Sentinel-2 L2a offset anomalies

[garnyt]

According to your documentation, the Sentinel-2-L2a collection for surface reflectance has an offset = 0 (collection "sentinel-2-l2a")
I was looking at various scenes in 2025 where this is not the case and when applying the offset = -0.1, half of the surface reflectance values are below 0, which, for vegetation, should not be at all. When comparing to other sensors it seems it should not have an offset at all.

Is this an error in the catalog?
Specific example: S2B_10SGF_20250808_0_L2A

[mhungen]

I can confirm that.

The metadata in your collection is incorrect in some cases. For example, this item claims the offset has not been applied by setting earthsearch:boa_offset_applied: false

When I apply the offset declared in assets.red.raster:bands[0].offest: -0.1 I get many negative reflectance values. Comparing the raw data of B04 with the JP2000 file from the Copernicus STAC item shows all values are off by -1000, indicating the Element84 STAC item is incorrect.

[mhungen]

These items seem to be affected as well:

• S2C_33UXP_20250126_0_L2A
• S2B_33UXP_20250210_0_L2A
• S2C_33UXP_20250218_0_L2A
• S2B_33UXP_20250302_0_L2A
• S2B_33UXP_20250305_0_L2A

[garnyt]

Would it be correct to assume ALL Sentinel-2-L2a offset values should be 0? I run hundreds of scenes and need some knowledge of when to ignore the offset and when not.

[mhungen]

According to #9 the earthsearch:boa_offset_applied flag is the relevant information. Its existence kind of implies that there are some items that were not harmonized.

~~You could switch to the newer "sentinel-2-c1-l2a" collection. It seems like they fixed some of this kind of issue there. Last time I checked this was still WIP though... ~~ since you have already completed processing the data, this is probably not a good option.

[davdma]

Can confirm I just encountered this exact issue with several sentinel-2-l2a products. I cross compared the exact same product from other data sources with verified values, but the one I got from the Earth Element 84 sentinel-2-l2a was incorrect. For example, the product with ID "S2B_16TCM_20220917_0_L2A" has scale=0.0001 and offset=-0.1, but the offset has already been applied. Applying the scaling and offset caused many raw DNs to go negative when they should not.