TY  - JOUR
A1  - Nodjoumi, Giacomo
A1  - Pozzobon, Riccardo
A1  - Sauro, Francesco
A1  - Rossi, Angelo Pio
T1  - DeepLandforms: A Deep Learning Computer Vision Toolset Applied to a Prime Use Case for Mapping Planetary Skylights
Y1  - 2023-01-04
VL  - 10
IS  - 1
JF  - Earth and Space Science
DO  - 10.1029/2022EA002278
PB  - 
N2  - Thematic map creation is a meticulous process that requires several steps to be accomplished regardless of the type of map to be produced, from data collection, through data exploitation and map publication in print, image, and GIS format. Examples are geolithological, and geomorphological maps in which most of the highest time‐consuming tasks are those related to the discretization of single objects. Introducing also interpretative biases because of the different experience of the mappers in identifying a set of unique characteristics that describe those objects. In this setting, Deep Learning Computer Vision techniques could play a key role but lack the availability of a complete set of tools specific for planetary mapping. The aim of this work is to develop a comprehensive set of ready‐to‐use tools for landforms mapping based on validated Deep Learning methodologies and open‐source libraries. We present DeepLandforms, the first pre‐release of a toolset for landform mapping using Deep Learning that includes all the components for data set preparation, model training, monitoring, and inference. In DeepLandforms, users have full access to the workflow and control over all the processes involved, granting complete control and customization capabilities. In order to validate the applicability of our tool, in this work we present the results achieved using DeepLandforms in the science case of mapping sinkhole‐like landforms on Mars, as a first example that can lead us into multiple and diverse future applications.
N2  - Plain Language Summary: The creation of maps is a complex set of several tasks that, regardless of the type of map, are often very time‐consuming. For instance, all the occurrences of a specific object, natural or man‐made in a defined area, need to be identified, drawn and classified manually. Mapping large objects in small areas is an easy task but may be unmanageable in cases such as small landforms on the entire surface of a planet. Nowadays, especially on Earth, researchers and professionals take advantages of highly specialized software based on a technique called Deep Learning. Such software are almost never free nor ready‐to‐use and often requires higher knowledge in computer programming languages. In this work, we present the first pre‐release of a novel open‐source computer software, nearly ready‐to‐use, that provides all the instruments for approaching Deep Learning for automatic landforms mapping. We present also the results obtained by trying this software using data of Mars's surface to map sinkhole‐like landforms.
N2  - Key Points: Instance Segmentation methodology is used to map landforms obtaining vectorial data in geopackage file format. A newly developed composite toolset to perform image pre‐processing, data labeling, model training and inference tasks is presented. The results of a prime case of mapping pit and skylights on Mars surface are showed.
UR  - http://resolver.sub.uni-goettingen.de/purl?gldocs-11858/11339
ER  -