THE IMPACT OF RAINFALL ON LANDSLIDE DYNAMICS: QUANTITATIVE ANALYSIS ON MOUNTAINOUS AREA IN NORTHERN ITALY USING MACHINE LEARNING ASSISTED APPROACHES

Landslides are a very common type of disaster. It happens in every state of the U.S and is defined as the movement of the mass of rock, debris, or earth down a slope. Debris flows, sometimes referred to as mudslides, mudflows, lahars, or debris avalanches, are common types of fast-moving landslides(Lynn et al. 1997). When a landslide takes place, it could bring down a large volume of mass which is enough to bury houses and buildings. Therefore, preventing or reducing the life and economic losses comes from landslides is an indispensable task for engineers. There are multiple factors that can cause landslides, including water level, stream erosion, changes in ground water. This paper will focus on examining the associations between rainfall and landslide displacement. The goal will be performing spatial and temporal estimation of landslide displacement in “Valle Febrraro” by using data during the slow-motion stage. We will analyze how we can possibly predict landslide dynamics in Valle Febrraro using precipitation data. We will adopt the concept of correlation coefficient to pinpoint at places where landslide dynamics might be sensitive to precipitation. In fact, we have identified dots with correlation coefficients close to negative 1 through calculation. Those dots are clustered in the lower left half of the selected region. After examining correlation coefficient for every single dot in the chosen space, we adopted kriging as a spatial estimation technique to predict the value of correlation coefficient at every place in the entire chosen space. Results indicate that for the regions where dots with correlation coefficient close to -1 are clustering have values close to -1 whereas those far away have values higher than -1. Besides kriging for the values of correlation coefficient for the entire chosen space, we also performed kriging for the displacement values everywhere inside the same space. We have seen that the results vary from one instant to another. The approach of kriging provides us with what are areas with high displacement values at each of the timestamp and therefore will provide useful insights for future landslide prediction. For temporal estimation, we use regression model to estimate landslide displacement values at one year. Besides, we will also modify the regression model in different ways to see how much better or worse the model will perform. In addition, we will also apply time series prediction technique auto-regression/auto-regressive models, make modifications to it and compare the results. The goal after creating and comparing these models is to perform some related error analysis. Although the results turn to be well in general, we have not noticed any obvious increase in the displacement when we increase the precipitation to several times of its original values when we are trying to optimize the model. Other ways for optimization do exist as we have found out that adjusting the model parameters or performing model for multiple times with each time predicting fewer values do help increase the accuracy.