Developing a Systematic Approach to Susceptibility Mapping for Landslides in Natural and Artificial Slopes in an Area Undergoing Land Use Change, Kota Kinabalu, Sabah, Malaysia

Abstract

The principal research effort of this thesis is directed towards understanding the degree to which various factors (both natural and human-induced) act together in time and space to influence landslide susceptibility and hazard within a tropical environment. The study is conducted in the Kota Kinabalu area of Sabah, Malaysia, an area which is undergoing rapid development. Thus, along with investigation of the natural terrain, the representation of susceptibility has had to recognise the impact of development transitions that contribute to slope instability.

The first aim of the analysis is to investigate how landslide distributions change over time in relation to development. The study area was assessed for three survey years; 1978, 1994 and 2010. These assessment years correspond to pre-, mid- and advanced stages of development. To achieve this first aim, two development indicators; land use and road density were used. The different land uses were simplified into four classes with three of these (barren, built up and other) referring to developed areas and forest representing areas still in their natural condition. Road density was classified into three density classes; <50m/40,000m² (low), 50-150m/40,000m² (moderate), and >150m/40,000m² (high). Results showed that as development progresses, landslide occurrence increased from 1978 to 2010. The landslide density in the built up class increased from 19 landslides/100km² to 50 landslides/100km² from 1978 to 2010. Areas that have been cleared (barren) also showed increasing landslide density over this period from 17 landslides/100km² to 70 landslides/100km². In terms of the influence of road density, landslide density sharply increased from 10 landslides/100km² to 62 landslides/100km² within the high road density class, from 1978 to 2010. These results indicate that development plays a major role in inducing landslides in the study area.

An important part of this thesis is the development and testing of suitable susceptibility models, formed the focus of second aim of the thesis. Landslide models representing two different approaches (heuristic and statistical) were employed in this study to model the landslide susceptibility. These models used are: the analytical hierarchy process (AHP), frequency ratio (FR), logistic regression (LR) and discriminant analysis (DA). Thirteen landslide contributing factors were included in each of the models; slope angle, slope aspect, elevation, slope curvatures (plan, profile and tangential), lithology, soil type, lineament density, road density, drainage density, land use and annual rainfall. Of these factors, slope angle, slope aspect, elevation, lithology, soil type and road density were found to be significant based on the statistical analyses using the following techniques: chi-square, Mann-Whitney, T-test, discriminant analysis, and logistic regression. The performance of each of the models was analysed using the degree of fit (DoF) and landslide density (Ld) techniques. The accuracy results indicated that the FR and LR were the most consistent and therefore, better suited the study area. In the same aim, this study also found that to accurately assess the performance of the landslide susceptibility model, the two assessment techniques; landslide density (Ld) and degree of fit (DoF) should be used together. The DoF is useful in giving the overall picture of the accuracy and misclassification of the model, whereas the Ld is useful in showing the number of landslides in each susceptibility class and also for calculating the landslide probability. The DoF and Ld showed a fair and in some cases good correlation in assessing the models’ performances. This shows that both assessment techniques can be used together to assess the model’s performance without producing contradictory results.

In the final stage of this study, the landslides from 1978, 1994 and 2010 were combined and later separated into two dataset: natural and artificial slopes. This was done to assess the differences of landslide contributing factors in each landslide dataset. To analyse both landslide datasets, seven landslide contributing factors were selected. These factors are slope angle, slope aspect, elevation, lithology, lineament density, soil type, and road density. These factors were selected based on their significant association with landslides across the three assessment years (1978, 1994, and 2010). The road density factor was (by definition) only used for the landslides in artificial slope where its contribution is most significant. From the sensitivity analysis, all the six contributing factors (except road density) were found to be important in producing the natural landslide susceptibility map. Conversely, only five factors (slope angle, slope aspect, lithology, road density and soil type) were considered significant in generating the artificial slope landslide susceptibility map. This indicates that the landslide distribution in both datasets was influenced by a slightly different set of landslide factors. Chi-square and Ld methods were used to assess the degree of susceptibility of each of the landslide factor classes as represented by landslide occurrences. Both of these techniques show a good correlation between the highest susceptibility class and the concentration of landsliding.

