Forecasting the surficial displacements of a landslide triggered by snow melting basing on LSTM and image processing algorithms

The majority of landslide-prone areas spread in mountainous areas with abundant rainfall. However, when high altitudes make areas prone to significant snowfall, the amount of such snowfall, as well as environmental temperature and humidity, should be taken into account to determine its effect on the condition of landslide stability. To pursue this aim, the present study focuses on the quantification of snow accumulation on the slope through approaches based on image analysis and on the prediction of surface displacements of the slope using a two-steps LSTM (Long short-term memory) algorithm. The main LSTM algorithm aims at forecasting the landslide displacement in the future 12 hours using as input the past 5 days data of rainfall, snowfall and movements of the slope, plus the weather prediction of the next day. The necessity of estimation of the trend of the snow condition makes it necessary to implement a secondary LSTM algorithm for estimating if the snow coverage is going to accumulate or melt in next 12 hours, again basing on the past 5 days environmental measurements (temperature and humidity) and a forecast of the future condition of the site. Both the algorithms are trained basing on the historical measurements of temperature, humidity, rainfall, snowfall and landslide displacement. The main code also includes a training based on the surficial movements of the slope measured by a topographical monitoring system. Within this model, the presence and the trend of the snow is evaluated by means of some image-processing algorithms aiming at evaluating the cover square percentage of white content in the RGB image, filtering out noises and false signals. The presented procedure is applied to the case of the Sant’Andrea landslide, located in Perarolo di Cadore (North Italy, Province of Belluno), whose bedrock is composed by dolomitic lithology and folded layers rich in anhydrides and gypsum easily erodible by water infiltration in the subsoil. The two-steps LSTM model implementation achieves the forecasting of the landslide displacements, focusing in particular on the effects of snow melting in the stability condition of the slope.