Clearly, the quality of knowledge on relationships between landslide susceptibility and predisposing factors from local experts is the key to the success of this knowledge-based approach. As described above, two types of knowledge were needed: (1) the list of predisposing factors that are important to assess landslide susceptibility, and (2) the relationships between the susceptibility and these predisposing factors. The latter, in turn, consists of (2a) the form of the individual relationships, and (2b) the critical values of the predisposing factors at which susceptibility is at its highest, and the critical values at which susceptibility is minimal. The provision of information (knowledge) on (1) is not a challenging task for most competent hillslope geomorphologists or geotechnical engineers, let alone for researchers who specialize in landslides, because this is basic knowledge in landslide research. Sub-type (2a) should not pose much of a challenge for an experienced landslide researcher because the approach described in this paper does not require the local expert to specify the mathematical forms of the relationships. Rather, the forms (in terms of three basic membership functions) are derived from the basic understanding of how these factors are related to landslide susceptibility as shown in Table 1. We expect that this level of understanding is not a challenge for someone with a few years of field-research experience in landslides.
Sub-type (2b) could be considered a challenge if the local expert was asked to provide the susceptibility value for every environmental value of a given variable. In this study, we adopted an approach based on personal construct theory (Zhu, 1999). This approach uses the concept of bipolar distinction, which is believed to be the key element in human learning (Kelly, 1955 and Kelly, 1970). The approach allows the expert to focus on the values for environmental conditions when the susceptibility is at its highest and the values for the environmental conditions when the susceptibility reaches 0. This makes knowledge acquisition from the local expert manageable.
We tested only this approach, using the knowledge from a single expert as an illustration of the idea for using expert knowledge in landslide-susceptibility mapping. It is possible to use the knowledge from multiple experts with this approach, but the issue of how to integrate knowledge from different experts under the fuzzy logic framework, particularly how to resolve the difference in knowledge among different experts, needs to be resolved. Knowledge integration has long been of interest in the Artificial Intelligence (AI) community, and many different frameworks to integrate knowledge have been proposed. Examination of these frameworks for landslide susceptibility using this expert-knowledge approach should be the subject of a separate study.
The portability of a model or an approach is important not only in terms of its usefulness over a wider area, but also in terms of how well we understand the processes the model or approach captured. The results of our case studies suggest that this knowledge-based approach holds up well when it is transferred without changes to an area that is about 19 times larger and much more complicated than the area in which the knowledge base was developed. How this method compares to data-driven approaches, which are widely applied in landslide-susceptibility mapping, in terms of portability is of great interest, but such a comparison is beyond the scope of this paper and merits a detailed study in its own right.
The expert knowledge approach in this study does not use past landslides to develop the knowledge base. It is essentially different from the statistical methods in that the expert knowledge approach does not use data of landslide occurrence and absence to extract the relationships between landslide susceptibility and predisposing factors. It does not have the false negatives during its model development as the statistical methods and some of the data mining methods do.