**6. Discussion**

Due to the observed upward tendency in landslide occurrence, authorities of all the administration levels (national, regional and local) are called on to collaborate with the scientific community to spatially determine potential landslide instances and mitigate, or even prevent, the damage and losses that they may cause. LS assessment and mapping is the first and most basic step for effective risk managemen<sup>t</sup> and disaster response [58]. Several LS assessment models have been developed and applied, with their own advantages and disadvantages [59]. A current tendency is the integration of these individual models to enhance their benefits and overcome their weaknesses. The consequent hybrid models are expected to reduce the uncertainty and improve the reliability of the output LS maps [60]. In order to address this statement, in the present study, a hybrid model based

on the integration of two different statistical analysis models, multivariate GeoDetector and bivariate IV, was proposed for LS assessment and mapping. In general, GeoDetector, as a new spatial model, has been rarely used in LS studies compared with other models. Hence, its integrated applications are even more limited. To the best of our knowledge and without ignoring the research works of Luo and Liu [7] and Yang et al. [61], the proposed integration had not been tested hitherto for the development of hybrid LS modeling.

A variety of natural and anthropogenic conditioning factors and a landslide inventory for a Greek wetland around the Pinios artificial lake, were analyzed as inputs in the hybrid model named GeoDIV. It can be stated that the advantages (or disadvantages) of GeoDIV model are "inherited" from the two individual models which it was based on. Under strict, prior defined data assumptions, the IV model is capable of evaluating the impact of each class of many conditioning factors due to the occurrence of past landslides; however, the mutual relationship between the factors is mostly neglected [62]. Without any assumptions on the distribution of data, the GeoDetector model is capable of exploring this relationship but not evaluating individually the impact of each factor class.

In addition to the models, factor selection also plays a major role in the LS results [14]. Too many redundant factors may lead to less realistic and reliable results. Therefore, the capability of significance statistics-based factor selection provided by GeoDetector makes it an ideal option for selecting the most proper factors and then assigning objective weights to them with regard to their different contributions to past landslide occurrence. By incorporating this property of GeoDetector in hybrid GeoDIV model, among the fourteen conditioning factors initially collected, four of them (slope aspect, profile curvature, SPI, and mean annual rainfall) were identified as statistically insignificant and were not finally included in LS assessment. Similar factors were also eliminated as redundant in [14,23,24].

Focusing on the factors qualified from factor selection, slope angle was highlighted by the factor-level weights (q values) of the GeoDIV model. In GeoDetector's terminology, slope angle can be characterized as the factor which most explains the spatial stratified heterogeneity of landslide occurrence in the study area. In simple words, its weight was found to be much higher than the rest of factors, revealing that slope angle has the greatest impact on landslide activity. This is in line with findings from other studies in Greece which, on the basis of using either qualitative or quantitative models at different scales (national and regional), also indicated slope angle as one of the most important factors [38,63,64].

When slope angle interacted to some degree with proximity to roads, an even greater impact was detected, according to the interaction weights. Generally, the single impacts of all other factors were shown to be significantly improved from their interactions with the slope angle. Except for the particularly influential role of the specific factor, this finding also confirms the "nature" of landslides as a phenomenon that, to a grea<sup>t</sup> extent, constitutes the result of interactions between multiple conditioning factors. From a sub-factor perspective, as it was derived from the class-level weights (IVs), the steep parts of study area being very close to roads and covered by scrub vegetation seem to be more prone to landslides.

The output map of the GeoDIV model illustrated the spatial distribution of the estimated LS. It shows that extensive parts, mainly located in south and north, are most likely to have landslides in the future. In comparison with the relevant map from the individual IV model, it can be mentioned that despite the preservation of the general spatial pattern, there was displacement of the pockets from low susceptibility in GeoDIV's map to higher susceptibility in IV's map. This "overestimation" from the IV model may have been due to the inclusion of the additional four conditioning factors, confirming the above statement about the negative impacts of redundant factors on the reliability of LS results.

Regarding the performance of proposed hybrid model, it has to be firstly noted that GeoDIV provided far more than satisfactory validation results in terms of accuracy (success rate) and prediction ability (prediction rate), considering the scale of analysis. Compared to the IV model, although both models seemed to converge to approximate results, the convergence of GeoDIV was found to be faster. This finding proves the expected "superiority" of hybrid against the individual modeling and is in agreemen<sup>t</sup> with previous

studies, concluding that the integration of bivariate with multivariate statistical models improved the performance of former ones [30,47].

Some assumptions and limitations of the present study have to be pointed out. The quality of LS assessment and mapping is highly related to both the landslide inventory and conditioning factors. By using Google Earth satellite imagery, the landslides with identifiable signs in the images were mainly mapped in landslide inventory. Hence, the inventory cannot totally represent the landslide-contributing factors of the study area. Moreover, although the preparation of different susceptibility maps for the various types of landslides can provide more realistic predictions [65], the different types of mapped landslides were not considered in this study. On the other hand, the differentiation between landslide source and deposition zones enabled the models to accurately identify source areas and hence to precisely define the factors that contribute to the initiation of a landslide. The lack of this differentiation resulted to a study's assumption concerning the existence of similar terrain conditions within these zones and thus the representation of each landslide by a single polygon feature. Considerable simplification of these polygons had to be then undertaken by converting them to grid pixels. In this way, an underestimation of landslide data may have taken place in some cases. Additionally, the sampling procedure for the creation of the landslide training and validation datasets can affect the model's efficiency. On the basis of appropriate sizes, a sufficient amount of data should be included in the training dataset, and a remaining "independent" amount of data in the validation dataset. From the perspective of conditioning factors, their spatial resolution and classification can affect the precision of the spatial matching between the landslide and factor data. Therefore, the examination of alternatives for these parameters could lead to different results.
