MASSIVE - MAchine learning, Surface mass balance of glaciers, Snow cover, In-situ data, Volume change, Earth observation
Glaciers are under pressure in the current climate warming trend. The aerial extent and the mass balance, a measure of mass loss or mass gain of a glacier are representative of the ?health state? of a glacier. They are recognized as Essential Climate Variables by the World Meteorological Organization (WMO). Due to the sheer number of glaciers of roughly 200.000 worldwide and their inaccessibility, satellite data has been a valuable source to map glaciers around the globe during the last decades.
Satellite imagery has enabled researchers to map the extent of glaciers and glacier surface types (glacier facies, i.e. snow, firn, ice). Multi-temporal information on the glacier surface elevation and proxies (snow cover, albedo) which are highly correlated with the glacier mass balance, have been used to estimate mass changes. Today, the ever-growing amount of satellite data motivates the development of new methodology to maximize the information retrieved (added-value) and ease the handling and storage of the data (e.g. data cubes). In MASSIVE, the project team aims at improving glacier mapping and surface glacier mass balance estimation techniques with the help of machine learning, especially deep learning. We will develop the methodology for glaciers in Norway, Svalbard, the European Alps and the Himalayas and then expand it to regions with different glacier characteristics. The project outcome will be a multi-temporal glacier inventory and a multi-annual time series of mass balance of the glaciers under investigation. This data can support national authorities to report on glacier mass balance and be used as input for hydrological run-off modelling.
MASSIVE Partners: University of Oslo (Norway), EURAC (Italy), University of Twente (The Netherlands), The Norwegian Water Directorate (Norway) and Statkraft (Norway).
The MASSIVE project team will revolutionize glacier area mapping and glacier mass balance estimation. We aim at vastly expanding the applicability of today’s glacier surface mass balance regression techniques using deep learning for image classification and regression. First we will harvest unprecedented amounts of freely available remote sensing data to build data cubes of a variety of glacierized regions in the world. These databases will contain information on relevant mass balance predictors including snow cover area, glacier facies, albedo and glacier elevation changes. In this big data approach we will use state-of-the-art database software necessary to efficiently handle the ever increasing amount of remote sensing data. To achieve the highest possible quality for the predictors we will design novel glacier facies and snow cover classification algorithms based on deep learning. Our method using convolutional networks will be superior compared to today’s commonly used classification based on band ratio and indices as they can be trained with multi-resolution and multi-sensor data (including optical and synthetic aperture radar data) in a single classification framework. Output of the classification will be an updated multi-temporal glacier inventory. Once the regional data cubes are build we will use advanced regression techniques e.g. deep regression to extract a consistent and decade-long time series of surface mass balance for a variety of different glacier types. We will first design the methodology for glaciers in Norway, Svalbard and European Alps and other regions with different glacier characteristics and with solid training and validation data for the period 2000-2020. We will then test the transferability to glacierized regions with less base data available. Based on the regression parameters we also aim at prolonging the mass balance time series in a sensor-independent approach solely from snow cover and albedo maps.