Occupancy Modelling of Biased Landslide Inventories: When Does Separating Occurrence from Recording Improve on Neutralisation?

Abdirashid Omar; Abdikadir Ugaas

cs.CVNov 9, 2026

Occupancy Modelling of Biased Landslide Inventories: When Does Separating Occurrence from Recording Improve on Neutralisation?

Authors: Abdirashid Omar, Abdikadir Ugaas

Abstract

Landslide susceptibility models are trained on inventories that record where landslides were seen, not where they occurred. The established remedy fits bias-describing variables and neutralises them when mapping. We adapt occupancy-detection modelling from ecology, treating several inventories as imperfect observers of one event and separating occurrence from recording, and ask when this improves on neutralisation. We use near-complete inventories of the 2018 Hokkaido Eastern Iburi, 2015 Gorkha and 2010 Haiti earthquakes with simulated reporting bias, and three real Gorkha inventories. Compared with standard practice, the occupancy model reduces road bias in 24 of 25 paired draws and raises AUC by 0.08-0.14 on Gorkha and 0.03 on Haiti, but only 0.01-0.02 on Iburi. Its median predicted-to-true landslide rate is 0.99 (range 0.44-1.89), against 0.24-0.30 for the baselines. Rescaling the baselines with the same anchor area corrects their median but leaves larger errors (range 0.20-4.34). Cell-level probabilities remain too extreme (calibration slope 0.59). Compared with the best neutralised model, AUC is higher on Iburi (by 0.03-0.07) and lower on Gorkha and Haiti (by 0.01-0.05). Transplant experiments tie the Iburi advantage to the structure of its true landslide pattern, not to its terrain or road network. In all eight settings with a low interaction gap, computable from biased records, neutralisation was at least as accurate, including a prespecified test on Haiti; a high gap did not guarantee an advantage. The real Gorkha inventories are slope-biased, not road-biased, and there the model brings no meaningful gain. Report-based catalogues remain the main untested case.

Figures & tables

Panel of Figure 1 — Figure 1: Model structure. The susceptibility head sees only terrain and the detection head sees only access. Both enter the occupancy likelihood, fitted to several inventories at once, with the anchor area fixing the scale of detection. Only $\psi$ is mapped.

	2018 Hokkaido Eastern Iburi	2015 Gorkha (Sindhupalchok window)	2010 Haiti
Reference inventory	Zhang et al. [26] , 5,625 polygons	Roback et al. [27] , Zekkos et al. [28] , source areas, 12,246 polygons	Harp et al. [29] , 23,567 polygons (21,666 in window)
Area and mask	convex hull of the inventory	mapped extent minus obscured areas, 27.65–28.05 ∘ N, 85.40–86.10 ∘ E	convex hull of the inventory, 18.18–18.50 ∘ N, 72.98–72.17 ∘ W
Grid	120 m, 56,399 cells, 16.3% landslide	210 m, 62,652 cells, 10.4% landslide	210 m, 63,995 cells, 10.4% landslide
Epicentres	mainshock	mainshock and 12 May $M_{w}$ 7.3 aftershock	mainshock
Real independent inventories	–	Gnyawali and Adhikari [30] , Valagussa et al. [31] , Zhang et al. [32]	–

Table 1: Study sites. Haiti was analysed only after the selection rule was fixed (Sect. 5.5 ).

Figure 2: Iburi semi-synthetic setup (scenario A, draw 0). (a) Reference inventory with public roads. (b) Simulated biased records, with satellite footprints in blue. (c) Reference landslides in the held-out test blocks. (d–f) Predicted risk percentile from three models. This draw is favourable to the occupancy model. Means over five draws are 0.754 for standard practice and 0.772 for the occupancy model.

Model	Iburi A	Iburi C	Gorkha A	Gorkha C	Haiti A
Standard practice	$0.754\pm 0.078$	$0.789\pm 0.097$	$0.685\pm 0.088$	$0.698\pm 0.033$	$0.739\pm 0.030$
Neutralised MLP	$0.736\pm 0.029$	$0.721\pm 0.068$	$0.806\pm 0.042$	$0.797\pm 0.038$	$\mathbf{0.784}\pm 0.012$
Neutralised GLM	$0.713\pm 0.035$	$0.619\pm 0.137$	$\mathbf{0.820}\pm 0.028$	$\mathbf{0.820}\pm 0.018$	$0.783\pm 0.015$
Neutralised GAM	$0.731\pm 0.028$	$0.703\pm 0.099$	$0.814\pm 0.018$	$0.792\pm 0.028$	$0.780\pm 0.019$
Occupancy model	$\mathbf{0.772}\pm 0.060$	$\mathbf{0.799}\pm 0.054$	$\mathbf{0.820}\pm 0.030$	$0.780\pm 0.027$	$0.773\pm 0.023$
Oracle	$0.822\pm 0.044$	$0.837\pm 0.051$	$0.857\pm 0.007$	$0.860\pm 0.010$	$0.802\pm 0.018$

Table 2: Test AUC against the reference inventory (mean

\pm

SD over 5 draws). Bold marks the best non-oracle model in each column. The GAM was added after the Haiti prediction was recorded.

