Machine learning and object-based image analysis for landside mapping using UAV-derived data

We are living in an increasingly changing environment. The last half-century was a turning point in scientific conception due to the radical change in the scientific approach to the analysis of natural-induced hazards. Every year, geohazards are a major cause of loss of life and property. More specifically, mass movements constitute one of the most widespread geohazards worldwide, often with limited reporting, responsible for remarkable socioeconomic impacts. For example, in the US, 3.5 billion dollars of damage and approximately 35 deaths are reported per year due to landslides and rockfalls. Various triggering factors such as rainfall, earthquakes, and anthropogenic activities, accompanied by the intrinsic factors of the slope, soil characteristics, and geomorphic processes contribute to slope failures, sometimes creating devastating problems. In addition, forest fires and droughts substantially amplify soil erosion due to vegetation loss. Landslide inventories constitute the foundation for hazard and risk assessments of catastrophic events. High-resolution satellite imagery and advanced spatial analysis techniques allow the development of reliable landslide inventory datasets, whereas field investigations can serve for validation mapping. Unmanned or Uncrewed Aerial Vehicles (UAVs) have emerged as indispensable tools for data collection because they offer ultra-high-resolution datasets with high repeatability and precision. Such platforms are particularly effective for the 2D and 3D documentation of geomorphological changes after extreme events. Manifold landslide information can be extracted on a centimeter scale from UAV platforms equipped with various sensors. Consequently, a range of undiscovered landslide patterns and related information become distinguishable and can be further incorporated as advanced semantic knowledge in the characterization phase. The reliability of the final landslide output depends mainly on the quality of the datasets, the scale, and the selection of the appropriate mapping approach. In particular, landslide assessments can be performed efficiently when integrated with Geographic Information System and Remote Sensing techniques. Over the past few years, landslide mapping has seen rapid development with techniques that use a combination of different Earth Observation data. In the past decade, object-based approaches have been increasingly implemented for landslide mapping because of their effectiveness in handling and analyzing multivariate high-resolution data and combining expert knowledge with machine learning algorithms.