FL-MHSM: Spatially-adaptive Fusion and Ensemble Learning for Flood-Landslide Multi-Hazard Susceptibility Mapping at Regional Scale
| Authors | Aswathi Mundayatt & Jaya Sreevalsan-Nair |
| Year | 2026 |
| Field | Machine Learning |
| arXiv | 2604.16265 |
| Download | |
| Categories | cs.LG |
Abstract
Existing multi-hazard susceptibility mapping (MHSM) studies often rely on spatially uniform models, treat hazards independently, and provide limited representation of cross-hazard dependence and uncertainty. To address these limitations, this study proposes a deep learning (DL) workflow for joint flood-landslide multi-hazard susceptibility mapping (FL-MHSM) that combines two-level spatial partitioning, probabilistic Early Fusion (EF), a tree-based Late Fusion (LF) baseline, and a soft-gating Mixture of Experts (MoE) model, with MoE serving as final predictive model. The proposed design preserves spatial heterogeneity through zonal partitions and enables data-parallel large-area prediction using overlapping lattice grids. In Kerala, EF remained competitive with LF, improving flood recall from 0.816 to 0.840 and reducing Brier score from 0.092 to 0.086, while MoE provided strongest performance for flood susceptibility, achieving an AUC-ROC of 0.905, recall of 0.930, and F1-score of 0.722. In Nepal, EF similarly improved flood recall from 0.820 to 0.858 and reduced Brier score from 0.057 to 0.049 relative to LF, while MoE outperformed both EF and LF for landslide susceptibility, achieving an AUC-ROC of 0.914, recall of 0.901, and F1-score of 0.559. GeoDetector analysis of MoE outputs further showed that dominant factors varied more across zones in Kerala, where susceptibility was shaped by different combinations of topographic, land-cover, and drainage-related controls, while Nepal showed a more consistent influence of topographic and glacier-related factors across zones. These findings show that EF and LF provide complementary predictive behavior, and that their spatially adaptive integration through MoE yields robust overall predictive performance for FL-MHSM while supporting interpretable characterization of multi-hazard susceptibility in spatially heterogeneous landscapes.
Engineering Breakdown
Plain English
This paper addresses a critical gap in natural disaster risk assessment by proposing a deep learning workflow for joint flood-landslide susceptibility mapping in Kerala, India. The key problem: existing approaches treat hazards independently and use spatially uniform models, missing how floods and landslides interact geographically. The authors combine two-level spatial partitioning with probabilistic Early Fusion (data-level integration) and a soft-gating Mixture of Experts model to predict multi-hazard risk across large areas while preserving local geographic variation. Early Fusion remained competitive with tree-based Late Fusion baselines while improving flood recall, demonstrating that joint modeling captures hazard dependencies better than independent predictions.
Core Technical Contribution
The core innovation is a hierarchical deep learning architecture that explicitly models spatial heterogeneity for multi-hazard prediction through two mechanisms: spatial partitioning into zones that respect geographic variability, and soft-gating Mixture of Experts that routes different regions through specialized sub-networks based on learned geographic features. Unlike prior MHSM work treating hazards independently or assuming uniform spatial models, this approach uses probabilistic Early Fusion to capture cross-hazard dependence at the feature level while the MoE gating mechanism adapts predictions to local terrain and environmental context. The use of overlapping lattice grids enables efficient large-area inference without losing fine-grained spatial detail. This is distinct from generic multi-task learning because the gating is spatially informed rather than task-agnostic.
How It Works
The pipeline begins with input features (topography, land cover, climate, geology) for a large geographic region divided into overlapping spatial partitions via lattice grids. Early Fusion concatenates flood and landslide-related features and passes them through shared dense layers to learn joint representations that capture dependencies—for example, saturation effects that make steep slopes both flood and landslide-prone simultaneously. In parallel, a tree-based baseline (likely XGBoost or similar) creates Late Fusion predictions by training independent hazard models then combining outputs. The final Mixture of Experts model learns soft gating weights—a neural network that outputs normalized weights for multiple expert sub-networks based on spatial location and feature context. Each expert specializes in predicting hazard susceptibility for a specific geographic or feature subspace; the gating network learns which expert to trust in each region, effectively creating spatially adaptive predictions. The overlapping lattice structure allows batching and parallelization: predictions slide across a large area using overlapping windows, reducing memory overhead while maintaining prediction continuity at tile boundaries.
Production Impact
For disaster risk teams deploying susceptibility maps at national scale, this approach directly reduces false negatives in hazard-prone regions by capturing multi-hazard interactions—critical because compound flooding and landslide events disproportionately affect communities. In production, you'd integrate this as a preprocessing stage in disaster early warning systems: feed satellite imagery and terrain data into the partitioned DL workflow monthly or seasonally, output pixel-level hazard probabilities that feed into evacuation routing and resource allocation algorithms. The spatial partitioning and lattice-grid design make this practical for resource-constrained agencies—you can process large regions in parallel on commodity GPUs without needing massive batch sizes or storing full country-sized tensors in memory. Trade-offs include: increased training complexity compared to single-hazard models (requires labeled data for both floods and landslides jointly), higher inference latency than independent models due to MoE routing overhead, and need for careful hyperparameter tuning of partition size and gate network architecture. Integration into existing GIS/hazard mapping pipelines is straightforward since outputs are standard raster formats.
Limitations and When Not to Use This
The paper abstracts away several production realities: it assumes high-quality labeled multi-hazard data exists (in practice, flood and landslide labels often have different quality, temporal coverage, and spatial resolution, creating label noise). The spatial partitioning strategy is fixed at design time—the paper doesn't explore whether learned partition boundaries might outperform fixed lattice grids, or how to choose partition size without extensive hyperparameter search. Kerala's specific topography and climate may not generalize to other regions with different hazard regimes (e.g., arid regions with rare but catastrophic floods); the authors don't test cross-domain transfer. The paper stops short of validating whether soft-gating MoE actually provides better uncertainty quantification than simpler ensemble methods or Bayesian approaches, which is critical for risk-averse applications where over-confident wrong predictions cause real harm. Computational cost of soft-gating (routing through multiple experts) versus simpler late fusion isn't quantified, leaving unclear whether the performance gains justify deployment complexity.
Research Context
This work extends multi-hazard susceptibility mapping (MHSM) literature, which traditionally used statistical models (logistic regression, random forest) treating hazards independently. It builds on recent deep learning success in single-hazard mapping and incorporates architectural innovations from multi-task learning and Mixture of Experts literature—soft gating specifically derives from Shazeer et al.'s sparse MoE work and extensions in large language models, adapted here to spatial modeling. The two-level partitioning approach echoes hierarchical spatial modeling in climate science and land-use mapping, where explicit geographic stratification often outperforms uniform global models. This opens a research direction toward spatially-aware routing architectures for other geographic risk modeling (wildfire, drought, heatwave) and raises questions about learning optimal partition strategies end-to-end rather than fixing them beforehand.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::