🤖 machine learning

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

This paper proposes a novel landslide early warning system that utilizes Rainfall-Motion-Aware Contrastive Learning to generate displacement-robust representations from rainfall and terrain data, significantly improving prediction accuracy under the spatial uncertainties inherent in rainfall forecasts.

Original authors: Ren Ozeki, Hamada Rizk, Hirozumi Yamaguchi

Published 2026-05-19

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ren Ozeki, Hamada Rizk, Hirozumi Yamaguchi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict when a heavy rainstorm will cause a hillside to collapse. You have a map of the land (the terrain) and a live video feed of the rain (the weather radar). The problem is, your weather forecast isn't perfect. It's like watching a movie where the actors (the rain clouds) sometimes move a few steps to the left or right compared to where the script says they should be.

In the real world, if a forecast says rain will hit a specific village, but the computer model thinks the rain is actually 5 kilometers away, the prediction might fail. If the model thinks the rain is over the village when it's actually next door, it might miss a landslide warning. Or, it might scream "danger" when the rain is actually safe.

This paper introduces a new "smart brain" for landslide warning systems that doesn't panic when the rain forecast is slightly off. Here is how it works, broken down simply:

The Problem: The "Moving Target" Issue

Most current landslide systems are like students who memorized a textbook perfectly but fail the test if the questions are slightly reworded. They assume the rain forecast is 100% accurate. But in reality, weather forecasts often have "spatial displacement"—the rain is predicted to be in one spot, but it actually moves to a nearby spot.

Because landslides are rare events (like finding a needle in a haystack), there isn't much data to train these systems. This makes them fragile. If the rain moves even a little bit in the forecast, the system gets confused and its accuracy drops.

The Solution: Training with "Blurred" Vision

The researchers created a new system that learns to be "displacement-robust." Think of it like training a security guard to recognize a suspect even if the suspect is wearing a hat, sunglasses, or standing slightly behind a pillar.

They used a technique called Rainfall-Motion-Aware Contrastive Learning (RMCL). Here is the analogy:

The Old Way: You show the computer a photo of a landslide and a photo of a safe hill. It learns to tell them apart.
The New Way (RMCL): You show the computer the photo of the landslide, but then you slide the rain part of the photo a few inches to the left, right, up, or down. You tell the computer, "This is still the same landslide scenario, even though the rain moved."
The Result: The computer learns that the relationship between the rain and the hill matters more than the exact pixel-perfect location of the rain. It learns the "vibe" of the danger rather than memorizing a specific map coordinate.

How the System Works

The Rain Predictor: First, the system looks at the last few hours of rain and uses a "motion tracker" (like following a ball in a video game) to guess where the rain clouds will move next.
The "Stress Test" Training: Before the system goes live, it is trained on thousands of examples where the rain data is intentionally shifted around. This is like practicing for a storm while wearing blinders that make the world look slightly blurry.
The Prediction: When a real storm comes, the system looks at the terrain and the forecasted rain. Even if the forecast is slightly "off" (displaced), the system recognizes the pattern and says, "Hey, this looks like a landslide risk," because it has learned to ignore small shifts in the rain's position.

The Results

The researchers tested this system using two years of data from 19 different regions in Japan where landslides actually happened. They compared their new system against the best existing methods.

The Score: When using real, observed rain, the new system was already very good. But when using forecasted rain (which has errors), the new system was 37% more precise than the other top methods.
What This Means: In a real emergency, this means fewer false alarms (wasting time and causing panic) and fewer missed warnings (saving lives). The system is much better at handling the "uncertainty" of weather forecasts.

Summary

This paper presents a landslide warning system that is tough enough to handle imperfect weather forecasts. By teaching the AI to recognize danger patterns even when the rain moves slightly off-target, it creates a more reliable safety net for communities living on hillsides. It's like teaching a guard dog to bark at a threat even if the intruder is hiding behind a slightly different bush than expected.

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

1. Problem Statement

Rainfall-induced landslides pose an escalating global risk, necessitating reliable Landslide Early Warning Systems (LEWS) to provide sufficient evacuation time. Operational LEWS must estimate near-future landslide risk by integrating observed rainfall with short-term rainfall forecasts. However, a critical challenge exists: rainfall forecasts are inherently uncertain and often exhibit spatial displacement of rainfall fields relative to the actual terrain.

This displacement alters local accumulated rainfall calculations, which are primary inputs for landslide prediction models. Consequently, models trained on accurate observed data often suffer significant performance degradation when deployed with forecast data containing spatial errors. Furthermore, landslide datasets are characterized by severe class imbalance (rare positive events vs. abundant negative samples), hindering the ability of deep learning models to learn comprehensive mechanisms and robust representations. Existing approaches either assume accurate rainfall inputs or train directly on forecast data, risking overfitting to forecast-specific artifacts rather than underlying physical mechanisms.

