Learning Displacement-Robust Representations for Landslide Early Warning under Rainfall Forecast Uncertainty

Ren Ozeki †¶ , Hamada Rizk †‡§ , Hirozumi Yamaguchi †§ · 2026

Read Learning Displacement-Robust Representations for Landslide Early Warning under Rainfall Forecast Uncertainty by Ren Ozeki †¶ , Hamada Rizk †‡§ , Hirozumi Yamaguchi †§ (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-17419.txt.

Introduction

Ren Ozeki, Rizk and Yamaguchi build a landslide early-warning system robust to displaced rainfall forecasts, which shift rain fields and corrupt local accumulation estimates. Their model learns latent representations from rain and terrain data that stay stable under field motion, trained with rainfall-motion-aware contrastive learning that mimics forecast displacement. Tests on two years of Japanese data across 19 landslide regions beat leading baselines in precision by up to 37 percent.

Catalog record

CREATORRen Ozeki †¶ , Hamada Rizk †‡§ , Hirozumi Yamaguchi †§
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE29462
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.17419

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Learning Displacement-Robust Representations for Landslide Early Warning under Rainfall Forecast Uncertainty Ren Ozeki †¶ , Hamada Rizk †‡§ , Hirozumi Yamaguchi †§ Abstract Rainfall-induced landslides pose a growing risk worldwide as climate change intensifies extreme rainfall events. To provide sufficient time for evacuation, landslide early warning systems (LEWS) used for real-time disaster monitoring must estimate near-future landslide risk by integrating observed rainfall with short-term rainfall forecasts from spatio-temporal environmental data streams. Although recent landslide prediction methods have improved predictive performance using statistical and deep learning approaches, most assume accurate rainfall inputs. In operational settings, however, landslide prediction relies on rainfall forecasts, which often contain spatial displacement of rainfall fields due to forecasting uncertainties. Such displacement can alter local accumulated rainfall and degrade landslide prediction accuracy. To address this challenge, we propose a novel LEWS that is robust to rainfall field displacement. The key idea is to learn latent representations from rainfall and terrain data that remain stable under displacement in rainfall field motion, enabling reliable geospatial data integration for landslide risk estimation. The landslide prediction model is trained using Rainfall-Motion-Aware Contrastive Learning (RMCL), which introduces temporally correlated rainfall field perturbations to emulate forecast-induced displacement in rainfall-driven spatio-temporal environmental data streams. Experiments were conducted using two years of rainfall and terrain data across Japan, covering 19 regions where landslides occurred. The proposed system achieved up to 37% higher precision than state-of-the-art baselines. These results demonstrate that modeling rainfall as a moving spatial field and addressing rainfall field displacement within the learning process significantly improve the reliability of short-term landslide prediction in operational early warning systems for real-time disaster monitoring. I Introduction Global climate change has intensified extreme rainfall events and increased the frequency of rainfall-induced landslides worldwide [ 22 , 2 ] . Rainfall-induced landslides occur when rainfall accumulates in the soil, increasing soil moisture and reducing slope stability [ 16 ] . Therefore, monitoring rainfall as part of spatio-temporal environmental data streams is critical for predicting landslide occurrence [ 16 , 20 ] . To provide sufficient time for residents to evacuate safely, early warning systems that assess landslide risk before failure occurs are an essential component of real-time disaster monitoring within disaster management systems [ 4 ] . Thus, in operational settings, landslide early warning systems (LEWS) commonly estimate future landslide risk by considering not only current and preceding rainfall conditions but also the expected rainfall over the next several hours [ 4 ] , as shown in Figure 1 . In such LEWS, rainfall observations collected from weather radars and rain gauges form large-scale spatio-temporal environmental data streams that support geospatial data integration and are continuously updated as time series on a spatial grid. These data represent the spatial distribution of rainfall as a rainfall field. Over short timescales of several hours, rainfall fields evolve through rainfall field motion, where coherent rainfall regions move across the landscape driven by atmospheric flow [ 25 ] . To estimate future landslide risk, rainfall forecasting models generate short-term rainfall forecasts that are incorporated into LEWS in the same spatio-temporal format as observed rainfall data