Evaluating the potential of spatial and remotely sensed data for geological mapping across exposed and covered terrains in northern Sweden
(2026) In Master Thesis in Geographical Information Science GISM01 20262Department of Earth and Environmental Sciences (MGeo)
Dept of Physical Geography and Ecosystem Science
- Abstract
- Geological mapping provides essential information for understanding the geology of the Earth's crust. Recent advances in remote sensing and machine learning have increased the potential of spatial data to support lithological mapping. However, the effectiveness of such approaches may vary depending on environmental conditions, particularly where geological signals are obscured by vegetation, sediments or other surface cover.
This study evaluates the potential of multisource spatial data for lithological mapping across exposed and covered terrain conditions in the Scandinavian Caledonides of northern Sweden. Predictor variables derived from Sentinel-2 imagery, a digital elevation model and airborne gravity data were integrated within a... (More) - Geological mapping provides essential information for understanding the geology of the Earth's crust. Recent advances in remote sensing and machine learning have increased the potential of spatial data to support lithological mapping. However, the effectiveness of such approaches may vary depending on environmental conditions, particularly where geological signals are obscured by vegetation, sediments or other surface cover.
This study evaluates the potential of multisource spatial data for lithological mapping across exposed and covered terrain conditions in the Scandinavian Caledonides of northern Sweden. Predictor variables derived from Sentinel-2 imagery, a digital elevation model and airborne gravity data were integrated within a Random Forests classifier. Model performance was assessed using spatial cross-validation to reduce the influence of spatial autocorrelation and provide a more realistic classification performance.
The final full area model achieves a macro-averaged F1 score of 0.339 and an overall accuracy of 0.363. Classification performance varied considerably between lithological groups, with Metasedimentary and Felsic Igneous units achieving the highest accuracies and Carbonate and Quartz-rich units showing comparatively poor performance. The Bouguer gravity anomaly was identified as the most influential predictor, while terrain variables such as elevation, slope and topographic position index also contributed strongly to model predictions. Sentinel-2 spectral predictors exhibited comparatively lower individual importance but were still relevant to the overall classification performance.
Contrary to the initial expectation that exposed bedrock would provide substantially stronger geological signals, only relatively small performance differences were observed between exposed and covered terrain. Covered terrain achieved slightly higher values for all three overall accuracy metrics. While the achieved classification performance was quite low, the findings suggest that utilising spectral, terrain and geophysical datasets can still support lithological mapping, even where geological signals are partially obscured by surface cover. Future work could focus on alternative lithological grouping schemes, integration of additional predictor datasets such as magnetic measurements, and improving the understanding of how terrain conditions influence lithological classification performance. (Less) - Popular Abstract
- Maps are not only used for orientation but are essential tools for visualising the Earth’s geology. Geological maps are widely used in fields such as mineral exploration, environmental protection and infrastructure development. Traditionally, these maps have been created by detailed fieldwork. But what if the rocks lie beneath soil, forest or sediment or in places that are hard to reach?
This project asked whether this process can be automated. Satellite images and measurements of the Earth’s gravity cover huge areas. The general idea was to combine these datasets to train a model to recognise different rock types. This could have several benefits to the original process and would make it faster, cheaper and more accessible.
Northern... (More) - Maps are not only used for orientation but are essential tools for visualising the Earth’s geology. Geological maps are widely used in fields such as mineral exploration, environmental protection and infrastructure development. Traditionally, these maps have been created by detailed fieldwork. But what if the rocks lie beneath soil, forest or sediment or in places that are hard to reach?
This project asked whether this process can be automated. Satellite images and measurements of the Earth’s gravity cover huge areas. The general idea was to combine these datasets to train a model to recognise different rock types. This could have several benefits to the original process and would make it faster, cheaper and more accessible.
Northern Sweden provided an ideal testing area. The varied geography includes mountain ranges with exposed bedrock but also areas which are partially covered by sediments or vegetation. Satellites mostly capture information about the surface, not what lies beneath it. Therefore, the expectation was clear that the model should struggle wherever the rock is covered. The project evaluated how well a machine learning model could perform under these different conditions.
What can machine learning tell us about the rocks beneath the surface?
The results revealed some surprising findings. Interestingly, the model performed slightly better in covered areas than in areas with exposed bedrock. The results also showed which data were most useful. Gravity measurements contributed the most to the predictions, whereas satellite imagery played a smaller role on their own. Some rock types were also easier to recognise than others. Felsic igneous rocks, which are light coloured rocks such as granite, were identified more reliably than carbonates and quartz-rich rocks.
The overall result was modest and with only 36.3% of rock types correctly identified, showing that there is considerable room for improvement. But the project shows what works and what does not which is a useful starting point. Adding further data, such as magnetic measurements, and grouping rock types in different ways are promising next steps. The learning outcome is that covered terrain does not necessarily restrict this approach and the real challenge lies in telling similar rock types apart. (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/student-papers/record/9253551
- author
- D'argent, Bastian LU
- supervisor
-
- Lars Eklundh LU
- Carl Alwmark LU
- organization
- course
- GISM01 20262
- year
- 2026
- type
- H2 - Master's Degree (Two Years)
- subject
- keywords
- Geographical Information Systems, Remote Sensing, Lithological Mapping, Random Forests, Spatial Cross-Validation, Scandinavian Caledonides
- publication/series
- Master Thesis in Geographical Information Science
- report number
- 223
- language
- English
- id
- 9253551
- date added to LUP
- 2026-11-16 09:13:42
- date last changed
- 2026-11-16 09:13:42
@misc{9253551,
abstract = {{Geological mapping provides essential information for understanding the geology of the Earth's crust. Recent advances in remote sensing and machine learning have increased the potential of spatial data to support lithological mapping. However, the effectiveness of such approaches may vary depending on environmental conditions, particularly where geological signals are obscured by vegetation, sediments or other surface cover.
This study evaluates the potential of multisource spatial data for lithological mapping across exposed and covered terrain conditions in the Scandinavian Caledonides of northern Sweden. Predictor variables derived from Sentinel-2 imagery, a digital elevation model and airborne gravity data were integrated within a Random Forests classifier. Model performance was assessed using spatial cross-validation to reduce the influence of spatial autocorrelation and provide a more realistic classification performance.
The final full area model achieves a macro-averaged F1 score of 0.339 and an overall accuracy of 0.363. Classification performance varied considerably between lithological groups, with Metasedimentary and Felsic Igneous units achieving the highest accuracies and Carbonate and Quartz-rich units showing comparatively poor performance. The Bouguer gravity anomaly was identified as the most influential predictor, while terrain variables such as elevation, slope and topographic position index also contributed strongly to model predictions. Sentinel-2 spectral predictors exhibited comparatively lower individual importance but were still relevant to the overall classification performance.
Contrary to the initial expectation that exposed bedrock would provide substantially stronger geological signals, only relatively small performance differences were observed between exposed and covered terrain. Covered terrain achieved slightly higher values for all three overall accuracy metrics. While the achieved classification performance was quite low, the findings suggest that utilising spectral, terrain and geophysical datasets can still support lithological mapping, even where geological signals are partially obscured by surface cover. Future work could focus on alternative lithological grouping schemes, integration of additional predictor datasets such as magnetic measurements, and improving the understanding of how terrain conditions influence lithological classification performance.}},
author = {{D'argent, Bastian}},
language = {{eng}},
note = {{Student Paper}},
series = {{Master Thesis in Geographical Information Science}},
title = {{Evaluating the potential of spatial and remotely sensed data for geological mapping across exposed and covered terrains in northern Sweden}},
year = {{2026}},
}