Unit:
Κατεύθυνση Βιοπληροφορική-Επιστήμη Βιοϊατρικών ΔεδομένωνΠληροφορική
Author:
Tsamkosoglou Konstantinos
Supervisors info:
Κουτρούμπας Κωνσταντίνος, Διευθυντής Ερευνών, ΙΑΑΔΕΤ, ΕΑΑ
Πικράκης Άγγελος, Επίκουρος Καθηγητής, Τμ. Πληρ/κής, ΠΑΠΕΙ
Συκιώτη Όλγα, Κύρια Ερευνήτρια, ΙΑΑΔΕΤ, ΕΑΑ
Original Title:
Lithological mapping of Koutala island (Lavrio, Attiki) using machine learning methods on multispectral data
Translated title:
Lithological mapping of Koutala island (Lavrio, Attiki) using machine learning methods on multispectral data
Summary:
In recent decades, there has been a rapid advancement in the utilization of Earth Observation (EO) data in geology, driven by a growing interest in its application to identify potential sites associated with hydrothermal alteration and ore deposits. This development has garnered increasing attention due to its potential for substantial time and cost savings. In the present study, the target of interest is a small island called Koutala near the city of Lavrion (Attiki, Greece) and the aim is (a) to identify granitoid intrusions and schist formations on its surface and (b) to detect the associated alteration minerals. To this end, two high-resolution satellite datasets depicting the area of interest, taken from the Sentinel-2 and WorldView-3 missions, are utilized (the data sets differ in their spatial and spectral characteristics). Two different machine learning methods, namely clustering and spectral unmixing, were applied to extract geological information from the island.
Clustering was applied to both datasets to delineate regions with similar spectral signatures, aiming to identify granitoid and schist formations, as is referred on previous research insights [1]. In this framework, a novel clustering algorithm named SHC was introduced. SHC has been tailored especially for multispectral data. It takes advantage of the derivative of each pixel’s spectral signature, and outperforms traditional off-the-shelf clustering algorithms, like K-means and hierarchical methods. The SHC algorithm demonstrated improved accuracy in identifying granitoid intrusion areas, especially in the challenging lower spatial resolution context of the Sentinel-2 dataset and in general yield to more homogeneous clusters (in terms of spectral characteristics).
Additionally, various linear spectral unmixing methods were explored in the Sentinel-2 dataset, taking into account its larger number of spectral bands and spectral positions compared to WorldView-3 data, to detect the associated alteration minerals on the surface of the island. Despite the dataset's relatively low spatial resolution for this type of study, alteration minerals with high probability of presence (having as reference previous search insights) were accurately identified by most algorithms.
Main subject category:
Technology - Computer science
Keywords:
Clustering, Spectral-unmixing, Sentinel-2, WorlView-3 VNIR
