📄 earth science

Leveraging geospatial foundation model embeddings for local-scale landslide susceptibility mapping via positive-unlabeled learning

This study proposes a deployment-oriented workflow that integrates AlphaEarth foundation model embeddings with positive-unlabeled learning to overcome label and redundancy challenges, achieving highly selective local-scale landslide susceptibility mapping in Hong Kong by effectively fusing transferable geospatial representations with traditional physical features.

Original authors: Mingxuan Li, Wenkai Li, Hongliang Liu, Changshan Liang

Published 2026-09-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mingxuan Li, Wenkai Li, Hongliang Liu, Changshan Liang

Original paper licensed under CC BY 4.0 (https://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

Every year, heavy rains and steep terrain conspire to send tons of rock and soil sliding down hillsides, threatening lives and infrastructure. In places like Hong Kong, where cities are built right up against rugged mountains, predicting where these slides might happen is a matter of urgent safety. For decades, experts have mapped these risks by measuring physical factors: how steep a slope is, what kind of rock lies beneath, how much water flows through the soil, and how far a hill is from a river. These measurements are reliable, but they often miss the subtle, complex textures of the ground that signal danger. In recent years, a new kind of artificial intelligence has emerged, trained on vast amounts of satellite imagery to recognize patterns in the Earth's surface that humans cannot easily see. These systems, known as geospatial foundation models, can generate a digital "fingerprint" of a landscape, capturing details about vegetation, soil, and human activity all at once. The challenge has been figuring out how to use these powerful, abstract fingerprints to predict landslides in specific local areas, especially when the data available is incomplete.

A team of researchers at Sun Yat-Sen University has tackled this problem by combining these advanced AI fingerprints with a clever way of handling incomplete data. In the real world, scientists know exactly where landslides have happened, but they rarely know for sure where they have not. A spot on a map without a recorded slide might be safe, or it might just be a place no one has looked at recently. Traditional computer models struggle with this uncertainty because they usually demand a clear list of "safe" spots to learn from. To get around this, the researchers used a method called positive-unlabeled learning. Instead of trying to guess which spots are safe, the model learns from the confirmed landslide locations and treats everything else as "unknown." This allows the system to learn the patterns of danger without needing a perfect list of safe zones. The researchers tested this approach in Hong Kong, feeding the model two types of information: the traditional physical measurements and the new AI-generated fingerprints from a system called AlphaEarth. They wanted to see if the AI fingerprints could stand alone, or if they worked best when paired with the old-school physical data.

The results showed that while the AI fingerprints alone could make a decent prediction, they were not the whole story. When the researchers combined the AI fingerprints with the traditional physical factors, the model became significantly more accurate and stable. This combined approach did something remarkable: it identified high-risk areas with much greater precision. In the most dangerous zones identified by the combined model, the density of actual landslides was about twenty-four times higher than the average background level. Even more impressively, the model managed to capture the majority of known landslide events while marking only 6.3 percent of the entire study area as high-risk. This level of focus is crucial for emergency planners, as it allows them to concentrate their resources on the most critical spots rather than spreading them thin over vast areas.

One of the most interesting findings was how the AI fingerprints helped the model avoid a common mistake. In crowded cities, traditional models often get confused by steep, man-made structures like highway embankments or the sides of tall buildings. Because these structures look steep on a map, the model would sometimes flag them as dangerous, even though they are built to be stable. The AI fingerprints, however, contain subtle information about the surface texture and material that tells the difference between a natural, unstable hillside and a solid, engineered wall. By including these fingerprints, the model stopped flagging these urban structures as high-risk, reducing false alarms while still correctly identifying the natural slopes that were truly in danger. This suggests that the AI is not replacing the physical measurements but is instead adding a layer of fine-grained detail that helps the model understand the context of the landscape.

The study also highlighted the importance of how these complex AI data are handled. The raw fingerprints generated by the AI are incredibly high-dimensional, meaning they contain hundreds of interconnected numbers that can confuse a computer model if not cleaned up first. The researchers found that if they fed the raw data directly into the system, the model's logic became unstable and harder to understand. By using a technique to simplify and organize these numbers before the model saw them, they ensured that the final predictions were not only accurate but also consistent and reliable. This step was essential for making the technology usable in the real world, where engineers need to trust that the map they are looking at makes sense.

Ultimately, this research offers a practical path forward for disaster management. It suggests a two-step approach for different situations. When a disaster strikes or when data is scarce, a model using only the AI fingerprints can provide a rapid, automated assessment to get a quick sense of the danger. When time and resources allow for a more detailed analysis, combining those fingerprints with traditional physical data creates a highly precise map that can guide engineering decisions and field investigations. The work demonstrates that while artificial intelligence is powerful, its greatest value often comes from working alongside human knowledge, filling in the gaps where traditional methods fall short and helping us see the hidden risks in our changing world.

Technical Summary: Leveraging Geospatial Foundation Model Embeddings for Local-Scale Landslide Susceptibility Mapping via Positive-Unlabeled Learning

Problem Statement
Landslide susceptibility mapping (LSM) at the local scale is critical for disaster management but faces two primary bottlenecks when deploying advanced artificial intelligence: effective feature representation and the reliability of training labels. Conventional LSM relies on physically interpretable conditioning factors (e.g., slope, lithology) which often fail to capture subtle surface textures and complex non-linear interactions. While deep learning offers stronger feature extraction, it typically demands large labeled datasets, which are often unavailable in local operational contexts. Furthermore, landslide inventories inherently suffer from a lack of reliable negative samples (non-landslides); the absence of a record does not guarantee stability, leading to biased training if conventional Positive-Negative (PN) sampling is used. Additionally, Geospatial Foundation Models (GeoFMs), while promising for transferable representations, introduce challenges regarding high-dimensional redundancy, multicollinearity, and the uncertainty of whether their embeddings can stand alone or must be fused with traditional physical factors.

