User-guide - pandas : Add alternative to xarray.Dataset.from_dataframe

doc/ecosystem.rst

@@ -74,6 +74,7 @@ Extend xarray capabilities
 - `Collocate <https://github.com/cistools/collocate>`_: Collocate xarray trajectories in arbitrary physical dimensions
 - `eofs <https://ajdawson.github.io/eofs/>`_: EOF analysis in Python.
 - `hypothesis-gufunc <https://hypothesis-gufunc.readthedocs.io/en/latest/>`_: Extension to hypothesis. Makes it easy to write unit tests with xarray objects as input.
+- `ntv-pandas <https://github.com/loco-philippe/ntv-pandas>`_ : A tabular analyzer and a semantic, compact and reversible converter for multidimensional and tabular data
 - `nxarray <https://github.com/nxarray/nxarray>`_: NeXus input/output capability for xarray.
 - `xarray-compare <https://github.com/astropenguin/xarray-compare>`_: xarray extension for data comparison.
 - `xarray-dataclasses <https://github.com/astropenguin/xarray-dataclasses>`_: xarray extension for typed DataArray and Dataset creation.

doc/user-guide/pandas.rst

@@ -110,6 +110,62 @@ work even if not the hierarchical index is not a full tensor product:
     s[::2]
     s[::2].to_xarray()
 
+Lossless and reversible converter
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The previous example shows that the conversion is not reversible (lossless roundtrip) and that the size of the ``datasets`` increases.
+
+Another approach is to use a lossless and reversible conversion (e.g Third party `ntv-pandas`__ libraries). A dataset can then be shared
+between several tools.
+
+__ https://github.com/loco-philippe/ntv-pandas/blob/main/README.md
+
+DataFrame to Dataset or DataArray
+---------------------------------
+
+The conversion is done without loss, by finding the multidimensional structure hidden by the tabular structure.
+
+By applying this conversion to the DataFame above, we find the initial ``Dataset``:
+
+.. ipython:: python
+
+    import ntv_pandas as npd
+
+    df.npd.to_xarray()
+
+Dataset or DataArray to Dataframe
+---------------------------------
+
+In the other direction, information that is not supported by the DataFrame must be transferred to the DataFrame (e.g. ``attrs`` data).
+
+For this, pandas provides the ``attrs`` attribute.
+
+.. ipython:: python
+
+    ds = xr.Dataset(
+        {"foo": (("x", "y"), np.random.randn(2, 3))},
+        coords={
+            "x": [10, 20],
+            "y": ["a", "b", "c"],
+            "along_x": ("x", np.random.randn(2)),
+            "scalar": 123,
+        },
+        attrs={"example": "test npd"},
+    )
+    ds
+
+After reverse conversion, we find the initial ``Dataset``:
+
+.. ipython:: python
+
+    df = npd.from_xarray(ds)
+
+    df.npd.to_xarray()
+
+.. note::
+
+    The pandas ``attrs`` attribute is still experimental (some operations remove it). The associated information must therefore be processed as a priority
+
 Multi-dimensional data
 ~~~~~~~~~~~~~~~~~~~~~~