Same computation, different scenes: a single-site comparison of Sentinel-2 L2A composites from Google Earth Engine and Microsoft Planetary Computer

Heiber Alexander Pantevez Villamil 

Google Earth Engine (GEE) and Microsoft Planetary Computer (MPC) distribute Sentinel-2 Level-2A surface reflectance, and their outputs can be combined in one time series. Over a 73 ha cattle farm in northern Colombia, we tested whether the two platforms yield the same parcel-scale vegetation and water indices and, where they differ, whether the difference originates in the product (the L2A scene), the processing chain (masking, 3 × 3 smoothing, NDVI, EVI, NDWI and monthly median) or the scene universe (which acquisitions enter a composite). For the ten pairs in which both platforms served the same product, scene classification and harmonised reflectance matched in every comparable pixel, and the same chain run locally or server-side in GEE agreed within 10⁻⁶. The catalogues, however, did not hold the same scenes. Between June 2024 and April 2025, 10 of 146 scene–tile records existed only in GEE, six of them Sentinel-2C records from around the satellite's entry into operations, and none only in MPC. In 22 of 136 common records, all from June–July 2024, the catalogues held different L2A generations; of the ten such pairs with valid pixels, one had a different cloud mask. Adding the GEE-only scenes shifted mean monthly NDVI over the farm by −0.0005 (December 2024) and −0.0087 (January 2025), and by up to −0.028 in a public 1 km² demonstration area (January 2025). For the operations tested, the platforms are computationally equivalent, but combining them requires reconciling scene inventories. The reconciliation table, composite scene lists and open-source tools are released.