INTEGRATION OF PASSIVE AND ACTIVE MULTIFREQUENCY DATA FROM AMSR-2 AND COSMO SKYMED FOR ALPINE SNOW MONITORING AT HIGH RESOLUTION

E. Santi¹, S. Pettinato¹, S. Paloscia¹, P. Pampaloni¹, F. Baroni¹, G. Fontanelli¹, E. Palchetti¹

¹CNR – IFAC

Thanks to the multifrequency observations, the sun and weather independent operation, and the frequent revisiting, microwave satellite radiometers as the JAXA’s AMSR2 have well demonstrated capabilities in monitoring snow parameters. Unfortunately, the potential of these sensors is partially hampered by the coarse spatial resolution, which makes challenging their use in extremely complex environments as the Alps, in which the snow properties suffer dramatic spatial variations at sub-pixel scale.

This study evaluates the possibility of disaggregating the AMSR-2 observations (namely brightness temperatures – Tb) using the ASI’s Cosmo SkyMED (CSK) X-band SAR up the SAR resolution for estimating the snow depth (SD) at ≃20m resolution. The proposed implementation is based on Machine Learning methods that play a twofold role: the algorithm that merges AMSR2 and CSK to disaggregate the Tb is based on Artificial Neural Networks (ANN), while the algorithm that estimates SD by using the disaggregated Tb is based on ANN and Random Forest (RF): this algorithm is a reappraisal of the “HydroAlgo” algorithm (Santi et al. 2014).

The method was implemented and validated in the eastern part of Italian Alps, over an area roughly corresponding to the Val D’Aosta region. Data acquired by AMSR2 and CSK during the winter seasons between 2020 and 2022 have been spatially and temporally co-located. The reference SD data needed for implementing and validating the method have been derived from a network of 36 weather stations Managed by ARPA Piemonte and providing hourly measurements of SD and other snow parameters, DEM and landuse maps have been also considered as auxiliary information for the SD retrieval.

The Tb disaggregation is carried out in three steps, one for each of the involved frequency bands, namely X, Ku, and Ka. Then the disaggregated Tb and the auxiliary information are used as input of the reappraised HydroAlgo to generate the output SD maps at ≃20 m resolution.

The SD retrievals based on the disaggregated AMSR2 data are finally validated against in-situ measurements from the weather stations: the obtained correlation coefficient between estimated and target SD is R= 0.85 for the ANN based retrievals and R=0.86 for the RF based retrievals. The corresponding RMSE is 28 cm for ANN and 27 cm for RF, respectively. Although validation should be spatially and temporally extended, these preliminary results suggest a promising potential of the proposed technique for high resolution mapping of SD in alpine environments.

References

Santi E., S. Pettinato, S. Paloscia, P. Pampaloni, G. Fontanelli, A. Crepaz, M. Valt, 2014. Monitoring of Alpine snow using satellite radiometers and artificial neural networks. Remote Sensing of Environment, 144, 179-186. http://dx.doi.org/10.1016/j.rse.2014.01.012