopandas(), expected)
+        assert_geoseries_equal(
+            local, gpd.GeoSeries(values, index=index, name=name, crs=inherited_crs)
+        )
+
+    @pytest.mark.parametrize("values", [[], [None, None], [Point(1, 0), None]])
+    def test_constructor_null_and_empty_controls(self, values):
+        from geopandas.testing import assert_geoseries_equal
+
+        _ = self.spark
+        assert_geoseries_equal(
+            GeoSeries(values).to_geopandas(),
+            gpd.GeoSeries(values),
+            check_index_type=False,
+        )
+
+    def test_constructor_leading_null_preserves_embedded_srid(self):
+        from shapely import wkb
+        from sedona.spark.sql.types import GeometryType
+
+        _ = self.spark
+        point = wkb.loads(wkb.dumps(Point(1, 2, 3), srid=4326))
+        result = GeoSeries([None, point])
+
+        assert result.spark.data_type == GeometryType()
+        assert result.crs is None
+        assert result.spark.transform(stf.ST_SRID).to_pandas().dropna().tolist() == [
+            4326
+        ]
+        assert result.spark.transform(stf.ST_Z).to_pandas().dropna().tolist() == [3]
+
     def test_constructor_pandas_on_spark(self):
         obj = ps.Series([Point(x, x) for x in range(3)])
         sgpd_series = GeoSeries(obj)