The study set out to answer a question that has long frustrated geoscientists working in the Hindu Kush and Karakoram regions: when data are scarce and terrain is extreme, which mapping approach best identifies dangerous ground? The researchers compared three widely used techniques—the Analytical Hierarchy Process (AHP), a knowledge-driven method that translates expert judgment into numerical weights; and two bivariate statistical models, the Frequency Ratio (FR) and the Statistical Index (SI), which derive weights empirically from the observed relationship between past landslides and terrain characteristics. What makes their comparison unusually rigorous is the experimental design. Rather than allowing each method to use its own data or assumptions, the team forced all three models to work from the same landslide inventory, the same eight conditioning factors, the same 12.5-meter spatial resolution, the same training and testing split, and a common validation dataset.

Building the foundation for this comparison was itself a substantial undertaking. Between 2017 and 2025, the researchers compiled a multi-temporal inventory of 90 landslides along the corridor, combining satellite image interpretation, documentary records, and repeated field visits to verify each mapped failure. The fieldwork documented a sobering variety of slope failures: rotational slides with well-defined scarps, shallow translational slides, debris slides accumulating talus at their bases, and rockfall debris piling up against the road at the toes of fractured rock faces. The geology of the region—part of the Kohistan Island Arc, squeezed between the Main Mantle Thrust and Main Karakoram Thrust—provides ample raw material for instability, with tectonically disturbed volcanic, metavolcanic, metasedimentary, and intrusive rocks all represented along the route.

With the inventory in hand, the team prepared eight landslide-conditioning factors from a 12.5-meter digital elevation model and geological mapping: slope angle, elevation, aspect, curvature, profile curvature, lithology, drainage density, and relative relief. Before modeling, they tested the factors for redundancy using Spearman’s rank correlation and the Variance Inflation Factor, a standard diagnostic for multicollinearity. The results were reassuring: VIF values ranged from just 1.01 to 2.43, far below the conventional threshold of 10, meaning no factor was duplicating the information carried by the others. Lithology showed the highest VIF at 2.43, followed by slope at 2.12, but all eight variables could be retained without concern that overlapping signals would distort the model weights.

The pattern of empirical associations that emerged from the data reads like a field guide to slope failure in the region. Slope angle showed one of the clearest relationships, with the 30–45 degree class recording the highest frequency ratio of 1.712, and slopes steeper than 45 degrees registering the highest statistical index value. Drainage density proved even more potent: the highest class, 0.309 to 0.5, returned both the highest FR (2.433) and the highest SI (1.673), reflecting how dense stream networks concentrate runoff, undercut slope toes, and raise pore-water pressures during rainfall. Relative relief—the local difference between maximum and minimum elevation—showed a similarly strong association, with the 2097–2277 meter class reaching an FR of 2.059. These relationships are geomorphologically intuitive: steep, deeply dissected terrain provides the gravitational energy and the water pathways that slope failures require.

Lithology added its own decisive fingerprint. The Barawal Banda Slate, a fine-grained, foliated slate and phyllite unit with pervasive cleavage, showed the strongest association with mapped landslides among the well-behaved comparisons, with an FR of 1.457. The rock’s low intact strength and well-developed anisotropic failure surfaces parallel to its foliation make it a natural candidate for instability, and the data confirmed that intuition. West- and southwest-facing slopes also showed elevated susceptibility, plausibly reflecting stronger afternoon solar heating, repeated thermal stress on fractured rock, and differences in moisture retention and vegetation. Convex profile curvature classes, where flow accelerates and lateral support diminishes, added a further local concentration of risk. No single factor told the whole story; the danger zones emerged where multiple unfavorable conditions stacked on top of one another.

When the three models were run and their outputs classified into five susceptibility levels from very low to very high, all three converged on the same broad geography: the northwestern sector of the corridor is the principal hotspot, where steep, highly dissected terrain, dense drainage, high relative relief, and weak lithological units coincide. But the models differed in how sharply they discriminated. In the AHP implementation, slope angle received the highest expert weight at 32.1 percent, followed by drainage density at 21.7 percent and relative relief at 20.1 percent, with a consistency ratio of 0.0737 confirming internally coherent expert judgments. The FR and SI models, by contrast, let the landslide inventory itself speak, assigning each factor class a weight based purely on its observed association with past failures.

