Validation of S3 OLCI observations using one year of semi-continuous WISPstation measurementsin the high dynamic area of the Eems Estuary

The Eems Estuary is a very dynamic area featuring highly variable turbidity and Chlorophyll-a values. There is an interest to decrease theturbidity and monitoring is being put into place to observe the current status and changes. Remote sensing using Sentinel 3 OLCIobservations is a candidate monitoring technique but should provide robust and validated results. Obtaining high quality turbidity estimates starts with validated Bottom of Atmosphere reflectances. To validate BOA reflectances a WISPstation (Peters et al., 2019) was placed on a fixed structure in open water at 53.4743N and 6.8216Wfrom 13-11-2018 until 05-11-2019. The WISPstation contains 2 sets of sensors (Lup: 40 degrees, Lsky: 40 degrees and Ed): measurementswere taken in two directions, N and NE. Some shadowing of the Ed sensors occurred and needed to be filtered out. The hyperspectral measurements (350-1100 nm; 0.44 nm/pixel or 4.65 nm FWHM) were convoluted to OLCI spectral bands usingappropriate spectral response functions. Rrs was calculated using a rho retrieved from the Mobley (1999) table in combination with thesimilarity spectrum approach (Ruddick et al, 2006). Because the WISPstation is measuring all channels with one spectrometer, it is veryunsensitive to any uncertainties in the radiometric calibration. In total 61 cloud free S3 OLCI matchup measurements could be taken (S3A+B) , some of which were later flagged out. The S3 FR matchupdata were collected in 3x3 pixel windows and filtered according to the criteria mentioned in the EUMETSAT validation recommendationsdocument (EUM/SEN3/DOC/19/1092968, v5B). Since the WISPstation takes a measurement every 15 minutes we were able to take validation measurements close in time to the overpass,although one of the conclusions remains that -for this area- validation at exact overpass times would be better. We tested the C2RCC-alt v1 and C2X neural networks processors (SNAP v6) together with Polymer v4.10. We will show a detailed analysis ofthe results of these atmospheric correction processors, in terms of statistical comparison (as e.g. in Warren et al., 2019) and by looking ataveraged spectral shape of the OLCI Rrs spectra. For "all datapoints" scatterplots we find R2 values of 0.6 to 0.8 with slopes from 0.5 to 0.76, Indicating that the atmospheric corrections allunderestimate the in-situ measured values with some spread. For "band averaged" scatterplots we find high R2 values for all three processors (above 0.95) but -again- with slopes significantly deviatingfrom unity: all processors underestimate the reflectance, with Polymer giving the biggest underestimation. The underestimations range from15% (C2X) to 30% (Polymer). For the neural networks the biggest underestimations are found from 490 to 681 nm. Plots of mean spectra reveal that especially the neural networks C2RCC-alt and C2X underestimate for wavelengths smaller than about 700nm. Above that range the performance is much better. Based on the time averaged WISPstation data series we propose tentative correction spectra to improve the accuracy of BOA reflectances forthe studied versions of the atmospheric correction methods. We conclude that the presented method based on stationary in-situmeasurements is suitable to validate the performance of current and future updated atmospheric correction methods. This research was funded by the Dutch Ministry of Infrastructure and Water Management and the H2020 project MONOCLE ( grant agreementNo 776480