AUTHOR=Twickler Silas , Herzfeld Ute , Trantow Thomas 

TITLE=GEOCLASS-image – a versatile machine learning environment for ice-surface classification from high-resolution image data

JOURNAL=Frontiers in Earth Science

VOLUME=Volume 13 - 2025

YEAR=2025

URL=https://www.frontiersin.org/journals/earth-science/articles/10.3389/feart.2025.1572982

DOI=10.3389/feart.2025.1572982

ISSN=2296-6463

ABSTRACT=GEOCLASS-image is an open source cyberinfrastructure (CI) for automated classification of spatial surface structures based on high-resolution image data, consisting of a data-driven and physically informed neural network (NN) system and a data analysis tool for, currently, submeter resolution satellite image data (Maxar WorldView data). The objective of this paper is to introduce GEOCLASS-image v2.0, which provides a solution for two important problems in machine learning in the geosciences: (1) Version 2.0 presents an approach for creating, exporting and sharing labeled training datasets for cryospheric classification tasks for which such datasets do not currently exist. GEOCLASS-image (v2.0) offers options for user-friendly, system-immanent application using a graphical user interface (GUI), and additionally for importing and exporting data sets to facilitate interoperability with other software, a key for advancing Open Science. (2) Combining the advantages of a purely data-driven convolutional NN and a physically driven NN, a new combined NN architecture, termed VarioNet, is derived using a weighted fusion approach that includes one or several additional blocks. The GEOCLASS-image CI, demonstrated here for classification of 11 different glacier surface types which include crevasse classes and water-based classes, extracted from Maxar WorldView1 and WorldView2 data, is expected to generalize to similar classification problems in other geoscience disciplines and any high-resolution satellite imagery.