The landslide density model (LDM) was introduced at the later stage of the analysis and was incorporated into the framework proposed in this study. The LDM was generated using the number of landslides/area. This formula was applied to all the landslide factor classes. Subsequently, the model was generated by spatially overlaying all the landslide factors. The LDM model is much easier to apply than the AHP, DA, LR and FR models. The LDM was used instead of the FR and LR models to represent the condition of the study area because both these models performed poorly in capturing the landslides in natural and artificial slope into their high susceptibility class. The LDM showed consistently greater accuracy in classifying different landslide susceptibility areas than the FR, NSG (Malaysian slope stability guidelines) and LR models for both natural and artificial slope landslides. In addition, the LDM also managed to correctly classify 87% of landslides collected from fieldwork compared to FR (70%), and LR (29%). Maps produced using pre-existing NSG were able to correctly capture only 48 % of the observable landslides.

The framework used in this study provided an opportunity to test the effectiveness of the pre-existing Malaysian national slope guidelines (NSG). As one of the objectives required in this investigation was the construction of a comprehensive landslide inventory, this provided the means of directly testing the NSG in terms of their ability to define areas that are demonstrably less susceptible to landsliding. The NSG were designed to help developers find areas less susceptible to landsliding. Both the slope and mapping guidelines were prepared in separate documents but the key factor used to define landslide susceptibility is the same: slope angle. The guidelines required that slope angle was divided into four susceptibility classes; low (<15°), moderate (15-25°), high (25-35°), and very high (>35°). Higher restriction on development is imposed on the high and very high susceptibility classes. When the effectiveness of the guidelines was compared with the results of susceptibility modelling developed in this thesis, it was found that the current guidelines were unable to accurately define landslide susceptibility for three reasons; insufficient key factors are included, data availability in Malaysia is limited, and the mapping techniques are not well developed.

The landslide susceptibility maps for both natural and artificial slope landslide datasets were converted to landslide hazard maps by including the probability of occurrence values (landslides/km²/year) for each susceptibility class. The hazard descriptors provided for the probability value are based on the descriptors by the Joint Technical Committee (JTC-1). The minimum and maximum probabilities were calculated for each hazard class. Based on the minimum and maximum probability values, the natural landslide hazard map has moderate to high hazard probabilities and the artificial slope landslide hazard map has very low to moderate hazard probabilities.

This study proposes that landslide susceptibility should be defined based on three classes, as opposed to four classes used in the NSG. These classes are low, moderate and high. It is concluded that development should be allowed in the low and moderate susceptibility classes, with higher restrictions placed on development in the moderate class. In order to minimise unwanted consequences caused by landslides, no development should be allowed in the high susceptibility class. This study also proposes that the results from field slope strength assessments (Slope Condition & Risk Rating, Geological Strength Index, and Rock Mass Rating), material properties and hazard probability to be included in each of the susceptibility classes in order to achieve a greater understanding of the susceptibility map. This information is useful in improving the reliability of medium scale (1:50,000) landslide susceptibility maps.

The research framework developed in this study includes methods for selecting and categorising interval data, selecting significant landslide variables, distinguishing the most important variables (sensitivity analysis), incorporating field and laboratory analysis for susceptibility characterisation, and identifying the most suitable landslide susceptibility model. Moreover, this framework is not restricted solely to the specific conditions of the study area and can be applied universally. Problems concerning the limitations of this framework were also addressed, so that developers can take precautionary measures when applying this framework in land use planning. Finally, the framework in this study has successfully demonstrated that it is possible to develop a landslide susceptibility model that is much more successful than the existing national slope and mapping guidelines currently practised in Malaysia.