Automatic regional identification of active and inactive landslides using satellite image analysis

Over the past decades, landslides have significantly affected extensive areas worldwide due to changing environmental conditions and human activities, causing major problems in the built environment and infrastructure and resulting in the loss of human lives and significant financial damages. The island of Cyprus and especially its southwestern part (which constitutes the study area) have experienced the severe impact of landslides due to the unfavorable geological/geotechnical conditions and mountainous geomorphology. According to the data obtained from the Geological Survey Department of Cyprus (GSD), 1842 landslides (active and inactive) of various types have been identified in an area covering 40% (546km²) of the Paphos District (3.4 landslides per km²).

Knowledge of the location and extent of existing landslides plays crucial role in the landslide susceptibility and hazard assessment. The primary aim of this research is to develop an algorithm for the automatic detection of landslides at regional scale. This is achieved through application of image recognition technology utilizing the cascade method on the hillshade of a region as produced by ArcGIS. The database of recorded landslides of the GSD was split in a algorithm training dataset and a validation dataset. The study also explores the effect of the resolution of terrain data, expressed by the size of the grid cells. To comprehensively assess landslides, the morphology is classified into three types: active, dormant, and relict. The use of hillshade instead of a raster image of the elevation map was chosen because the latter usually results in relatively minor color variations between adjacent pixels, thus hindering the most striking geomorphological features of landslides, which are the main scarp and the enveloping streams.

The results obtained suggest that a hillshade produced using a high-resolution Digital Elevation Model (DEM), i.e. based on elevation contour interval of 1m and a cell size 1 x 1 m (obtained from the Department of Land and Surveys of the Republic of Cyprus), yields better results for landslides with gentle geomorphology (relict). Nonetheless, analysis based on such a high-resolution DEM requires substantial computational resources and time. On the contrary, landslides associated with steeper geomorphologies (active) exhibited optimal performance with a cell size of 2 x 2 m, achieving success rates (80%), for DEMs based on contour intervals of 1m and 5m. In this case, the computational time is significantly reduced.  Depending on the specific landslide types investigated in a particular area, the appropriate processing model can be selected, ultimately leading to significant time savings.

This research was funded by the European Commission (Marie Sklodowska-Curie Actions, Hybland-Society and Enterprise panel, Project No.: 101027880).