Rock and soil classification using thermal and SAR data in deep learning models
J. Martinez (TU Delft - Architecture and the Built Environment)
Azarakhsh Rafiee – Mentor (TU Delft - Digital Technologies)
Remi J.G. Charton – Mentor (TU Delft - Applied Geology)
More Info
expand_more
Other than for strictly personal use, it is not permitted to download, forward or distribute the text or part of it, without the consent of the author(s) and/or copyright holder(s), unless the work is under an open content license such as Creative Commons.
Abstract
Fields like geology, mining, and construction have long relied on manual rock and soil classification methods, which remain time-consuming and labour-intensive. Accurate rock and soil classification is necessary for hazard assessment, urban management or resource exploration, which the rapid advances in artificial intelligence and remote sensing might facilitate. Though optical imagery has dominated remote sensing in the last decades, thermal data offers unique advantages such as sensitivity to material properties, sunlight independence or seasonal and diurnal changes responsiveness. This study explores the use of thermal imagery, along with SAR and NDVI data as complementary data, to leverage and automate the classification process with deep learning. Therefore, the aim is to develop two deep learning models capable of rock and soil segmentation with multi-modal datasets, and evaluate the performance of these models, Convolutional Neural Networks (CNNs) and Convolutional
Long Short-Term Memory networks (ConvLSTMs) across diurnal, seasonal and multi-source
scenarios as well as the effect of vegetation on the prediction results.
Results show CNN tends to have a strong overall performance, especially with thermal and SAR data. Instead, ConvLSTM excels at capturing temporal dependencies, improving the classification of most datasets and variations. The findings demonstrate the potential of thermal imagery to be used as a powerful classification tool when combined with deep learning methods for rocks and soils. They also demonstrate the slightly negative influence of vegetation on the predictability of the outputs. By combining spatial and temporal dynamics, the models offer an automated, scalable, fast and cost-effective approach to more traditional workflows. This work intends to contribute towards modernizing geological practices for more informed decision-making in geology, mining, urban planning, hazard and resource management.