The large-scale landslide susceptibility assessment (LSA) is an important tool for reducing landslide risk through the application of resulting maps in spatial and urban planning. The existing literature more often deals with LSA modelling techniques, and the scientific research very rarely focuses on acquiring relevant thematic and landslide data, necessary to achieve reliable results. Therefore, the paper focuses on the crucial step of classifying continuous landslide conditioning factors for susceptibility modelling by presenting an innovative comprehensive analysis that resulted in 54 landslide susceptibility models to test 11 classification criteria (scenarios which vary from stretched values, partially stretched classes, heuristic approach, classification based on studentized contrast and landslide presence, and commonly used classification criteria, such as natural neighbour, quantiles and geometrical intervals) in combination with 5 statistical methods. The large-scale landslide susceptibility models were derived for small and shallow landslides in the pilot area (21 km²) located in the City of Zagreb (Croatia), which occur mainly in soils and soft rocks. Some of the novelties in LSA are the following: scenarios using stretched landslide conditioning factor values or classification with more than 10 classes prove more reliable; certain statistical methods are more sensitive to the landslide conditioning factor classification criteria than others; all the tested machine learning methods give the best landslide susceptibility model performance using continuous stretched landslide conditioning factors derived from high-resolution input data. The research highlights the importance of qualitative assessments, alongside commonly used quantitative metrics, to verify spatial accuracy and to test the applicability of derived landslide susceptibility maps for spatial planning purposes.

The objective of the study is to show that landslide conditioning factors derived from different source data give significantly different relative influences on the weight factors derived with statistical models for landslide susceptibility modelling and risk analysis. The analysis of the input data for large-scale landslide hazard assessment was performed on a study area (20.2 km²) in Hrvatsko Zagorje (Croatia, Europe), an area highly susceptible to sliding with limited geoinformation data, including landslide data. The main advantage of remote sensing technique (i.e., LiDAR, Light Detection and Ranging) data and orthophoto images is that they enable 3D surface models with high precision and spatial resolution that can be used for deriving all input data needed for landslide hazard assessment. The visual interpretation of LiDAR DTM (Digital Terrain Model) morphometric derivatives resulted in a detailed and complete landslide inventory map, which consists of 912 identified and mapped landslides, ranging in size from 3.3 to 13,779 m². This inventory was used for quantitative analysis of 16 input data layers from 11 different sources to analyse landslide presence in factor classes and thus comparing landslide conditioning factors from available small-scale data with high-resolution LiDAR data and orthophoto images, pointing out the negative influence of small-scale source data. Therefore, it can be concluded that small-scale landslide factor maps derived from publicly available sources should not be used for large-scale analyses because they will result in incorrect assumptions about conditioning factors compared with LiDAR DTM derivative factor maps. Furthermore, high-resolution LiDAR DTM and orthophoto images are optimal input data because they enable derivation of the most commonly used landslide conditioning factors for susceptibility modelling and detailed datasets about elements at risk (i.e., buildings and traffic infrastructure data layers).

A preliminary landslide susceptibility analysis on a regional scale of 1:100 000 using bivariate statistics was conducted for the City of Karlovac. The City administration compiled landslide inventory used in the analysis based on recorded landslides from 2014 to 2019 that caused significant damage to buildings or infrastructures. Analyses included 17 geofactors relevant to landslide occurrence and classified them into four groups: geomorphological (elevation, slope gradient, slope orientation, terrain curvature, terrain roughness), geological (lithology-rock type, proximity to geological contacts, proximity to faults), hydrological (proximity to drainage network, proximity to springs, proximity to temporary, permanent and to all streams, topographic wetness) and anthropogenic (proximity to traffic infrastructure, land cover using two classifications). Five scenarios were defined using a different combination of geofactors weighted by the Weights-of Evidence (WoE) method, resulting in five different landslide susceptibility maps. The best landslide susceptibility map was selected upon the results of a ROC curve analysis, which was used to obtain success and prediction rates of each scenario. The novelty in the presented research is that a limited amount of thematic data and an incomplete landslide inventory map allows for the production of a preliminary landslide susceptibility map for usage in spatial planning. Also, this study provides a discussion regarding the used method, geofactors, defined scenarios and reliability of the results. The final preliminary landslide susceptibility map was derived using ten geofactors, which satisfied the pairwise CI test, and it is classified in four zones: low landslide susceptibility (57.05% of the area), medium landslide susceptibility (20.63% of the area), high landslide susceptibility (13.28% of the area), and very high landslide susceptibility (9.03% of the area), and has a success rate of 94% and a prediction rate of 93% making it a highly accurate source of preliminary information for the study area.

