An inventory-driven rock glacier status model (intact vs. relict) for South Tyrol, Eastern Italian Alps

• Published in: Geomorphology (Amsterdam, Netherlands) Netherlands), 2020-02, Vol.350, p.106887, Article 106887
• Main Authors: Kofler, Christian, Steger, Stefan, Mair, Volkmar, Zebisch, Marc, Comiti, Francesco, Schneiderbauer, Stefan
• Format: Article
• Language: English
• Subjects: Classification algorithms; Machine learning; Permafrost; Probabilistic modelling; Rock glacier activity; Rock glacier inventory
• ISSN: 0169-555X; 1872-695X
• DOI: 10.1016/j.geomorph.2019.106887

•Data-driven models allow to estimate the rock glacier status (intact vs. relict).•Generated classification rules assigned the status to 235 unclassified rock glaciers.•Vegetation cover and elevation were found to be the most relevant predictors.•Logistic regression performed slightly better than machine learning models.•For most of the unclassified rock glaciers all models agreed on the status. Ice presence in rock glaciers is a topic that is likely to gain importance in the future due to the expected decrease in water supply from glaciers and the increase of mass movements originating in periglacial areas. This makes it important to have at ones disposal inventories with complete information on the state of rock glaciers. This study presents a method to overcome incomplete information on the status of rock glaciers (i.e. intact vs. relict) recorded in regional scale inventories. The proposed data-driven modelling framework can be used to estimate the likelihood that rock glaciers contain frozen material. Potential predictor variables related to topography, environmental controls or the rock glacier appearance were derived from a digital terrain model (DTM), satellite data and gathered from existing data sets. An initial exploratory data analysis supported the heuristic selection of predictor variables. Three classification algorithms, namely logistic regression (GLM), support vector machine (SVM) and random forest (RF), were trained on the basis of the available information on the status of rock glaciers within the territory of South Tyrol (Eastern Italian Alps). The resulting classification rules led to assign a binary label – intact or relict – to 235 unclassified rock glaciers present in the inventory. All models were validated quantitatively on spatially-independent test samples (spatial cross validation) and achieved highly satisfactory performance scores. Hereby, the less flexible statistically-based classifier (GLM) performed slightly better than the more flexible machine learning algorithms (SVM and RF). Spatial permutation-based variable importance assessment revealed that elevation and vegetation cover (based on NDVI) were the most relevant predictors. For more than 80% of the unclassified rock glaciers, all of the three models agreed on the spatially predicted rock glacier status. Only for a minor portion (12.3%), one model differed from the remaining two.