Figure 3: Standard practice, three neutralised models, the occupancy model and the oracle in each setting. Large markers are means and small dots are the 5 draws. Horizontal axes do not start at zero.

Model	Pred. $\div$ true rate,	Mean	ECE	Log loss	Brier	Calib.
	median (range)	$\|\log\text{rate}\|$				slope
Standard practice	0.30 (0.13–0.61)	1.24	0.088	0.486	0.109	0.41
Neutralised MLP	0.24 (0.09–0.53)	1.48	0.094	0.448	0.113	0.61
Neutralised GLM	0.29 (0.18–0.58)	1.22	0.089	0.412	0.112	1.14
Standard, scaled	1.01 (0.20–2.64)	0.60	0.109	0.816	0.131	0.15
Neutralised MLP, scaled	0.85 (0.21–2.39)	0.61	0.091	0.604	0.124	0.41

Table 3: Probability estimates on test blocks, 25 re-fitted draws over the five settings of Table 2 . “Scaled” divides a baseline’s prediction by the share of anchor landslide cells recorded by any inventory.

Figure 4: Reliability curves on test blocks, averaged over 5 draws per setting (dashed line = perfect calibration); the neutralised GLM stands in for the other neutralised baselines. Bottom right: predicted $\div$ true overall rate for every draw and model, on a logarithmic axis; vertical ticks are medians.

Label	True pattern	Gap,	Gap,	Occ. $-$	Occ.
		reference	records	best	higher
I-add	Iburi terrain: additive GLM fit of the inventory	$-0.005$	$-0.006$	$-0.037$	0/5
I-rand	Iburi terrain: GLM fit + synthetic interactions	$+0.056$	$+0.037$	$-0.019$	1/5
I-real	Iburi terrain: flexible fit on terrain	$+0.046$	$+0.018$	$+0.038$	5/5
I-real+roads	Iburi terrain: flexible fit on terrain + roads	$+0.050$	$+0.019$	$+0.047$	5/5
I-no-epi	Iburi terrain: flexible fit, no epicentral distance	$+0.004$	$-0.005$	$-0.009$	2/5

Table 4: Interaction gaps and the AUC of the occupancy model minus the best neutralised model (MLP, GLM or GAM), 5 draws each. Upper part: transplant worlds, with labels as in Fig. 5 . Lower part: real settings of Table 2 .

Figure 5: Interaction gap against the AUC difference between the occupancy model and the best neutralised model, for real settings (filled circles) and transplant worlds (open squares; labels as in Table 4 ). The shaded band is the no-prediction range of the rule. Haiti, in bold, is the only setting not used in developing the rule.

Model	AUC	Remote	Rate	Bias
Standard practice	0.739	0.768	0.24	1.70
Neutralised MLP	0.784	0.791	0.19	0.65
Neutralised GLM	0.783	0.803	0.29	0.39
Neutralised GAM	0.780	0.773	0.11	0.65
Occupancy model	0.773	0.767	0.97	0.28
Oracle	0.802	0.819	0.99	0.45

Table 5: Haiti 2010, scenario A, means over 5 draws. “Remote” is AUC in the most remote quarter; “Rate” is predicted

\div

true rate; “Bias” is road bias.

Model	AUC vs reference	AUC vs consensus	Slope bias	Pred. $\div$ true rate
Standard practice	0.830	0.867	0.44	1.11
Standard, roads dropped	0.836	0.872	0.41	1.13
Neutralised MLP	0.829	0.863	0.25	1.15
Neutralised GLM	0.825	0.858	0.28	1.05
Occupancy model	0.837	0.873	0.47	1.38
Occupancy + visibility detection	0.835	0.872	0.31	1.35

Table 6: Real Gorkha inventories (5 draws, held-out blocks). Slope bias is

|\log|

of the steep/gentle predicted-rate ratio relative to the reference. The consensus includes the training records, so it is not an independent test.

