Abstract

We present results of the application of structured geological maps to the prediction of landslide occurrence based on statistical mechanics. The statistical mechanics approach shows that the probability of landslide occurrence is derived from the energy associated with geological and topographical variables. Landslides involve the irreversible conversion of potential energy into kinetic energy. The activation energy required for landslide rupture can be expressed in terms of the strength of rocks and strata, and is influenced by geological variables. Because the energy cannot be explicitly expressed as a function of the variables, a neural network algorithm was used to estimate the probability based on training datasets of these variables. Three geological variables were obtained from structured geological maps: formation age, petrological classification, and formation environment. In contrast, traditional geological maps, which are typically unstructured, contain only a single geological variable. We compared the performance of landslide predictions derived from structured versus unstructured geological maps.　Receiver operating characteristic (ROC) and precision–recall (PR) analyses show that structured geological maps outperformed their unstructured ones. These results indicate that the three geological variables incorporated in structured geological maps capture the geological diversity more effectively than the single variable used in unstructured maps. We also evaluated the performance of the neural network trained on two different spatial extents: (1) the Sasebo–Imari area in northwestern Kyushu, and (2) the entire northern Kyushu region, southwest Japan, using the Sasebo–Imari area as a common validation region. Neural network modeling assumes the training data are drawn from a probability distribution defined over explanatory and objective variables. The data-generating probability distribution is, in principle, assumed to be spatially invariant, irrespective of the analysis area. This markedly degraded performance indicates that important local explanatory variables may have been omitted or insufficiently represented in the broader-scale training data. Our results suggest that the local tectonic evolution, low-P/T metamorphic overprint induced by magma intrusion, and hydrothermal alteration in the shallow volcanic fields are likely candidates for the omitted important explanatory variables.