Preliminary Comparison of Frequency-Ratio Landslide Susceptibility and Previously Calculated Proforma-Based Severity Levels Along the Pekan Nabalu–Kundasang Corridor, Sabah

Landslides are among the most significant geological hazards in Sabah, Malaysia, affecting communities, transportation infrastructure, and the natural environment. The Pekan Nabalu–Kundasang corridor is particularly susceptible to slope instability because of its steep topography, complex geological setting, tropical rainfall, and active tectonic environment. The effects of the 2015 Ranau earthquake further demonstrated the sensitivity of slopes surrounding Mount Kinabalu to earthquake-induced ground disturbance and mass movement [1-4].

Previous studies in the Kundasang region have primarily focused on landslide susceptibility mapping using statistical and GIS-based approaches. Researchers have applied Weights of Evidence and Frequency Ratio (FR) models to identify areas with different levels of spatial susceptibility [5-7]. Other studies have examined the influence of lithology, slope geometry, structural discontinuities, land use, runout characteristics, and debris-flow processes [8-11]. These studies provide valuable information on where landslides are more likely to occur but do not directly describe the severity characteristics of individual observed landslides.

Landslide susceptibility and landslide severity represent different concepts. Susceptibility describes the relative spatial likelihood that a landslide may occur, whereas a complete hazard assessment additionally considers temporal occurrence and landslide intensity. The severity or intensity of an observed landslide may be influenced by its magnitude, velocity, displacement, runout, slope geometry, and material characteristics [12, 13]. Therefore, the severity class assigned to an observed landslide does not necessarily correspond directly to the susceptibility class of its location.

The reliability of susceptibility assessments can also be affected by the completeness of landslide inventories, model-development procedures, sampling design, and uncertainty in the conditioning data [14-16]. Consequently, a severity rating should not be used as an independent measure of the predictive accuracy of a susceptibility map. Instead, comparison between the two can be used to identify areas of agreement and divergence.

The Landslide Hazard Identification (LHI) Proforma provides a structured approach for documenting geological, geomorphological, hydrological, land-use, and geotechnical characteristics of observed landslides. The recorded parameter ratings can be combined to calculate a Landslide Severity Level (LSL) [3].

Therefore, this study presents a preliminary comparison of FR-based landslide susceptibility classes and previously calculated Proforma-based LSL scores for twenty-five observed landslides along the Pekan Nabalu–Kundasang corridor. The objective is to quantify agreement, divergence, and ordered association within the 25-site dataset. The comparison is not used to validate FR predictive accuracy; instead, it identifies concordant and discordant records that can support prioritisation for field verification, geotechnical investigation, or monitoring. The analysis uses the final LSL scores because the original Proforma input matrices are not part of the dataset provided for this study.

This study analyses a 25-site dataset containing FR susceptibility classes and previously calculated LSL scores and classes. The analysis consists of three main components:

(1) compilation of the FR susceptibility class and final LSL score and class for each of the twenty-five analysed sites;

(2) descriptive comparison of the FR and LSL classifications, including class-wise LSL summaries;

(3) agreement and ordered-trend analyses applied to the 25-site dataset.

3.1 Frequency ratio susceptibility map and site selection

The existing FR map provides the regional susceptibility classification for the 25 site records analysed in this study. The LSL results are not used to validate the FR model because susceptibility and severity represent different landslide characteristics, and the 25-site dataset is not an independent validation inventory. The dataset contains five observed landslides in each FR susceptibility class, producing a balanced 25-site dataset. No reproducible probability-based random sampling procedure is documented for these records. The dataset is therefore treated as a balanced stratified sample by FR class, not as a random sample. Equal representation across the five FR classes supports class-to-class comparison but does not represent the natural frequency distribution of landslides in the mapped area. Sampling design and the completeness and spatial representativeness of landslide inventories can influence susceptibility modelling and interpretation [17-20]. Some sites occur within the same general road corridor and may exhibit spatial clustering. Each site identifier, L1–L25, represents one recorded slope-failure observation.

Appendix Table A1 reports a separate 25-record landslide inventory with approximate coordinates, landslide type, activity status, and origin; coordinates are rounded to three decimal places. The files provided for this study contain no verified crosswalk between the Appendix Table A1 records and manuscript site identifiers L1–L25. For this reason, the coordinates are not assigned to individual L1–L25 records. This limitation prevents site-by-site spatial replication.

3.2 Previously calculated Landslide Severity Level data and source records

The dataset contains final LSL scores and severity classes previously calculated using the LHI Proforma developed by Roslee [3]. This study does not repeat the original site-level Proforma assessment. Project documentation states that the original assessment used field observations, laboratory analyses, published literature, and geotechnical records.

