Accurate records of snow avalanche activity are crucial for hazard assessment, avalanche forecasting, and estimating the probability of extreme runout distances, yet such datasets are limited in many mountain regions. This study combines dual-polarization Sentinel-1 Synthetic Aperture Radar (SAR) imagery with a deep learning model to automatically detect avalanche deposits in the Central Rhaetian Alps over a ten-year period. Two independent SAR image series, acquired at different times of day, are analysed to assess the effects of acquisition timing and snow conditions on detection performance. An Avalanche Activity Index, derived from both SAR and manual observations, allows us to classify scenes according to valid, false-positive, or false-negative predominance, mitigating gaps in manual datasets. Both dry and wet snow avalanche cycles are reproduced without significant preferential performance, while systematic detection errors are mainly associated with snow conditions at the time of post-event acquisition, and snow and weather evolution during SAR revisits. False negatives are related to snowmelt and the presence of dry, spread deposits, with new snow in the scene. On the other hand, false positives primarily result from general snow refreezing and, more rarely, from wet snow on tree canopies. A comparison with potential avalanche paths simulated with a dynamic model reveals recurrent biases in areas such as lakes, ski slopes, and anthropic terrain. Finally, we conduct the first probabilistic runout distance estimation based on avalanche annual maximum Runout Ratios automatically extracted from SAR data and fitted to a generalized extreme value distribution. This pilot study demonstrates that the combination of SAR imagery and deep learning reliably reconstructs historical avalanche activity, complements traditional records, and supports probabilistic hazard assessment in regions with sparse ground-based monitoring.