Independent validation on the withheld 25 percent testing subset delivered the study’s headline result. The Statistical Index model achieved the highest area under the receiver operating characteristic curve, an ROC–AUC of 0.903—considered excellent discrimination—followed by the Frequency Ratio model at 0.881 and the expert-based AHP at 0.847. In other words, the simplest, purely empirical approach outperformed structured expert judgment in this setting. The SI model also proved remarkably efficient in its spatial allocation: it classified only 12.2 percent of the study area as very high susceptibility while capturing 57.7 percent of the mapped landslides within that class, precisely the kind of concentrated warning that road managers need. The authors are careful, however, to note that without confidence intervals or formal pairwise significance tests, SI should be described as the best performer in this experiment rather than as statistically proven superior.

The practical implications extend well beyond an academic comparison of methods. The maps give engineers and disaster-management authorities in Upper Dir a defensible, reproducible screening tool for prioritizing slope monitoring, drainage improvement, detailed geotechnical investigation, and road-maintenance budgets along a corridor where alternative access routes are scarce and every closure carries real consequences for isolated communities. Sections of road crossing contiguous high and very high susceptibility cells should be first in line for inspection and stabilization, particularly after major rainfall or snowmelt episodes. The authors also stress the limits of the analysis: with only 90 mapped events, no rainfall or seismic predictors, and no temporal validation, the maps indicate where failures are likely, not when or how large. Still, the framework—three interpretable models tested on identical data—offers a template that other data-poor mountain regions, from the Himalaya to the Andes, can adapt. In an era when climate change is intensifying the rainfall that triggers landslides across high-mountain Asia, knowing precisely which kilometers of road to watch first may prove to be the cheapest insurance available.

The distinction between susceptibility and hazard is worth emphasizing for readers encountering these maps. Susceptibility, as produced here, is a purely spatial statement: given the terrain and geological conditions observed today, which locations possess the combination of attributes most conducive to failure. It deliberately excludes timing, magnitude, and runout, which would require rainfall thresholds, seismic triggers, and dynamic runout modeling that the current inventory cannot support. This is why the authors frame their product as a screening tool rather than a forecast, and why they caution against reading the very high class as a prediction of imminent failure.

The statistical logic underlying the two bivariate models also merits brief explanation. Both FR and SI operate class by class: each factor class receives a weight reflecting the proportion of landslide pixels it contains relative to its share of the study area. The Frequency Ratio expresses this as a simple ratio, while the Statistical Index takes its logarithm, which compresses extreme values and can stabilize model behavior when class areas vary widely. Because both models treat each factor independently, they cannot capture interactions—for instance, a steep slope may be far more dangerous under one lithology than another—a limitation that machine-learning approaches address at the cost of greater data demands and reduced transparency.

The regional geological setting amplifies the value of such screening. The Kohistan Island Arc records the collisional history between the Indian and Eurasian plates, and the rocks it preserves—slates, volcanics, and batholithic intrusions—have been repeatedly sheared, fractured, and altered. Tectonic fabrics such as cleavage and foliation create planes of weakness that orient failure surfaces, meaning geology and topography interact rather than act independently. In corridors like Sheringal–Kumrat, where road cuts expose these weakened materials directly to weathering and infiltration, even modest increases in seasonal precipitation can translate into measurable slope instability, reinforcing the case for targeted, map-guided maintenance.

Subject of Research: Landslide susceptibility assessment along the Sheringal–Kumrat Road corridor in Upper Dir, northern Pakistan, comparing AHP and bivariate statistical models

Article Title: Landslide susceptibility assessment along Sheringal Road, Upper Dir Northern Pakistan using AHP and bivariate statistical models

Article References: Khan, S., Anjum, N., Bibi, H., Rauf, M., Ullah, W., Khan, A., Jadoon, H. K., & Yaqoob, A. (2026). Landslide susceptibility assessment along Sheringal Road, Upper Dir Northern Pakistan using AHP and bivariate statistical models. Discover Geoscience, 4(1), Article 353. https://doi.org/10.1007/s44288-026-00725-w

Image Credits: AI Generated

DOI: 10.1007/s44288-026-00725-w

Keywords: landslide susceptibility, AHP, frequency ratio, statistical index, GIS, remote sensing, Upper Dir, Pakistan, Hindu Kush, ROC–AUC, road infrastructure, geohazard mapping