Abstract:
Mudflows are some of the most destructive geological phenomena, and their prediction is
challenging due to their complexity and the strong nonlinear relationships between the various factors that
contribute to their formation. Traditional modeling methods have limitations in their ability to interpret
and account for the complex interactions between different factors, and this lead to the need for the
development of more advanced approaches.
Aim. The study aims to develop and test a sigma-pi neural network architecture for mudflow clustering
based on morphometric and genetic characteristics as well as to identify the key factors and their
combinations that contribute to the formation of different mudflow types.
Materials and methods. Cadastral data on mudflows in the southern European part of Russia is
used as the initial data. A sigma-pi neural network capable of accounting for both linear features and
their second-order interactions is employed for analysis. A silhouette coefficient is used to determine
the number of clusters. The results are compared with those obtained using Kohonen's self-organizing
maps (SOM).
Results. The model identified three stable clusters corresponding to mud, rock, and mud-rock types of
mudflows. Analysis of the significance of features has revealed that the basin area, channel slope, and
maximum sediment volume make the greatest contributions to cluster formation, as well as their various
pairwise combinations. Comparison with the SOM (self-organizing map) confirmed the improved
interpretability of the proposed model and its ability to identify hidden, nonlinear relationships.
Conclusions. The use of sigma-pi neural networks not only improves the accuracy of mudflow
clustering, but also ensures the interpretability of the results by analyzing the significance of features and
their combinations. This approach is promising for engineering geology and can be used in geoecological
monitoring systems and forecasting of hazardous processes.
Fulltext:
PDF file (719 kB)