Source-linked AI summary
Synergising Local Geo-Environmental Characteristics with Spatial Context for Enhancing Landslide Susceptibility Mapping
Yusen Cheng, Lei Fan, Qinfeng Zhu, Cheng Zhang, Yangyang Li, Ron Mahabir
TL;DR
Existing landslide susceptibility models either focus on local geo-environmental attributes or incorporate spatial context with potentially irrelevant neighbouring information. This study proposes LGSCF, which fuses both representations through feature-wise modulation across CNN branches. LGSCF-based models consistently improved susceptibility mapping performance and concentrated known landslides more strongly in very high susceptibility zones.
Problem
Pixel-based models neglect surrounding spatial context, whereas patch-based models may include spatially irrelevant pixels around target landslide locations.
Method
LGSCF uses dual CNN branches to encode central-pixel geo-environmental characteristics and image-patch spatial context, then integrates them through feature-wise modulation.
Results
LGSCF-based models consistently outperformed their corresponding baselines, with improved AUC stability and stronger concentration of landslides in very high susceptibility zones.
Takeaways & Limitations
LGSCF provides a flexible strategy for enhancing different CNN-based landslide susceptibility models by jointly using local geo-environmental and neighbourhood information.
Takeaways & Limitations
The experiments used a single study area and leave sensitivity to patch sizes and negative-sample selection strategies for future evaluation.
Abstract
from arXiv · showhide
Data-driven methods are widely used in landslide susceptibility mapping (LSM) because they can effectively model the complex relationships between landslides and geo-environmental conditions. Existing data-driven approaches generally follow two types of data representations. Pixel-based models focus solely on the geo-environmental characteristics of a specific landslide but neglect the influence of its surrounding environment. Patch-based models incorporate surrounding spatial context but may include pixels with weak or no spatial relevance to the target landslide location. To address this limitation, this study proposes a Local-Geo and Spatial Context Fusion (LGSCF) strategy, which synergises the geo-environmental characteristics of landslide points with their corresponding spatial context through a feature-wise modulation mechanism. We tested the LGSCF strategy by integrating it into several representative convolutional neural network (CNN) architectures, creating nine different LGSCF-based models. The study area covers approximately 2644 km2 across Jenai and Sinyi Townships in Nantou County, Taiwan, and the dataset comprises 5332 landslide samples and an equal number of non-landslide samples. The results show that LGSCF-based models consistently outperform their original versions, achieving F1-scores up to 87.09% and AUC values up to 0.9472. Furthermore, the susceptibility maps produced by LGSCF-based models show that known landslides are more accurately concentrated in "very high" susceptibility zones with fewer misclassifications. These findings demonstrate that our fusion strategy can significantly improve the accuracy of landslide susceptibility mapping.
1 Introduction
Landslide susceptibility mapping increasingly uses data-driven and CNN-based methods, but existing pixel- and patch-based representations incompletely balance local geo-environmental characteristics with surrounding spatial context. LGSCF addresses this gap by fusing both representations through feature-wise modulation.
- Motivation: Accurate landslide-prone-area prediction is important for disaster prevention because landslides threaten lives, infrastructure, and socioeconomic stability.Landslide occurrence is associated with gravity, intense rainfall, earthquakes, volcanic activity, and human interventions.
- Data-driven landslide susceptibility mapping: LSM delineates spatial zones of differing landslide susceptibility from relationships between historical landslides and geo-environmental conditioning factors.Its underlying assumption is that future landslides are more likely under conditions resembling those of past events.
- Existing representations: Pixel-based CNN representations learn nonlinear relationships from local conditioning factors but overlook surrounding spatial context.This limitation can produce unstable predictions near slope boundaries and weaken representation of terrain continuity and hydrological connectivity.
- Existing representations: Patch-based representations incorporate neighbourhood context but may add redundant background information and dilute the central pixel’s class-relevant contribution.Standard convolution and pooling progressively aggregate and downsample neighbourhood signals without explicitly prioritising critical information.
- Proposed strategy: LGSCF fuses a central-pixel local-geo branch with an image-patch spatial-context branch through feature-wise linear modulation.The framework can pair different CNN architectures while prioritising predictive features and reducing redundancy.
2 Study area and data
The study uses Jenai and Sinyi Townships in eastern Nantou County as its study area and constructs a landslide inventory from Taiwan’s 2022 annual map. Fifteen conditioning factors were selected based on regional LSM use, physical relevance, and consistent data availability, with some factors excluded because of data limitations or redundancy.
