Cross-Mission Methodology for Masking Validation: EnMAP cloud mask with Sentinel-5P

Kurzfassung

The validation of masking algorithms for hyperspectral imagery requires a precise reference, often referred to as ground truth. Currently, most validation exercises for masking algorithms rely on hand-made annotations, but these demand significant qualified manpower, as each validation requires the creation of a new dataset. Additionally, misinterpretation of class definitions often causes overlaps in annotations between classes. Furthermore, other sources of information, such as in-situ measurements, are difficult to obtain on a global scale. We propose a validation methodology that uses more sensitive and global sources of information, employing physical properties to validate masking products. The main objective is to retrieve reference data from missions specifically designed to sense a particular characteristic of the atmosphere or Earth's surface and use it to validate the masking products of other missions, particularly optical remote sensing. To achieve this, it is necessary to implement an interface that transforms the masking products into physical properties that match the format of the mission products used as a reference. This work addresses a use case for a spatial-spectral feature extraction-based neural network, trained with a pre-classification from the Python-based Atmospheric Correction (PACO) software developed at the German Aerospace Center (DLR). It validates the cloud masking of the Environmental Mapping and Analysis Program (EnMAP) that uses PACO as L2A land processor, using the European Space Agency's (ESA) Sentinel- 5P TROPOspheric Monitoring Instrument (TROPOMI) cloud products generated by DLR, such as Cloud Optical Thickness (COT) and cloud cover, as a reference. This validation involves finding matching location overpasses with short time differences between the two missions. Thus, the EnMAP cloud mask is resampled to the spatial resolution of TROPOMI products (5.5 km × 3.5 km) and then validated with the cloud cover product of TROPOMI, using robust metrics to compare the differences.