Project documentation states that geotechnical and laboratory information supported parameters including Standard Penetration Test values, cohesion, friction angle, shear strength, point-load index, compressive strength, and Rock Quality Designation. However, the original report titles, borehole identifiers, sample identifiers, and site-by-site source records for these parameters were not available to the authors. Therefore, these details could not be reconstructed from the final LSL totals and were not inferred in this study.

Table 1. Parameters and sub-parameters of the Landslide Hazard Identification (LHI) Proforma adapted from Roslee [3]

The LHI Proforma framework contains seven major parameter groups: geological characteristics, geodynamic features, geomorphometric features, hydrology and hydrogeology, land use, soil mechanical characteristics and rock mechanical characteristics. Table 1 lists the sub-parameters defined in the source framework to document the method used in the original assessment; it does not represent recovered site-level input data.

Under the original LHI Proforma framework [3], individual sub-parameter ratings contribute to seven component scores (S1–S7), which are summed to obtain the total LSL score. The dataset contains the final LSL scores and classifications, but not the original site-specific sub-parameter scoring sheets or S1–S7 component matrices. The project files contain no record explaining the absence of these source records. Because a final LSL total does not uniquely determine its component scores, worked site examples and summary statistics cannot be reconstructed without introducing hypothetical values. No reconstructed values are reported.

3.3 Landslide Severity Level aggregation framework and classification

According to the original LHI-LSL framework [3], the total LSL is defined as the sum of the seven major parameter-group scores:

LSL = S1+S2+S3+S4+S5+S6+S7                (1)

where,

S1 = Geological characteristics

S2 = Geodynamic features

S3 = Geomorphometric feature

S4 = Hydrology and hydrogeology

S5 = Land use

S6 = Soil mechanical characteristics

S7 = Rock mechanical characteristics

The severity thresholds were adopted from the original LHI-LSL model developed by Roslee [3]. In this secondary analysis, the final LSL scores, reported to one decimal place, were classified using the explicit boundaries in Table 2. The scores were not recalculated from S1–S7 because the original component matrices are not part of the dataset.

The comparison examined agreement and divergence between the FR susceptibility class recorded for each site and the corresponding LSL class. Agreement statistics quantify correspondence within the 25-site dataset and are not interpreted as validation of FR predictive accuracy or as evidence that the two classifications measure the same construct.

Table 2. Landslide Severity Level (LSL) classification adapted from Roslee [3]

3.4 Statistical analysis

Agreement between FR susceptibility classes and LSL classes was quantified using percentage exact agreement and Cohen's kappa (κ). Because both classifications consist of five ordered categories, linearly weighted Cohen's kappa was also calculated to reflect the magnitude of class disagreement. The ordered trend in LSL scores across increasing FR susceptibility classes was tested using the Jonckheere-Terpstra test, with p < 0.05 as the statistical threshold. Because the site-selection procedure was not probability-based, the statistics characterize the 25-site dataset and are not generalized to the wider landslide population.

This preliminary comparison compared FR landslide susceptibility classes with previously calculated Proforma-based LSL scores for twenty-five landslide records along the Pekan Nabalu-Kundasang corridor. The analysis used the FR classes, final LSL scores and classes, and locality information recorded for these sites. It did not recalculate the original Proforma component scores.

Within the 25-site dataset, two sites were classified as Very Low LSL severity, eleven as Low, seven as Moderate, three as High, and two as Very High. The Very Low FR class contained two Very Low- and three Low-severity sites; the Low class contained five Low-severity sites; the Moderate class contained two Low- and three Moderate-severity sites; the High class contained four Moderate- and one High-severity site; and the Very High class contained one Low-, two High-, and two Very High-severity sites.

Exact class agreement occurred at 13 of 25 sites (52%), with Cohen's kappa of 0.40 and linearly weighted kappa of 0.61. LSL scores also showed a significant ordered increase across FR classes (Jonckheere-Terpstra J = 231, p < 0.001). These results demonstrate partial correspondence and an ordered relationship within the 25-site dataset. They do not independently verify the original LSL calculations or validate FR predictive accuracy.

The comparison supports preliminary screening by distinguishing sites where regional susceptibility and site-level severity are concordant from sites where they diverge. The study does not verify the original LSL arithmetic or serve as an independent validation of the FR model because the original Proforma score matrices, detailed geotechnical source records, and verified site crosswalk are not part of the dataset. Future work should use fully documented source data and a reproducible sampling procedure.