for real-time disaster monitoring. However, forecast errors are not avoidable and often appear as spatial displacements of the rainfall field. Such displacement changes whether and how long rainfall remains over a particular location. Since rainfall forecasts are directly used to compute accumulated rainfall and other indicators in landslide prediction models, displacement of the rainfall field can alter the inputs to landslide prediction models and consequently affect prediction accuracy. Additionally, since landslides are extreme events, the number of occurrence cases is limited, resulting in a scarce and highly imbalanced dataset. Such characteristics of landslide data hinder landslide prediction models from acquiring comprehensive knowledge of landslide mechanisms and reduce their robustness to input perturbations caused by rainfall forecasting. Due to these two challenges, accurate future landslide prediction under operational settings remains difficult. Fig. 1: The overview of the landslides early warning system. Previous studies on landslide prediction have explored statistical and machine learning models based on rainfall history and terrain conditions [ 27 , 12 , 23 , 19 ] . Representative approaches include rainfall threshold models that estimate triggering conditions using the relationship between rainfall intensity and duration, as well as machine learning-based risk estimation methods that extend these frameworks [ 4 , 16 ] . More recently, predictive performance has been improved by integrating high-resolution rainfall data and various environmental variables with CNNs or Transformers [ 11 , 31 ] . However, many studies assume that the rainfall inputs used in landslide prediction models are accurate. Consequently, although these models perform well when evaluated using observed rainfall data, their predictive performance may degrade when forecast rainfall is used as input, as investigated in [ 13 ] . To address spatial displacement in rainfall forecasts, some studies train landslide prediction models directly using forecast rainfall data [ 29 , 6 ] . However, since these models rely on forecast rainfall, they may overfit to forecast-specific features rather than the underlying mechanisms of rainfall-induced landslides. Furthermore, accumulated forecast errors over longer prediction horizons can further degrade model robustness and predictive performance. To address these challenges, we propose a novel landslide early warning system designed to maintain predictive performance under rainfall forecast uncertainty. The key idea is to learn latent representations from rainfall and terrain data that are robust to displacement in rainfall field motion. To achieve this, we model rainfall as a moving spatial field over grid-based spatio-temporal data and emulate plausible displacement of rainfall fields using a temporally correlated perturbation process. These augmented rainfall fields are used in Rainfall-Motion-Aware Contrastive Learning (RMCL) to learn representations that remain stable under the displacement of rainfall field motion. By integrating rainfall field perturbation with representation learning, the proposed approach enables landslide prediction models to maintain stable performance even when rainfall forecasts contain spatial displacement errors. Furthermore, the proposed contrastive learning enables landslide prediction models to learn meaningful representations by utilizing pairwise relationships among samples within the same class, allowing them to acquire comprehensive knowledge even from imbalanced data. To evaluate the proposed framework, we collected two years of rainfall and terrain data across Japan, including 19 regions where landslides occurred. Using a chronological train-test split to prevent time-series leakage and reflect operational forecasting conditions, 70% of the earlier data were used for training and the remaining 30% for testing. Model performance was assessed using precision at 80% recall to account for severe class imbalance. Experimental results demonstrate that the proposed system achieves up to 43% higher precision compared with state-of-the-art baselines. These findings highlight the importance of modeling rainfall as a moving spatial field and addressing rainfall field displacement within the learning process for reliable short-term landslide prediction. II Related Work II-A Landslide Prediction Systems Recent global climate change has led to more frequent disasters, even in previously low-risk areas, increasing the demand for accurate and reliable landslide prediction. Traditional landslide prediction approaches are algorithm-based methods [ 27 , 12 , 23 , 19 ] . For instance, the authors of [ 27 ] proposed a statistical method to predict landslides and evaluated their method using a dataset from India as a case study. However, these approaches do not sufficiently consider

Read the complete text