Landslide recognition by deep learning of non-standard multi-source images

Abstract

<p>The recent development of mobile surveying platforms and crowd-sourced information has produced a huge amount of non-validated data <br>that are now available for research. In the field of landscape analysis, with particular reference to geomorphology and engineering geology, images generated by autonomous platforms (such as UAVs, ground-based acquisition systems, satellite sensors) and pictures obtained from web data-mining can be easily gathered and contribute to the fast surge in the amount of non-organised information that engulf data storage facilities. The high potential impact of such methods, however, may be severely impacted by the need of a massive amount of Human Intelligent Tasks (HIT), which is necessary to filter and classify the data, whatever the final purpose.<br>In landslide hazard analysis, both UAV-surveys and the gathering of crowd-sourced information generate big-data that would require HITs before becoming usable in early warning, vulnerability assessment, residual risk estimation, model parametrisation and mapping. Very often, this an important limitation to the real-world applications that are actually feasible with the support of such systems. Examples of such HITs are the intelligent guidance of drones, the classification of fake news, the validation of post-disaster information.<br>Computer vision can be of great help in fostering the autonomous capability of intelligent systems to complement, or completely substitute, HITs. Image and object recognition are at the forefront of this research field. They are based on a number of computer-aided methods that rely on different degrees of interaction with the user, ranging from semi-automated object-based detection to deep learning by neural networks. <br>In this work, we present a new set of convolutional neural networks specifically designed for the automated recognition of landslides and mass movements in non-standard pictures that can be used for supporting UAV automated guidance and data-mining filtering. The deep learning has been accomplished by resorting to transfer learning of some of the top-performers CNNs available in the literature. Results show that the deep learning machines, calibrated on a relevant dataset of validated images of landforms, are able to supply reliable predictions with computational time and resource requirements compatible with most of the UAV platforms and web data-mining applications for landslide hazard studies.</p><p> </p>