By using a number of recorded soil slips and their relevant parameters, it is possible to infer dependencies between the parameters and the danger level of a particular slope. These dependencies can then be employed for predicting landslide hazard. To accomplish this goal in practice, the entire study area is divided into cells (5 x 5 m), each represented by a vector of measurable parameters. In particular, the following characteristics are considered: slope gradient, slope orientation, geological conditions, vegetation, precipitation, etc.. In total, there are approximately 25K registered cells (data were collected during the project "Vegetationswirkungen und Rutschungen" of the Swiss Federal Institute for Forest, Snow and Landscape Research, of two diploma theses of the Chair of Applied Geology, and during field trips in the Sachsler Mountains/Switzerland), of which some 5K are marked as landslides and the rest is landslides-free.

To issue predictions of the above kind, we trained a statistic classifier to map parameter vector of a specific cell in a binary danger/no-danger response. It is important to note that many of the employed parameters come from non-metric definition domains (e.g. vegetation type). Therefore they must be encoded as groups of binary subparameters by means of a "1 out of N" algorithm. With this altered representation employed, the parameter vectors used in the baseline experiments counted about 60 elements.

In order to boost the significance of the classification experiments, we availed ourselves of the Leave-One-Out (LOO) methodology and created several training and test corpora out of the original data set. Two kinds of classifiers were tested: Linear Discriminant Analysis (LDA) classifier and Support Vector Machines. While the LDA classifier was generally much faster (maximum of 30 minutes per experiment), the SVM classification resulted in higher prediction rates.

Another substantial improvement came from taking into account the direct neighborhood of the cells. The brute-force solution is to concatenate the parameter vectors of all cells pertaining to the neighborhood of the cell in question into one long parameter vector that will represent it. Also, derived measures such as the mean and variance of each parameter over the entire neighborhood can be added to this vector. The optimal neighborhood geometry and optimal choice of these derived measures are subject to validation.

The best results attained in the experiments so far were obtained by using the LDA to select some 960 classification parameters out of the total of 3200 available, and feeding them into the SVM classifier to make a final prediction. When discriminating between two classes (landslide/no-landslide) this strategy brought classification rates of more than 82%.

For the moment a three-class classification problem where the landslide origins and landslide bodies are regarded as different classes is our next objective in this research field.