Share Email Print

Proceedings Paper

A multitemporal probabilistic error correction approach to SVM classification of alpine glacier exploiting sentinel-1 images (Conference Presentation)
Author(s): Mattia Callegari; Carlo Marin; Claudia Notarnicola; Luca Carturan; Federico Covi; Stephan Galos; Roberto Seppi

Paper Abstract

In mountain regions and their forelands, glaciers are key source of melt water during the middle and late ablation season, when most of the winter snow has already melted. Furthermore, alpine glaciers are recognized as sensitive indicators of climatic fluctuations. Monitoring glacier extent changes and glacier surface characteristics (i.e. snow, firn and bare ice coverage) is therefore important for both hydrological applications and climate change studies. Satellite remote sensing data have been widely employed for glacier surface classification. Many approaches exploit optical data, such as from Landsat. Despite the intuitive visual interpretation of optical images and the demonstrated capability to discriminate glacial surface thanks to the combination of different bands, one of the main disadvantages of available high-resolution optical sensors is their dependence on cloud conditions and low revisit time frequency. Therefore, operational monitoring strategies relying only on optical data have serious limitations. Since SAR data are insensitive to clouds, they are potentially a valid alternative to optical data for glacier monitoring. Compared to past SAR missions, the new Sentinel-1 mission provides much higher revisit time frequency (two acquisitions each 12 days) over the entire European Alps, and this number will be doubled once the Sentinel1-b will be in orbit (April 2016). In this work we present a method for glacier surface classification by exploiting dual polarimetric Sentinel-1 data. The method consists of a supervised approach based on Support Vector Machine (SVM). In addition to the VV and VH signals, we tested the contribution of local incidence angle, extracted from a digital elevation model and orbital information, as auxiliary input feature in order to account for the topographic effects. By exploiting impossible temporal transition between different classes (e.g. if at a given date one pixel is classified as rock it cannot be classified as glacier ice in a following date) we here propose an innovative post classification correction based on SVM classification probabilities. Optical data, i.e. Landsat-8 and Sentinel-2, have been employed, when available, for training sample collection. Detailed field observations from two glaciers located in the Ortles-Cevedale massif (Eastern Italian Alps) have been employed for validation.

Paper Details

Date Published: 8 December 2016
PDF: 1 pages
Proc. SPIE 10003, SAR Image Analysis, Modeling, and Techniques XVI, 100030F (8 December 2016); doi: 10.1117/12.2242416
Show Author Affiliations
Mattia Callegari, EURAC (Italy)
Carlo Marin, EURAC (Italy)
Claudia Notarnicola, EURAC (Italy)
Luca Carturan, Univ. degli Studi di Padova (Italy)
Federico Covi, Univ. of Innsbruck (Austria)
Stephan Galos, Univ. of Innsbruck (Austria)
Roberto Seppi, Univ. degli Studi di Pavia (Italy)

Published in SPIE Proceedings Vol. 10003:
SAR Image Analysis, Modeling, and Techniques XVI
Claudia Notarnicola; Simonetta Paloscia; Nazzareno Pierdicca; Edward Mitchard, Editor(s)

© SPIE. Terms of Use
Back to Top
Sign in to read the full article
Create a free SPIE account to get access to
premium articles and original research
Forgot your username?