Methodology
This study proposes a deployment-oriented workflow integrating AlphaEarth (a GeoFM) embeddings with Positive-Unlabeled (PU) learning to address these constraints. The study uses Hong Kong as a testbed, utilizing the Enhanced Natural Terrain Landslide Inventory (ENTLI) for positive samples (1,059 events from 2017–2024) and unlabeled background data.

The methodology proceeds through three core phases:

Feature Preprocessing and Optimization:
- Data Sources: 81 predictor variables were derived, including 64 AlphaEarth embedding bands (A00–A63), topographical, hydrological, geological, surface activity, and satellite-derived factors.
- Collinearity Diagnosis: Variance Inflation Factor (VIF) and Pearson correlation analyses revealed severe internal multicollinearity within the 64 embedding dimensions and redundancy between embeddings and specific satellite indices (NDVI, NDBI).
- Feature Sets: Three configurations were defined: Embedding-only, Traditional-only (all non-embedding factors), and Fused (all factors).
- Optimization: To mitigate multicollinearity, Principal Component Analysis (PCA) retaining 95% variance was applied to the Embedding-only and Fused sets. NDVI and NDBI were excluded from the Fused set due to redundancy with embeddings.
Modeling with Positive-Unlabeled Learning:
- Algorithm: The Positive and Background Learning with Constraints (PBLC) algorithm was employed. Unlike standard supervised learning, PBLC trains on confirmed positive landslides and unlabeled background samples, estimating the posterior probability of landslide occurrence while accounting for label frequency ( $c$ ).
- Architecture: A fully connected Artificial Neural Network (ANN) with three hidden layers (32, 16, 8 neurons) served as the classifier.
- Loss Function: A composite loss function ( $L_{total}$ ) combined Focal Loss (to handle class imbalance), a PBLC constraint term (assuming max landslide probability $P_{max} \approx 1$ ), and L2 regularization.
- Validation: A robust evaluation strategy involved 100 independent experimental runs (10 datasets $\times$ 10 training runs) to assess stability. Performance was measured using AUC and the $F_{pb}$ score (a proxy for F1-score in PU settings).
Interpretability and Stability Analysis:
- SHAP Analysis: SHapley Additive exPlanations were used to quantify feature contributions and interpret the decision logic of the high-dimensional embeddings.
- Stability Metrics: Spearman's Rank Correlation Coefficient was calculated across the 100 runs to evaluate the consistency of feature importance rankings.

Key Contributions

PU-Learning Deployment Strategy: The study establishes a workflow for engineering-geological susceptibility assessment under unreliable negative labels, utilizing PBLC to learn from positive inventories and unlabeled data without manual negative sampling.
Systematic Benchmark of GeoFM Embeddings: It provides a comparative analysis of Embedding-only, Traditional-only, and Fused feature sets, quantifying the baseline capability of GeoFMs and the marginal gains achieved through fusion.
Guidelines for High-Dimensional Embeddings: The research explicitly diagnoses redundancy in GeoFM embeddings and evaluates optimization strategies (PCA, regularization), offering practical guidance on balancing predictive accuracy with model robustness and interpretability.

Results

Performance: The Fused model (Embeddings + Traditional factors) achieved the highest median AUC and $F_{pb}$ scores with the narrowest interquartile ranges, indicating superior precision and stability. The Embedding-only model served as a competitive, low-preprocessing baseline but showed lower stability.
Spatial Selectivity: The Fused model delineated the most concentrated high-susceptibility zones. It captured the majority of landslides within only 6.3% of the study area, achieving a landslide density approximately 24 times the background level.
False Positive Reduction: In highly urbanized areas (e.g., Kowloon Peninsula), the Traditional-only model exhibited "elevation deception," assigning high susceptibility to artificial steep surfaces (embankments, elevated roads). The Fused model, leveraging GeoFM surface semantics, significantly reduced these false positives while retaining accuracy in natural mountainous terrain.
Feature Importance: SHAP analysis confirmed that while traditional factors (Slope, Lithology, Land Use) remained dominant, the inclusion of embeddings shifted the weighting toward geological and hydrological variables. Embedding contributions were more evenly distributed than physical factors.
Stability vs. Accuracy: Ablation studies revealed that removing PCA (using raw embeddings) yielded marginal gains in test metrics but caused a substantial decline in feature-importance stability (lower SHAP Spearman correlation) and increased overfitting gaps. Thus, dimensionality reduction is critical for robust deployment.

Significance and Claims
The paper claims that this workflow supports rapid engineering-geological screening and detailed susceptibility assessment under realistic label constraints. It proposes a dual-mode framework:

Level 1 (Rapid Response): Utilizing Embedding-only PU models for emergency situations or data-scarce regions where negative labels are unavailable and preprocessing time is limited.
Level 2 (Detailed Assessment): Utilizing Fused models with dimensionality reduction for post-disaster planning and site selection, ensuring results are statistically accurate, physically interpretable, and robust.

The study concludes that GeoFM embeddings do not replace traditional physical factors but complement them by providing fine-grained surficial semantics that resolve failures of purely physical logic (e.g., in urban environments). While the implicit spatio-temporal encoding in AlphaEarth embeddings streamlines the representation of cumulative environmental change, the authors modestly note this is a supplement to, not a replacement for, explicitly dynamic triggering variables (e.g., real-time rainfall). The framework is presented as a practical tool for prioritizing field investigations rather than a substitute for site-specific geotechnical analysis.

More like this