This paper focuses on large-scale landslide susceptibility modelling in NW Croatia. The objective of this research was to provide new insight into stable and unstable area sampling strategies on a representative inventory of small and shallow landslides mainly occurring in soil and soft rock. Four strategies were tested for stable area sampling (random points, stable area polygon, stable polygon buffering and stable area centroid) in combination with four strategies for unstable area sampling (landslide polygon, smoothing digital terrain model derived landslide conditioning factors, polygon buffering and landslide centroid), resulting in eight sampling scenarios. Using Logistic Regression, Neural Network, Random Forest and Support Vector Machine algorithm, 32 models were derived and analysed. The main conclusions reveal that polygon sampling of unstable areas is an imperative in large-scale modelling, as well as that subjective and/or biased stable area sampling leads to misleading models. Moreover, Random Forest and Neural Network proved to be more favourable methods (0.804 and 0.805 AUC, respectively), but also showed extreme sensitivity to the tested sampling strategies. In the comprehensive comparison, the advantages and disadvantages of 32 derived models were analysed through quantitative and qualitative parameters to highlight their application to large-scale landslide zonation. The results yielded by this research are beneficial to the susceptibility modelling step in large-scale landslide susceptibility assessments as they enable the derivation of more reliable zonation maps applicable to spatial and urban planning systems.

The study presents analyses of input data impact on the quality of the landslide susceptibility large-scale maps. For comparison, two input data sets were used to produce two landslide susceptibility maps. The first input data set included free-available, small-scale data with low spatial accuracy, while the second set included high-resolution remote sensing data. The same nine types of landslide causal factors were derived and used for susceptibility analyses. Furthermore, LiDAR-based landslide inventory and bivariate statistical method, i.e. Information Value method, were used for susceptibility modelling. The resulting landslide susceptibility maps were compared with ROC curves. Success and prediction rates showed that the landslide susceptibility model based on causal factors derived from high-resolution remote sensing data is approximately 10% more accurate than the model based on causal factors derived from small-scale input data. Furthermore, based on the conducted research, it can be concluded that susceptibility modelling based on small-scale data and LiDAR-based inventories enables reliable landslide susceptibility assessments at the regional level.

The large scale landslide susceptibility assessment (LSA) is an important tool for reducing landslide risk through the application of resulting maps in spatial and urban planning. The existing literature more often deals with LSA modelling techniques and the scientific research very rarely focuses on acquiring relevant thematic and landslide data, necessary to achieve reliable results. Therefore, the paper focuses on the crucial step of classifying continuous landslide conditioning factors for susceptibility modelling by presenting an innovative comprehensive analysis that resulted in 54 landslide susceptibility models to test 11 classification criteria (scenarios which vary from stretched values, partially stretched classes, heuristic approach, classification based on studentized contrast and landslide presence, and commonly used classification criteria, such as Natural Neighbor, Quantiles and Geometrical intervals) in combination with five statistical methods. The large scale landslide susceptibility models were derived for small and shallow landslides in the pilot area (21 km²) located in the City of Zagreb (Croatia), which occur mainly in soils and soft rocks. Some of the novelties in LSA are the following: scenarios using stretched landslide conditioning factor values or classification with more than 10 classes prove more reliable; certain statistical methods are more sensitive to the landslide conditioning factor classification criteria than others; all the tested machine learning methods give the best landslide susceptibility model performance using continuously stretched landslide conditioning factors derived from high-resolution input data. The research highlights the importance of qualitative assessments, alongside commonly used quantitative metrics, to verify spatial accuracy and to test the applicability of derived landslide susceptibility maps for spatial planning purposes.