- 2.1 Study area description: Jenai and Sinyi Townships in eastern Nantou County form the study area, covering approximately 2,644 km².Nantou’s terrain ranges from western lowlands to eastern mountains with elevations approaching 3800 m.
- 2.2 Landslide inventory: 5332 landslide points were extracted from Taiwan’s 2022 Annual Landslide Inventory Map.The inventory was produced by visually interpreting SPOT satellite imagery, using a minimum mapping unit of 0.1 hectares and crown areas as representative points.
- 2.3 Conditioning factors: The selected LCFs reflect established regional LSM practice, physical relevance to landslide occurrence, and consistent spatial coverage across Nantou County.Selection also considers Nantou’s steep terrain, complex geology, monsoon rainfall, variable land cover, and road development.
- 2.3 Conditioning factors: Soil type was excluded because no region-wide dataset offered comparable spatial resolution and attribute consistency.Lithology and LULC were retained as surface-condition descriptors, while landform classes were omitted because related terrain attributes were already represented.
- 2.3 Conditioning factors: The 15 LCFs include terrain, hydrological, geological, vegetation, land-use, rainfall, and distance-related variables represented spatially across the study area.The mapped factors include elevation, slope, aspect, curvature, SPI, TWI, TRI, lithology, NDVI, LULC, rainfall, and distances to faults, roads, and rivers.
3 Methodology
The methodology combines local geo-environmental information with spatial context through LGSCF feature-level fusion, while using diagnostic sampling, training, and evaluation procedures for landslide susceptibility mapping.
- Preliminary diagnostics: Tolerance and variance inflation factor diagnostics were used to assess whether conditioning factors exhibited strong linear dependency.TOL below 0.1 and VIF above 10 were regarded as indicating multicollinearity.
- Data preparation: Frequency ratio, relative frequency, and prediction rate were used to select non-landslide samples while reducing potential mislabelling of unstable areas.The product of each factor’s prediction rate and class relative frequency guided sampling from very low susceptibility zones.
- Data preparation: 5332 non-landslide points were selected to match the 5332 landslide points and maintain class balance.Non-landslide points were randomly selected from very low susceptibility zones using frequency-ratio-based criteria.
- LGSCF architecture: Local-geo features generate multiplicative and additive modulation parameters for spatial-context features, producing f = γ⨀s + β before final classification.This feature-wise affine modulation avoids direct concatenation and conditions contextual interpretation on central-location characteristics.
- Model construction: Nine LGSCF-based models were constructed by pairing three spatial-context branch models with three local-geo branch models.Combinations within the same branch type were excluded because they were not considered complementary.
- Training and evaluation: Each model was tuned through a grid search over two batch sizes and five learning rates, with each configuration trained eight times using different random seeds.Configurations were ranked by mean validation AUC, and all other training settings were held constant for comparison.
- Training and evaluation: Model evaluation used validation-set classification metrics derived from a binary confusion matrix and ROC-based AUC.The ROC curve compares true-positive and false-positive rates across decision thresholds, while AUC measures overall discriminative ability.
4 Results
The results indicate that LGSCF-based models improve predictive performance over their corresponding baselines while preserving coherent susceptibility patterns. Model interpretation further shows that fusion redistributes the contributions of core conditioning factors rather than introducing a different dominant factor set.
- 4.1 Results of LCFs multicollinearity diagnostics: 15 LCFs showed insignificant multicollinearity, with TOL values of 0.487–0.953 and VIF values of 1.049–2.053.All TOL values exceeded 0.1 and all VIF values remained below 10, supporting their use in subsequent susceptibility analysis.
- 4.2 Configurations of best-performing models: Batch size 32 consistently yielded competitive performance across most model configurations.Configurations were evaluated using mean AUC across repeated runs and ranked by predictive performance.
- 4.3 Susceptibility map of the study area: Very high susceptibility zones concentrated in the southeast and northeast, while the western region was dominated by low and very low susceptibility levels.These broad patterns aligned with the mountainous terrain and observed landslide distribution, although models differed in local susceptibility assignments.
- 4.4 Model performance analysis: Spatial-context baselines outperformed local-geo baselines, with Models B and C exceeding 86% accuracy and F1-score and 0.945 AUC.Models M and N had precision above 88% but recall below 78%, indicating less balanced predictions.
