Prediction of precipitation-induced landslides and sediment discharge at the basin scale using machine learning

Kido, Riho; Inoue, Takuya; Yamanoi, Kazuki

doi:10.5194/egusphere-2026-2689

Preprints

https://doi.org/10.5194/egusphere-2026-2689

Preprints

09 Jul 2026

| 09 Jul 2026

Status: this preprint is open for discussion and under review for Earth Surface Dynamics (ESurf).

Prediction of precipitation-induced landslides and sediment discharge at the basin scale using machine learning

Riho Kido, Takuya Inoue, and Kazuki Yamanoi

Abstract. During heavy rainfall, sediment discharge from mountainous regions is exacerbating damage in downstream urban areas. Therefore, predicting sediment discharge from mountainous area is of critical importance. A large proportion of discharged sediment is produced by slope failures and subsequently transported through channel networks. Although topographic and geotechnical conditions vary within a watershed, both the susceptibility to slope failure and the volume of sediment produced depend strongly on these conditions. In this study, we develop a machine learning model that predicts slope failure occurrence and landslide volume from topographic and geotechnical parameters. By coupling this model with a rainfall and sediment runoff, we propose an integrated framework that simulates the entire process from slope failure to sediment production and downstream transport at the watershed scale. The proposed model incorporates uncertainties associated with unaccounted variability through a probabilistic representation, enabling the evaluation of multiple plausible scenarios. The model was applied to the Pekerebetsu basin for the 2016 Hokkaido heavy rainfall event. Repeated simulations under identical topographic, geotechnical, and rainfall conditions produced slightly different spatial patterns and numbers of slope failures. However, all simulations reproduced sediment production and discharge close to observed values. These results demonstrate that the proposed framework can consistently capture watershed-scale sediment dynamics while accounting for inherent variability in slope failure processes.

Received: 12 May 2026 – Discussion started: 09 Jul 2026

Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims made in the text, published maps, institutional affiliations, or any other geographical representation in this paper. While Copernicus Publications makes every effort to include appropriate place names, the final responsibility lies with the authors. Views expressed in the text are those of the authors and do not necessarily reflect the views of the publisher.

Download & links

Riho Kido, Takuya Inoue, and Kazuki Yamanoi

Status: open (extended)

Post a comment Subscribe to comment alert

RC1:
'Comment on egusphere-2026-2689', Anonymous Referee #1, 11 Sep 2026 reply
General comments: The manuscript presents a useful and timely framework that integrates rainfall infiltration–slope stability analysis, machine learning, and watershed-scale sediment transport modelling to predict rainfall-induced landslides and sediment discharge. The coupling of Random Survival Forest and Random Forest models with the SiMHiS framework is potentially valuable, particularly because the approach attempts to represent uncertainty in failure occurrence while reducing the computational cost of applying detailed slope-stability analyses across an entire basin. However, the manuscript would benefit from further clarification and validation of several simplifying assumptions, particularly those related to uniform soil hydraulic properties and soil depth, representation of unsaturated soil behaviour, model transferability beyond the Pekerebetsu basin, and the physical interpretation of the probabilistic machine-learning framework.
Specific comments
The novelty of the proposed framework should be stated more explicitly in the Introduction. The manuscript appropriately identifies the computational difficulty of applying coupled infiltration–stability analyses to every slope in a watershed and proposes machine learning as a surrogate. However, the distinction between the present RSF/RF–SiMHiS framework and previous hybrid physically based–machine learning approaches should be summarized more clearly, preferably in a short paragraph or comparison table describing previous model inputs, outputs, treatment of rainfall infiltration, uncertainty, landslide volume, and sediment routing.

The treatment of unsaturated soil behaviour requires stronger justification. Rainfall infiltration is modelled using Richards' equation, but the hydraulic parameters θs\theta_s, θr\theta_r, ψm\psi_m, σ\sigma, and KsK_s are assigned uniformly to all simulated slopes. Because the manuscript emphasizes spatial heterogeneity at the watershed scale, the assumption of identical unsaturated hydraulic properties needs to be justified and its influence quantified through sensitivity analysis. In particular, variations in the soil-water characteristic curve and unsaturated hydraulic conductivity can strongly control rainfall infiltration, suction dissipation, and the timing of slope failure.

I recommend adding the following reference: DOI: 10.1631/jzus.A1800593. This reference is directly relevant because it presents measurement of both the SWCC and unsaturated permeability over different suction ranges, which are precisely the hydraulic functions required by rainfall infiltration models. The citation would strengthen Sect. 2.1 when discussing C(ψ)C(\psi), K(ψ)K(\psi), and the need for representative hydraulic parameters rather than assuming a single set of values for all slopes.

