Snow extend and snow change mapping with Sentinel-1 imaging using SVM

Snow dynamics is a key parameter for the hydrological model predicting the river flow rate used in dam management. In the MORDOR model used by EDF, the information of the daily snow extent is an input to improve the flow prediction. This information is extracted from MODIS NDSI daily product. Due to cloud cover, this information can be lacking or imprecise for multiple consecutive days over one catchment, reducing the precision of the prediction.

The goal of this study is to detect the snow extent using SAR data, since it can acquire images through clouds. We focus over the Guil catchment in the French Alps. Sentinel-1 interferometric stacks from June 2018 to August 2019 are used for three different orbits.

Previous studies showed the capacity of the ratio between the current image and a reference image acquired in summer to detect wet snow [Nagler2016], or that the ratio between VH and VV could be linked to the height of snow [Lievens2019]. Interferometry has be shown capable to detect snow since the snow covered area can exhibit a lower coherence [Singh2008].

To compare these parameters using a ground truth, we projected the MODIS NDSI data on our S1-stack using a 1m DEM and considered pixels as snowy if the NDSI is above 0.4.

As pointed in other studies [Löw2002, Wang2015], it is very hard to set a threshold for these parameters, mostly because the vegetation exhibits volume scattering and changes the same way as snow. Using SVM, we investigated the capability of these parameters to detect snow in two setups:
- snow detection: the goal is to classify the pixels as snow or snow-free for all the image, using Nagler parameter in VV and VH, the ratio between VH and VV at each date and the polarimetric coherence at each date. For Nagler parameter, the reference image is the temporal average of the images over July and August 2018.
- change detection: the goal is to classify the pixels into 4 classes, snow-free to snow-free, snow to snow, snow-free to snow and snow to snow-free. Considering two consecutive images, this was done using the variation of the VV and VH ratio, the interferometric coherence between these images, and the ratio between the polarimetric coherences of the images.
For each setup, the learning and the testing were done on two samples of 20000 randomly selected pixels, equally distributed between the classes.

For the snow detection method, between 54% and 59% of the pixels are correctly classified, for the three orbits. This result is stable with the choice of the learning sample. For the change detection setup only 30% of the pixels are correctly classified. Moreover, the per-class metrics vary widely from one experience to the other. This variability as well as the low classification results underline the difficulty of the task but can also be linked to the resolution difference between MODIS used as ground truth and S1. To robustify the detection, spatial and temporal regularization seems necessary.