Performance of deep learning algorithms on obtaining geotechnical properties of unconsolidated material to improve input data for regional landslide susceptibility studies

The presented study implements existing deep learning (DL) algorithms, an artificial intelligence approach, to extract geotechnical properties about unconsolidated material from photographs. The ultimate goal of this approach lies in facilitating, aiding and simplifying the collection of often missing data about unconsolidated bedrock cover relevant in regional landslide susceptibility studies. Current research aims at answering, if existing DL algorithms (e. g. Buscombes (2020) Sedinet algorithm), developed for granular, often well-sorted sediments, can also perform well with poorly-sorted sediments. It also inquires, if, which and how well geotechnical properties, as described in soil classification standards like ISO 14688-1:2017-12 (EU) and ASTM D2487-17e1 (USA), can be directly or indirectly obtained through DL analysis of photographs. The study approaches these questions by initially building a DL model based on several thousand photographs of 240 samples of unconsolidated material plus their several hundred laboratory sieve residue samples. In a previous project, the 240 samples of mostly alluvial, colluvial, eolian and glacial sediments had been collected from different geological environments within the state of Styria, Austria. Grain size distribution (GSD) and other soil classification parameters, obtained through field and laboratory testing, exist for these samples and have been provided as courtesy by the Land Steiermark (State of Styria). In the current study this knowledge about geotechnical properties of the samples allows attribution of this information to each of the several thousand photographs, which were taken with three different cameras under controlled conditions. The DL model uses several hundred of these photographs with their associated attributes as training and test data to build a prediction model. The validation of thus derived model in regard to its performance is achieved with selected photographs, not yet used in the training and testing. Results of this approach allow a discussion about applicability, emerging limitations and possible improvements in regard to predicting geotechnical parameters, particularly GSD, for unconsolidated material using existing DL algorithms. As a consequence the results and drawn conclusions also warrant an outlook and contemplation on how, if and in what way the method can aid and simplify field mapping and the collection of relevant input data for regional landslide susceptibility studies.