Item: FORECASTING AVALANCHES IN DATA-SPARSE CENTRAL ASIA USING SNOWPACK AND WEATHER MODELING WITH AVALANCHE OBSERVATIONS
-
-
Title: FORECASTING AVALANCHES IN DATA-SPARSE CENTRAL ASIA USING SNOWPACK AND WEATHER MODELING WITH AVALANCHE OBSERVATIONS
Proceedings: Proceedings, International Snow Science Workshop, Whistler, BC, Canada, 2026
Authors:
- Itai Sheleg [ Lake County High School, Leadville, CO, USA ]
- Doug Chabot [ Latok, LLC, Bozeman, MT, USA ]
- John Snook [ Colorado Avalanche Information Center, CO, USA ]
- Jaime Peters [ Lake County High School, Leadville, CO, USA ]
- Walter Steinkogler [ Wyssen Avalanche Control AG, Switzerland ]
- Ron Simenhois [ Colorado Avalanche Information Center, CO, USA ]
Date: 2026-09-28
Abstract: Avalanches in Central Asia, particularly in Afghanistan, Tajikistan, and Pakistan, are a persistent hazard, causing hundreds of fatalities in severe winters and routinely destroying homes, infrastructure, and livestock essential for survival. To address this risk, in 2015, Aga Khan Agency for Habitat established a remote avalanche forecasting program supported by external expertise. Following its implementation, it became clear that a primary constraint on forecasting capability is the limited availability of snowpack, weather, and avalanche observations, driven by both geographic remoteness and government restrictions on data sharing. We address this gap by developing a proof-of-concept framework that fuses heterogeneous, publicly available data sources to support avalanche forecasting in data-sparse regions. Specifically, we combine numerical weather prediction (NWP)-forced SNOWPACK simulations (4 km grid spacing) with opportunistically extracted avalanche observations from social media posts shared by local residents. This approach treats informal reports as weakly labeled observations, enabling the reconstruction of avalanche occurrence in the absence of systematic records. We compiled 69 valley-floor avalanche observations from social media and AKAH field logs across two winters, and matched each to the nearest SNOWPACK station sharing its predominant aspect, reconstructing the associated meteorological and snowpack conditions at 30 virtual stations. Five daily variables drive the model: total snow depth, 24-hour new snow, maximum air temperature (TAmax), the whole-profile minimum natural stability index (Sn38), and the burial depth of the weakest layer. To distinguish avalanche from non-avalanche days, we fit a hierarchical Bayesian logistic regression that gives each station its own baseline avalanche rate while sharing the influence of the five variables across the network. Trained on 2024–2025 and evaluated on the unseen 2025–2026 season, the model ranks avalanche days above quiet ones (AUC-ROC 0.715), and its warnings are about seven times more precise than chance. Total snow depth carries the most weight, followed by weak-layer burial depth, which lowers risk as it deepens. New snow and TAmax both raise risk, with warming weighing about as much as new-snow loading, suggesting that warm, likely rain-influenced storms are at least as important as loading in driving these events. Detection climbs with a station's accumulated history, from 56% with one recorded avalanche to 100% with three, counting events from any season rather than only the current winter. At a fixed threshold, false alarms do not fall along that same curve, though four of the stations already meet an operational standard of roughly one false alarm per month. This work incorporates a novel data-fusion and weak-supervision framework for avalanche forecasting, integrating NWP-driven snowpack modeling with opportunistic, publicly sourced observations. However, this framework is constrained by a limited event dataset and a model chain that lacks rigorous, component-wise validation, introducing uncertainty in both inputs and predictions. As such, the framework is best interpreted as a decision-support tool rather than a fully autonomous forecasting system. Coupled with expert validation and interpretation, it provides a scalable pathway for developing operational guidance in regions where conventional data streams are sparse or unavailable.
Object ID: ISSW2026_O7.4.pdf
DOI: https://doi.org/10.15788/1790098653
Language of Article: English
Presenter(s): Itai Sheleg
Keywords: Avalanche Forecasting, SNOWPACK, Bayesian Modeling, Central Asia
Page Number(s): 694 - 701
-