Why do some interior 3x3 rolling windows in my DataArray return NaN instead of 0?

The behavior you’re seeing is due to how xarray’s rolling methods handle windows with missing data via the min_periods parameter. By default, when you call

PYTHONCopied
result = arr.rolling(center=True, x=3, y=3).reduce(lambda x, axis: np.all(~np.isnan(x), axis=axis))

xarray requires that each rolling window have a full 3×3 (i.e. 9) valid (non-NaN) entries before it applies the reduction. This means that even though your lambda function would return False (which would naturally convert to 0) when one or more entries in the window are NaN, the rolling operation itself “gives up” on that window because it doesn’t meet the min_periods requirement (which defaults to the window size).

In other words, whenever a single NaN is present in an otherwise interior window, the rolling mechanism refuses to compute the reduction and outputs NaN instead of 0. Only those windows that already have 9 non-NaN values have the reduction applied, and if all of them are non-NaN, the lambda returns True (which you later see as 1).

If you want windows that contain at least one valid data point—even if that means some values are missing—to be computed (and have your lambda return 0 when any NaN is present), you need to adjust min_periods. For example, if you set:

PYTHONCopied
result = arr.rolling(center=True, x=3, y=3, min_periods=1).reduce(lambda x, axis: np.all(~np.isnan(x), axis=axis))

then every window, even at the borders (although you might still get NaN there for other reasons), will be reduced. However, be careful with lowering min_periods because it changes the meaning of the reduction: you now evaluate windows that may not contain the full 3×3 block.

To summarize:

• The NaNs on the border come from the fact that a full 3×3 window isn’t available.
• The interior NaNs occur because when any window has one or more NaNs from the original data, the rolling reduce method doesn’t apply your lambda (as the number of valid points is less than the default min_periods required), so it leaves the result as NaN instead of computing a false (0) value.

This is why your output shows a mixture of NaNs (for windows that didn’t meet the full window criteria) and 1s (for windows that were computed successfully according to your reduction function).