[Python] Extreme trends in GSL (Growing Season Length) due to missing years in ERA5-Land based...

I calculated the Growing Season Length (GSL) index for the 1950–2023 period in Turkey using ERA5-Land daily mean temperature data. According to the definition:

• First, find the first occurrence of at least 6 consecutive days with daily mean temperature > 5 °C.
• Then, after July 1st (NH), find the first occurrence of at least 6 consecutive days with daily mean temperature < 5 °C.
• The number of days between these two dates is recorded as GSL.

The issue:

• In coastal areas or warmer regions, the second condition (6 consecutive days < 5 °C) may never occur in some years, so GSL cannot be calculated and remains NaN.
• For example, between 2000–2023, in some pixels GSL is missing for years like 2004, 2007, 2009, and 2012.
• When I compute trends (e.g., Mann-Kendall), these missing years result in artificially high positive or negative trends in some pixels, which are not realistic.

My questions:

• How should I handle such missing years before performing trend analysis?

What I tried:

• Implemented GSL following the ETCCDI definition with ERA5-Land daily mean temperatures (1950–2023).
• Calculated GSL pixel-wise using xarray + numpy.
• Applied Mann-Kendall trend test on the yearly GSL time series. What I expected:
• A realistic spatial distribution of GSL trends, e.g., gradual lengthening or shortening in certain regions.
• Missing years to have minimal influence on long-term trends. What happened instead:
• Pixels with missing years (due to no 6-day <5 °C period) show very large artificial trends (sometimes >±10 days/decade), which are not climatologically reasonable.

Continue reading...