Landslide Susceptibility for a Hill Road Alignment

The validation is circular. The inventory set the weights and then tested the map; the model has already seen the answer. 91% is a measure of fit, not of predictive skill, and will not survive a new rainfall event.

Accuracy is the wrong metric here. Landslides occupy a minuscule fraction of the 1,150 km2. A map declaring the entire district stable would score above 99% accuracy while being useless. Under such class imbalance, precision-recall and the spatial agreement of the highest susceptibility class matter far more than overall accuracy.

Inventory bias: records cluster along roads and settlements, where people observe and report failures; the 2018 event dominates the sample, so the model largely learns one rainfall scenario. Failures smaller than 0.5 ha are barely a few 30 m pixels, and their location is uncertain relative to the DEM cell - position error alone degrades any slope-based signal.

Model issues: eight factors with equal weights is an arbitrary assumption presented as a result, and slope, lithology, road distance and soil depth are mutually correlated, so weights double-count the same underlying control. No sensitivity analysis is reported, so the user cannot know whether the high-susceptibility class is stable or an artefact of the weight choice.

Split by time, not at random. Train on 2015-2017 landslides and test on 2018-2022, or hold out the 2018 event entirely as an out-of-sample scenario. Random point splitting leaks spatial information because neighbouring pixels share terrain, giving falsely optimistic scores.

Split by space as well. Use catchment- or block-wise cross-validation: hold out whole terrain units so the model is tested where it has not seen the terrain. Report the spread of scores across folds, not just the mean - instability across folds is itself a finding worth reporting.

Choose non-landslide samples carefully. Pseudo-absences drawn from flat valley floors make the problem trivial and inflate skill. Draw them from the same domain as the failures - similar slope and road-distance ranges - so the model must discriminate within the terrain where landslides actually occur.

Metrics: ROC/AUC for ranking, but pair it with precision-recall and with the success-rate and prediction-rate curves, which show what fraction of future landslides fall inside the top 10% or 20% of area flagged. For an alignment decision, that is the number that matters: "how much of the district must we treat as hazardous to capture most failures". Also report performance separately for small failures, since those dominate the inventory.

Model changes:

• Reduce redundancy. Check variance inflation or pairwise correlation and drop or combine duplicates (for example slope and soil depth, or road distance and land cover). Fewer, better-justified factors beat eight equal ones.
• Replace equal weights with either a data-driven model (regularised logistic regression for interpretability, or a tree ensemble with monotonic constraints where the physical direction is known) or an explicit expert weighting with sensitivity testing across plausible weight sets.
• Report the ensemble spread. If the high-susceptibility class moves substantially between weight sets, that instability is the headline, not a nuisance.
• Carry the rainfall scenario explicitly: state that the map represents the 2018-type event, and produce a second version for an ordinary monsoon, because alignment decisions differ between them.

Presentation for alignment: aggregate to road-segment support rather than pixels - for each candidate segment, give the mean susceptibility, the proportion of the segment in the top class, and the variability. Add a stated false-negative rate for the high class, derived from the held-out test. Recommend field verification at every segment crossing the top class, and a rainfall-threshold monitoring plan for construction and operation. Present the map as a screening tool with quantified skill, not as a deterministic hazard zone.