The definition of slope failure through a “critical saturated layer thickness” should be discussed more carefully from an unsaturated-soil mechanics perspective. The model defines hcrh_{cr} from the portion of the soil profile where pressure head becomes zero or positive. Rainfall-induced slope failure, however, can occur before complete local saturation because the reduction of matric suction causes progressive loss of unsaturated shear strength. The authors should clarify whether the stability analysis explicitly includes suction-dependent shear strength before saturation, or whether unsaturated conditions influence stability only indirectly through pore-water pressure.

The paper would benefit from adding DOI: 10.4417/IJGCH-05-01-04. This work demonstrates through field monitoring and numerical analysis how rainfall infiltration modifies negative pore-water pressure and slope factor of safety in tropical residual soils, and how controlling infiltration can substantially improve slope stability. It would be particularly appropriate in the Introduction where the manuscript describes rainfall infiltration, pore-water-pressure increase, and reduction in slope stability, and also in the Discussion when interpreting the role of infiltration processes in landslide initiation.

The assumption of a spatially uniform 3-m soil depth needs stronger validation. The authors acknowledge that soil depth and failure width vary spatially and recommend future three-dimensional analyses. Because landslide volume is calculated directly using predicted failure length multiplied by soil depth and slope width, uncertainty in soil depth propagates almost linearly into sediment-production estimates. I suggest conducting at least a simple sensitivity analysis using several plausible soil depths, for example 1, 2, 3, 4, and 5 m, to quantify the effect on estimated landslide volume and basin-scale sediment discharge.

The probabilistic interpretation of unresolved heterogeneity needs additional justification. The manuscript states that factors not explicitly represented—such as vegetation effects and spatial variability in soil depth—are indirectly incorporated through probabilistic failure prediction. This is not necessarily guaranteed: stochastic model output does not automatically reproduce the statistical effects of omitted physical variables. The authors should make this claim more cautious and ideally demonstrate it by testing whether introducing plausible distributions of cohesion, hydraulic properties, soil depth, or vegetation-related root cohesion produces uncertainty comparable with the RSF-derived probabilistic range.

The suggested reference, Hamdany et al. (2022), 586–596, DOI: 10.1016/j.iswcr.2022.01.007, should also be incorporated. The study demonstrates field measurement of soil suction and water content and the derivation of field SWCC behaviour using long-term instrumentation. This is highly relevant to the present manuscript because future practical implementation of the proposed watershed-scale framework will require field observations of temporal soil-moisture/suction changes for calibration and validation. The reference could therefore be cited in the Methodology or Discussion when recommending monitoring-based validation of predicted saturation levels and failure thresholds.

Model validation should be strengthened beyond a single historical rainfall event and aggregate sediment volumes. The RSF achieves an AUC of 0.89, and the simulated sediment production and discharge for the 2016 Hokkaido event are close to reported values. However, the same watershed and event are used to demonstrate the integrated framework. Validation against an independent storm event, another sub-basin, or spatially observed landslide inventories would provide much stronger evidence of transferability. At minimum, the authors should report additional quantitative metrics such as precision, recall, F1 score, Brier score/calibration error for failure probability, RMSE or MAE for failure length, and error percentages for sediment production/discharge.

The Discussion and Conclusion should more clearly separate demonstrated findings from model assumptions and future applicability. The manuscript concludes that the framework successfully reproduces sediment production and discharge during the 2016 event, while important assumptions remain regarding uniform hydraulic properties, constant soil depth, fixed friction angle, calibrated cohesion, slope width, and river-sediment transport parameters. The conclusion should therefore explicitly state that the reported performance is site- and event-specific at this stage, and future work should examine spatially distributed SWCC/hydraulic conductivity, variable soil depth and vegetation, additional rainfall events, independent basins, and three-dimensional landslide geometry before the method is considered broadly transferable.

Reply
Citation: https://doi.org/10.5194/egusphere-2026-2689-RC1

Riho Kido, Takuya Inoue, and Kazuki Yamanoi

Viewed

Total article views: 311 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
201	86	24	311	20	24

HTML: 201
PDF: 86
XML: 24
Total: 311
BibTeX: 20
EndNote: 24

Views and downloads (calculated since 09 Jul 2026)

Month	HTML	PDF	XML	Total
Jul 2026	120	42	12	174
Aug 2026	22	14	2	38
Sep 2026	59	30	10	99

Cumulative views and downloads (calculated since 09 Jul 2026)

Month	HTML	PDF	XML	Total
Jul 2026	120	42	12	174
Aug 2026	22	14	2	38
Sep 2026	59	30	10	99

Viewed (geographical distribution)

Total article views: 221 (including HTML, PDF, and XML) Thereof 221 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 29 Sep 2026

Short summary

It is important to predict sediment production by slope failures and its transport at the basin scale. In this study, we used the results of rainfall infiltration–slope stability analyses to develop machine-learning models that estimate failure probability and failure length from topographic and geotechnical parameters. By incorporating these models into a sediment runoff model, we successfully reproduced the large-scale sediment discharge event.


Total:	0
HTML:	0
PDF:	0
XML:	0