- 4.4 Model performance analysis: LGSCF-based models consistently achieved higher F1-score and AUC than their corresponding baselines, including AUC values of 0.9473 for B+N and 0.9472 for C+P.Improvements were especially evident for weaker local-geo baselines and remained marginal but steady for stronger spatial-context baselines.
- 4.5 Analysis of the Importance of Factors: Distance to faults, DEM, distance to rivers, slope, and rainfall formed the dominant factor group across Models B, M, and B+M.LGSCF retained distance to faults as the leading factor while increasing the relative contributions of slope and rainfall; SHAP importance does not establish direct causality.
5 Discussions
The discussion evaluates LGSCF through susceptibility-class distributions, spatial robustness, branch interactions, and limitations. LGSCF generally concentrates landslides in very high-susceptibility zones and improves AUC, but benefits vary by branch pairing and study setting.
- 5.1 Comparison of landslide susceptibility maps: 66.84% of landslides reached the very high susceptibility class for Model C+N, while combined very low and low allocations fell to roughly 5%.Model C+N exceeded both baselines, and LGSCF models generally reduced lower-category misclassifications.
- Spatial robustness: All nine LGSCF-based models achieved higher mean AUC values than their corresponding spatial-context baselines in five-fold spatial evaluation.Model B+M remained best with a mean AUC of 0.9506; eight models had lower standard deviations than their baselines.
- Spatial robustness: Seven branch combinations showed statistically significant AUC improvements at p < 0.05, whereas Model B+P and Model C+N did not.The results support robustness, but improvement magnitude and statistical significance varied among branch combinations.
- Branch interactions: LGSCF gains did not follow a simple additive or rank-preserving pattern across baseline and local-geo branches.Model B produced the most pronounced error reduction despite not having the strongest baseline, suggesting that complementary feature sources and cross-branch modulation shaped performance.
- Limitations and future work: The framework's mechanism was qualitatively characterised, and stronger baselines such as Model C showed relatively modest gains.The authors call for systematic ablation and analysis of how LGSCF behaves with more advanced architectures.
- Limitations and future work: LGSCF was evaluated in a single study area, so cross-regional testing under different geological, climatic, and landslide conditions remains necessary.Additional uncertainty arises from patch-size selection, negative-sample selection, and imperfections in landslide inventories.
6 Conclusion
The conclusion presents LGSCF as a flexible strategy for combining local geo-environmental characteristics with neighbourhood spatial context in CNN-based LSM. Across baseline comparisons, the strategy improved predictive performance and susceptibility-class concentration, while its broader applicability still requires further validation.
- 6 Conclusion: LGSCF integrates dual-branch CNNs through feature-wise modulation to jointly capture local-geo characteristics and neighbourhood spatial context.The framework can be applied across different CNN architectures rather than defining one fixed network structure.
- 6 Conclusion: LGSCF-based models consistently improved F1-scores and AUC values over their single-branch counterparts across six baseline models.They also concentrated more landslides in the very high susceptibility category.
- 6 Conclusion: The results suggest that existing CNN-based LSM models can be enhanced by incorporating LGSCF.The conclusion frames this as an improvement in predictive accuracy within the evaluated setting.
- 6 Conclusion: The study's experiments were conducted in a single study area, and further validation across geological and climatic conditions is needed to assess generalizability.The authors also note that improvement magnitude and the factors driving it require further exploration.
Appendix A 1
Appendix A presents susceptibility maps for six base models and nine LGSCF-based model combinations. The figures organize these maps by model and corresponding fusion pairing.
- Appendix A 1: Six base-model susceptibility maps are shown for Models A, B, C, M, N, and P.The maps are labelled panels (a) through (f).
- Appendix A 1: Nine LGSCF-based susceptibility maps are shown for combinations pairing Models A, B, and C with Models M, N, and P.The combinations include A+M through C+P across panels (a) through (i).
Appendix B 7
Appendix B documents hardware configuration, training time, and computational complexity for baseline and LGSCF models. Training time reflects early stopping rather than a fixed epoch budget.
- Appendix B 7: Table B1 reports hardware configuration, training time, and computational complexity for baseline and LGSCF models.The comparison covers both model groups.
- Appendix B 7: Training terminated after 10 consecutive validation epochs without improvement, so reported time measures the path to early stopping.LGSCF models generally had higher computational complexity.