Figure 6: Detection learned for each simulated observer (Iburi, scenario A, draw 0). Curves show the additive effect of each feature on detection log-odds, constrained to the known direction. The last panel compares learned and true detection probability (dashed line = equality; colours as in the other panels).

Landslide Susceptibility Mapping

See all 11 papers

Which Site, and When: A Free-Satellite-Data Test of Himalayan Glacial Lake Bursts, Landslides, and Ice Floods
Aug 12, 2026Matthew Kahn, Milan Arjel, Nirmala Adhikari +2Event Time Prediction Landslide Susceptibility Mapping
Applied Economics and Management, Cornell University, New York, United States of America · Department of Electronics and Computer Engineering, Purwanchal Campus, Institute of Engineering, Tribhuvan University, Dharan, Nepal · Department of Physics, University of Alabama at Birmingham, Alabama, United States of America +2
From Landslide Conditioning Factors to Satellite Embeddings: Evaluating the Utilisation of Google AlphaEarth for Landslide Susceptibility Mapping using Deep Learning
Jan 12, 2026Yusen Cheng, Qinfeng Zhu, Lei FanGeospatial Foundation Models Geospatial Representation Learning
Department of Civil Engineering, Xi’an Jiaotong-Liverpool University, Suzhou, 215123, China · Department of Computer Science, University of Liverpool, Liverpool, L69 3BX, UK
Spatial Heterogeneity-Aware Multi-Hazard Susceptibility and Risk Mapping at Regional Scale
Aug 8, 2026Aswathi Mundayatt, Siddharth Anil, Hitanshu Seth +1Spatial Generalization Landslide Susceptibility Mapping
Graphics-Visualization-Computing Lab, International Institute of Information Technology Bangalore, Karnataka 560100, India

Aug 12, 2026J/K move · Enter open · S save

Which Site, and When: A Free-Satellite-Data Test of Himalayan Glacial Lake Bursts, Landslides, and Ice Floods

Matthew Kahn, Milan Arjel, Nirmala Adhikari, Mingmar Sherpa+1

Applied Economics and Management, Cornell University, New York, United States of America · Department of Electronics and Computer Engineering, Purwanchal Campus, Institute of Engineering, Tribhuvan University, Dharan, Nepal · Department of Physics, University of Alabama at Birmingham, Alabama, United States of America+2

Event Time Prediction Landslide Susceptibility Mapping Hydrology

Open paper AskPDF arXiv

Two free satellite signals carry real information about glacial-lake outburst risk in the Nepal Himalaya: radar interferometry sees a moraine dam slowly sagging, and satellite weather marks the weeks when a primed lake is under stress. A companion feasibility study found that deformation indicates which lake is destabilizing and weather indicates when it is at risk, but proposed no predictive model. To address this gap, we propose and evaluate models that predict which site is susceptible and when a trigger arrives. We test three related hazards on free data alone: large moraine- and ice-dammed bursts, rainfall-triggered landslides, and smaller floods from ponds on and around a glacier. Each hazard gets two questions, never blended. Using 589 dated outbursts from HMAGLOFDB and several thousand catalogued landslides, we match each event against similar but unfailed sites, and hold every model to a strong simple baseline under spatial cross-validation that withholds whole map tiles, so no model succeeds by recognising a trained-on neighbourhood. Antecedent weather times the trigger at ROC 0.73 for big bursts, 0.83 for landslides, and 0.82 for small floods. Terrain ranks susceptibility only in part: scored naively it appears near 0.9, largely because catalogued failures cluster in wetter ranges; matched against comparable nearby sites the honest figures are 0.76, 0.71, and 0.54 (no better than chance). The burst signal holds within single regions, reaching 0.89 in Nepal alone. Five deep-learning models do not decisively beat a simple gradient-boosted baseline. Three score marginally higher on landslides, a hint too small to confirm. For the lake hazards the baseline wins outright, reproduced by a three-rule decision tree on ruggedness and monsoon rainfall. We close with a ranked Nepal watchlist, a prioritisation aid, not a prediction, and note where free data reaches its limits.

Abstract

Landslide Susceptibility Mapping

Explore similar work

Which Site, and When: A Free-Satellite-Data Test of Himalayan Glacial Lake Bursts, Landslides, and Ice Floods

From Landslide Conditioning Factors to Satellite Embeddings: Evaluating the Utilisation of Google AlphaEarth for Landslide Susceptibility Mapping using Deep Learning

Spatial Heterogeneity-Aware Multi-Hazard Susceptibility and Risk Mapping at Regional Scale