pandas 2.0 removals leave four entry points dead

[EnvDroneSense]

Describe the bug
Four code paths call pandas APIs removed in pandas 2.0 (.append() & .iteritems()), so they raise AttributeError on any currently supported pandas. litpop.py:640 is probably the root cause of #826. None of the four is exercised by the test suite, so CI is green.

Location	Removed call
climada/entity/exposures/litpop/litpop.py:640	gdf.append(...)
climada/engine/impact_data.py:422	hit_countries.append(...)
climada/engine/calibration_opt.py:406	df_result.append(df_out, input)
climada/engine/calibration_opt.py:94	years_in_common.iteritems()

To Reproduce
Steps to reproduce the behavior/error:

1. Confirm the APIs are absent on a supported pandas.
2. Call LitPop.from_shape_and_countries with a GeoSeries or list shape (reaches litpop.py:640), or calib_all (reaches calibration_opt.py:406).

Code example:

import pandas as pd, geopandas as gpd
print(pd.__version__)                                  # 2.2.3
print(hasattr(pd.DataFrame, "append"))                  # False
print(hasattr(gpd.GeoDataFrame, "append"))              # False
print(hasattr(pd.Series, "iteritems"))                  # False

# litpop.py:640, in the GeoSeries/list branch of `shape`:
#     gdf = gdf.append(exp.gdf.loc[exp.gdf.geometry.within(shp)])
# -> AttributeError: 'GeoDataFrame' object has no attribute 'append'

Expected behavior
These four functions run. pyproject.toml pins pandas with no upper bound, so any supported version hits this.

Climada Version: develop @ dea46fa (6.1.1-dev)

System Information (please complete the following information):

• Operating system and version: Windows 11 Home 10.0.26200
• Python version: 3.12.14 | packaged by conda-forge | (main, Sep 2 2026, 23:18:20) [MSC v.1944 64 bit (AMD64)]

Additional context

The fix differs per site:

• litpop.py:640 and calibration_opt.py:406 accumulate DataFrames; collect into a list and pd.concat once. This also removes the O(n²) copy the loop currently pays, since .append reallocated the accumulated frame on every iteration.
• impact_data.py:422 appends a dict, not a frame; collect the rows into a list and build one pd.DataFrame(records, columns=[...]) at the end.
• calibration_opt.py:94 needs no restructuring: Series.iteritems() -> Series.items().

calibration_opt.py:406 also passes the builtin input as the second positional argument (ignore_index). Since bool(input) is True, it happened to mean ignore_index=True, so this was a latent bug rather than a visible one.

Two adjacent observations in the litpop.py loop, neither part of the fix above:

• exp.gdf.geometry.within(shp) scans every exposure point per shape with no spatial index, so the loop is O(n_shapes x n_points). The Shape branch at :626 already uses geopandas.sjoin for the same job.
• A point inside two overlapping shapes is matched twice and appears twice in the result, so its value is double-counted in the exposure total. Reproduced on a 5-point toy case with two overlapping squares: 6 rows summing to 180.0, where the unique total is 150.0. geopandas.sjoin(..., predicate="within", how="inner") returns the identical row multiset, so switching would not change this. Whether it should be deduplicated is a question for the maintainers, since it would alter exposure totals for overlapping input shapes.

The test-coverage gap is arguably the more significant finding: four user-facing entry points are dead and nothing detected it.

[emanuel-schmid]

👀 many thanks for pointing this out, @EnvDroneSense 🙌