Speaker
Description
The estimation of snow parameters, such as snow water equivalent (SWE) and liquid water content (LWC), is an important task to support the water management and avalanche warning applications. The purpose of this study is twofold: in the first case we intend to retrieve the SWE using X-band SAR data, in the second case we want to retrieve LWC from X- and C-band SAR data. Many measurement campaigns have been carried out to acquire ground truth data on snow in the South Tyrol and Valle d’Aosta (Italy), and in-situ measurements, as snow depth, density, snow grain radius, temperature, were collected. In order to proceed with the in-depth analysis for the retrieval of SWE, the acquired in-situ snow information was used to simulate the X-band backscatter with the Dense Medium Radiative Transfer (DMRT) electromagnetic model [1]. The sensitivity of the CSK X-band ° to in-situ measurements of SWE was analyzed. Backscattered data were separated by ascending and descending orbits in order not to mix dry with wet snow. The correlation between ° and SWE is confirmed for ascending orbits (early morning), when snow is dry, with rather high correlation coefficient (R = 0.64); whereas the correlation becomes negligible for descending orbits (late afternoon), due to some presence of wet snow. Consequently, two machine learning models were considered to implement the SWE retrieval algorithm. The first was based on Artificial Neural Networks (ANN), [2], whilst the second exploited the Support Vector Regression (SVR) theory [3]. These algorithms were trained with both experimental data and DMRT model simulations, and finally were applied to a selection of CSK HIMAGE HH polarized scenes collected on the South Tyrol test area. The results of the test performed on SWE parameter of the two selected algorithms showed a correlation coefficient R = 0.86 and R = 0.83 for ANN and SVR, respectively. These results demonstrated that the obtained SWE data agrees with the SWE measurement at the control points and in line with the season and the meteorological conditions. Validation should be obviously improved in the future, with a more consistent measurement dataset, in order to verify the replicability of the results.
The LWC retrieval from X- and C-band SAR data was based on ANN and Random Forest (RF) algorithms. In this case, the setup of the experiment considered CSK and Sentinel-1 (S1) data, acquired on the test site of Valle d’Aosta (Italy). The available product mode in this area is Interferometric Wide Swath (IW) and VV polarization. The training was based on data simulated by the Strong Fluctuation Theory (SFT), [4], that was modified to account for the scattering contribution from the soil under snow in the computation of the total backscattering of wet snow, thus significantly improving the simulation accuracy. The scope of the SFT reappraisal was to create a forward electromagnetic model capable of accurate simulation of wet snow backscattering with smaller computational cost than other popular models as the DMRT [1]. The fast computation allowed generating a training set of some thousands backscattering values by using pseudo random input soil and snow parameters whose distribution has been derived from the experimental data available. In this case too, the training was carried out by using the simulated σ° as inputs and the corresponding LWC as target. The trained algorithms were then validated against the in-situ and simulated snow parameters from SNOWPACK model, [5]. The validation of the ANN and RF algorithms showed R = 0.69 and R = 0.79 for C-band data, respectively. Concerning the X-band data, the validation results retrieved R = 0.60 and R = 0.69 for ANN and RF algorithms respectively. In both cases, RF outperformed ANN, despite several attempts for retraining and the iterative processing for defining the ANN best architecture.
[1], Tsang L., J. Pan, D. Liang, Z. Li, D. W. Cline, and Y. Tan, “Modeling active microwave remote sensing of snow using dense media radiative transfer (DMRT) theory with multiple-scattering effects,” IEEE Trans. Geosci. Remote Sens., 45, 4, 2007, pp. 990-1004.
[2] SWE retrieval in Alpine areas with high-resolution COSMO-SkyMed X-band SAR data using Artificial Neural Networks and Support Vector Regression techniques,” Proc. 2020 XXXIII General Assembly and Scientific Symposium of the Int. Union of Radio Science, URSI GASS, 2020, pp. 1-4.
[3] Santi E., De Gregorio L., Pettinato S., Cuozzo G., Jacob A., Notarnicola C., Gunther D., Strasser U., Cigna F., Tapete D., Paloscia S., “On the Use of COSMO-SkyMed X-Band SAR for Estimating Snow Water Equivalent in Alpine Areas: A Retrieval Approach Based on Machine Learning and Snow Models,” (2022), DOI: 10.1109/TGRS.2022.3191409
[4] Jin, Y. Q., Electromagnetic scattering modelling for quantitative remote sensing, World Scientific, 1993.
[5] Lehning, M., Bartelt, P., Brown, B., Fierz, C., & Satyawali, P. (2002a). A physical SNOWPACK model for the Swiss avalanche warning Part II. Snow microstructure. Cold Regions Science and Technology, 35(3), 147–167. https://doi.org/10.1016/S0165-232X(02)00073-3.
