Information-theoretic modeling of landslide susceptibility using mostly categorical spatial data
Abstract. Landslide susceptibility mapping with machine learning should make full use of inherently informative categorical layers such as lithology, soil type, and land use, yet these variables usually enter models in a diminished, encoded form. Reconstructability Analysis (RA), an information-theoretic method based on Shannon entropy, uses categorical data directly and produces transparent, interpretable models. We applied RA to two contiguous Oregon Coast Range (USA) watersheds (3,500 km²) rated Very High Risk for landslides by the Federal Emergency Management Agency, using ~9,000 landslide deposits mapped mostly from LiDAR by the Oregon Department of Geology and Mineral Industries (DOGAMI), with DOGAMI lithology, Soil Survey, and National Land Cover Database layers; terrain variables were discretized into optimized bins. Variable importance analysis led to stratification into volcanic and sedimentary lithologies, with landslide rates of 44 % and 60 %. The best RA models reached 75 % (volcanic) and 71 % (sedimentary) test accuracy, nearly as good as Random Forest, while remaining interpretable in terms of named soil, landform, and rock units. Their probability tables give landslide likelihood for specific combinations of landform, geology, and elevation; in the sedimentary stratum, two andic soil subgroups emerged as most landslide-prone, consistent with high water retention. The tables are provided as a supplement.