<p>As glaciers shrink, high interest in their near real-time mass balance arises. This is mainly for two reasons: first, there are concerns about water availability and short-term water resource planning, and second, glaciers are one of the most prominent indicators of climate change, resulting in a high interest of the broader public.</p><p>To satisfy both interests regarding information on near real-time mass balance, we are running the project CRAMPON &#8211; &#8220;Cryospheric Monitoring and Prediction Online&#8221;. Within this project, we set up an operational assimilation platform where it is possible to query daily mass balance estimates in near real-time, i.e. updated with a lag of max. 24 hours. During the operational alpha phase, we increase the amount of modelled glaciers and assimilated observations steadily. We start with about 15 glaciers from the Glacier Monitoring Switzerland (GLAMOS) program, for which time series of seasonal mass balances from the glaciological method are available. After that, we expand our set of modelled glaciers to about 50 glaciers that have frequent geodetic mass balances in the past, and finally to all glaciers in Switzerland. The assimilated observations reach from the operational GLAMOS seasonal mass balance observations via daily point mass balances from nine in situ cameras providing instantaneous ablation rates to satellite-derived albedo and snow distribution on the glacier.<br>As basis for the platform, we run an ensemble of three temperature index and one simplified energy balance melt models. This ensemble takes gridded temperature, precipitation and radiation as input and aims at quantifying uncertainties of the produced daily mass balances. To determine uncertainties in the model prediction of a current mass budget year correctly, we run the models with parameter distributions we have fitted on individual parameter sets calibrated in the past. Since a purely model-based prediction can reveal high uncertainties though, we choose a sequential data assimilation approach in the form of a Particle Filter to constrain this uncertainty with observations, whenever available. We have customized the standard Particle Filter to (1) use a resampling method that is able to keep models in the ensemble despite a temporary bad performance, and (2) allow parameter and model probability evolution over time.</p><p>In this contribution, we focus on giving a holistic overview over the already existing platform features and discuss the future developments. We plan to make the calculated mass balances publicly available in summer 2021, and to extend this platform to the global scale at a later stage.</p>
Temporal prediction of in-hospital falls using tensor factorisation
