Abstract

The current study aims to investigate landslide susceptibility in Western and Central Greece and to enrich our existing knowledge in slow-moving deformation phenomena, mainly triggered by prolonged or excess precipitation events. For this purpose, SAR interferometry and AI are implemented on Earth Observation data.
Landslide susceptibility mapping was performed in a broad area within the most landslide prone geotectonic zones of Greece, extending from Crete Island up to the Greek - Albanian borders. These are the Ionios, Gavrovo and Pindos geotectonic zones. The variety of geological formations and morphological features, crucial infrastructure and the plethora of settlements provide a great variety of conditions and an exceptionally large landslide inventory.
By exploiting a national scale inventory of Line of Sight surface displacements in Greece, the so-called InSAR Greece project, more than 3000 landslides were detected in Western Greece. Topographical, geological, meteorological, hydrological parameters and vegetation, were introduced to the model, as landslide causative factors.
The prediction problem was approached as a binary classification problem defining the classes landslide, no landslide for each landslide point and was solved training robust classic machine learning algorithms using the aforementioned dataset. The annotation of the dataset was performed by experts in the field. A strict ML methodology was employed for training, tuning and testing the machine learning, comprising of 5-fold cross validation, hyperparameter tuning, feature ranking and selection and strict train validation dataset split to avoid spatial autocorrelation. XGBoost algorithm showed the best performance among the traditional ML algorithms, thus it was preferred for producing susceptibility maps.
The exploitation of big volumes of EO data and AI, in landslide susceptibility mapping, can provide a valuable tool in risk reduction. Also, validation of susceptibility mapping with ground truth investigations, provides an additional advantage in establishing a landslide susceptibility system and in the adoption of mitigation measures.