2. Methodology

The authors propose a novel LEWS framework designed to maintain predictive performance under rainfall forecast uncertainty. The system comprises two main components: a rainfall forecasting module and a landslide prediction module trained via Rainfall-Motion-Aware Contrastive Learning (RMCL).

A. System Architecture

Rainfall Forecasting Model:
- Utilizes an optical-flow-based approach (dense Lucas–Kanade algorithm) to estimate the horizontal motion of rainfall fields.
- Extrapolates the most recent observations to generate short-term forecasts (up to 8 hours ahead), simulating the transport mechanism of rainfall.
- Outputs a continuous sequence of observed and forecast rainfall fields.
Landslide Prediction Model:
- Inputs: Combines (i) spatio-temporal rainfall sequences (observed + forecast) and (ii) static geospatial features (terrain, soil properties, vegetation, elevation, slope).
- Encoder:
  - Spatial Features: Processed via a two-layer CNN.
  - Rainfall Sequences: Processed via a CNN–Transformer architecture (two-layer CNN + three-layer Transformer) to capture temporal dependencies and complex interactions.
- Output: A lightweight Multi-Layer Perceptron (MLP) maps the fused latent representation to landslide probability.

B. Rainfall-Motion-Aware Contrastive Learning (RMCL)

To address spatial displacement and class imbalance, the model is pretrained using a supervised contrastive learning framework.

Data Augmentation (Rainfall Motion Perturbation):
- Instead of independent time-step translations, the method applies temporally correlated perturbations to the rainfall field ( $R_i$ ) while keeping terrain features ( $G_i$ ) fixed.
- This emulates the accumulation of positional errors found in real-world forecasting, creating augmented samples that simulate plausible forecast-induced displacement.
- Gaussian noise is added to both rainfall and terrain features.
Loss Function:
- Supervised Contrastive Loss: Encourages the model to map original and perturbed samples with the same landslide label to nearby points in the latent embedding space, ensuring representations remain stable under displacement.
- Focal Loss: Used during the fine-tuning stage to address class imbalance by emphasizing hard and minority samples.

3. Key Contributions

Displacement-Robust Representation Learning: The paper introduces a method to learn latent representations from rainfall and terrain data that remain stable despite spatial displacement in rainfall field motion, a common issue in operational forecasting.
Temporally Correlated Perturbation: A novel data augmentation strategy that mimics the specific nature of forecast errors (accumulated positional errors over time) rather than random noise, improving the realism of the training process.
Integration of Contrastive Learning for Imbalanced Data: The framework leverages pairwise relationships within the same class to acquire comprehensive knowledge even from scarce positive landslide samples, mitigating the bias toward the majority class.
Operational LEWS Framework: A complete system integrating optical-flow-based nowcasting with a robust prediction model, designed specifically for settings where future rainfall is not directly observable.

4. Experimental Results

The system was evaluated using two years of data (2021–2022) across 19 regions in Japan, covering 253 confirmed landslide events and 3,567 total samples (including negative rainy days). The evaluation used a chronological train-test split to prevent time-series leakage.

Metric: Precision at 80% Recall (to balance detection of rare events with the avoidance of false alarms).
Performance vs. Baselines:
- The proposed method achieved 0.90 precision under observed rainfall and 0.69 precision under forecasted rainfall.
- Compared to state-of-the-art baselines (SWI-EWS, CTLGNet, LiteTransNet), the proposed system showed significant improvements.
- Under forecasted rainfall conditions, the proposed system achieved up to 37% higher precision than the strongest alternative baseline.
Ablation Study:
- Comparing RMCL against standard end-to-end training and end-to-end training with forecast rain inputs, the RMCL approach consistently yielded the highest accuracy.
- Under forecast rainfall settings, RMCL improved accuracy by approximately 50% compared to the baseline training strategies, confirming its effectiveness in handling forecast uncertainty.

5. Significance and Claims

The paper claims that modeling rainfall as a moving spatial field and explicitly addressing rainfall field displacement within the learning process significantly improves the reliability of short-term landslide prediction. By learning representations that are invariant to the spatial shifts inherent in forecasting, the proposed system bridges the gap between idealized model training (using observed data) and operational deployment (using forecast data).

The authors assert that this approach enables landslide prediction models to maintain stable performance in real-time disaster monitoring systems, even when inputs contain spatial displacement errors. The results suggest that addressing the specific nature of forecast uncertainty through contrastive learning is crucial for capturing early indicators of landslide occurrence in operational early warning systems. Future work is noted to potentially extend this framework to other rainfall-induced hazards like floods.