Enterprise AI Analysis
Remote sensing-based landslide prediction and risk assessment using a hybrid CNN-LSTM deep learning model
Authored by: Fei Teng, Seyed Saeid Ekraminia, Amirreza Zarei & Yican Li
Landslide susceptibility assessment is essential for risk mitigation in regions affected by complex terrain and variable environmental conditions. This study proposes a hybrid deep learning framework based on a Convolutional Neural Network and Long Short-Term Memory (CNN-LSTM) architecture to integrate spatial and temporal information for landslide susceptibility mapping using remote sensing data. Spatial features were extracted from satellite imagery and digital elevation models, while temporal patterns were characterized using rainfall and reservoir-level time series. The proposed model was applied to Kerman Province, Iran, and evaluated using an independent test dataset of historical landslide events. The CNN-LSTM model achieved an accuracy of 95.6%, an F1-score of 93.5%, and an AUC of 0.98, outperforming traditional machine learning models and standalone deep learning approaches. The resulting susceptibility maps effectively identified high-risk zones consistent with historical landslide occurrences, demonstrating the benefit of integrating spatiotemporal information for regional-scale landslide susceptibility assessment.
Executive Impact & Key Findings
Leveraging a hybrid CNN-LSTM model, this research significantly advances landslide prediction, offering robust and reliable insights crucial for high-stakes enterprise decisions.
0
Overall Accuracy
0
F1-Score
0
AUC Score
0
AUC Improvement over RF
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Methodology & Innovation
Performance & Validation
Real-world Impact
Limitations & Future Work
This study introduced a novel hybrid CNN-LSTM deep learning framework for landslide susceptibility assessment. The core innovation lies in its ability to integrate both spatial and temporal information. The CNN component excels at extracting hierarchical spatial patterns from static remote sensing data (e.g., satellite imagery, DEMs), while the LSTM component effectively models temporal dependencies using time-series data like rainfall and reservoir levels. This end-to-end architecture enables joint spatiotemporal learning, overcoming limitations of models that consider these factors separately and providing a more comprehensive representation of landslide processes.
The proposed CNN-LSTM model demonstrated superior performance, achieving an accuracy of 95.6%, an F1-score of 93.5%, and an AUC of 0.98 on an independent test dataset. This significantly outperforms traditional machine learning models (SVM, DT, RF, MLP) and standalone deep learning approaches (CNN-only, LSTM-only). Specifically, the hybrid model showed an absolute AUC improvement of 6% points over the best-performing traditional benchmark (Random Forest), underscoring the benefits of integrated spatiotemporal modeling for enhanced predictive reliability.
Applied to Kerman Province, Iran, the model generated susceptibility maps that effectively identified high-risk zones, consistent with historical landslide occurrences. These maps classify 24.7% of the area as high susceptibility (primarily in zones with steep slopes, high cumulative rainfall, and significant reservoir-level fluctuations), 41.3% as medium, and 34.0% as low. This provides critical insights for land-use planning, early warning systems, and disaster risk management, enabling informed decisions to mitigate landslide risks in regions affected by complex terrain and variable environmental conditions.
Despite its robust performance, the study acknowledges several limitations. Model accuracy is dependent on the quality and resolution of remote sensing data, which can be affected by factors like cloud cover or sensor noise. The temporal component was limited to rainfall and reservoir-level fluctuations, lacking variables such as soil moisture or seismic activity. The landslide inventory might be incomplete or biased, and the framework's transferability to other regions remains untested. Future work should focus on applying spatially independent validation strategies, incorporating additional dynamic variables, and testing the framework across diverse geographical settings to enhance robustness and generalizability.
0.98
AUC Score on Independent Test Set
CNN-LSTM Model Implementation Workflow
Data Collection (Remote Sensing & Meteorological Data)
→
Data Preprocessing (Cleaning, Feature Extraction, Time-Series Structuring)
→
Model Construction (CNN, LSTM, Fully Connected Layers)
→
Model Training (Hyperparameter Tuning, Optimization, Regularization)
→
Model Validation (Metrics, Confusion Matrix)
→
Comparative Evaluation (Against ML/DL Benchmarks)
→
Results & Visualization (Susceptibility Maps, Feature Importance)
Performance Comparison of Models
| Model | Accuracy | Precision | Recall | F1-Score | AUC |
|---|---|---|---|---|---|
| Hybrid CNN-LSTM | 0.956 | 0.928 | 0.943 | 0.935 | 0.980 |
| Random Forest (RF) | 0.92 | 0.89 | 0.88 | 0.88 | 0.92 |
| Long Short-Term Memory (LSTM) (Standalone) | ~0.91 | ~0.88 | ~0.87 | ~0.88 | 0.832 |
| Convolutional Neural Network (CNN) (Standalone) | ~0.90 | ~0.87 | ~0.86 | ~0.87 | 0.821 |
| Support Vector Machine (SVM) | 0.90 | 0.87 | 0.86 | 0.87 | 0.88 |
| Multilayer Perceptron (MLP) | 0.88 | 0.84 | 0.83 | 0.83 | 0.90 |
| Decision Tree (DT) | 0.85 | 0.82 | 0.80 | 0.81 | 0.86 |
| AUC values for standalone CNN/LSTM are cited from Hamedi et al. (2022) in Table 1; other metrics are estimated from Figure 5. | |||||
Case Study: Landslide Susceptibility in Kerman Province, Iran
Kerman Province presents a challenging environment for landslide prediction due to its complex topography, variable lithology, active tectonics, and arid to semi-arid climate with low annual precipitation. These conditions favor both rainfall- and earthquake-induced landslides. The region's landscape includes mountain ranges, valleys, and plains, further complicated by human activities like mining and agriculture.
The hybrid CNN-LSTM model was applied to Kerman Province to generate reliable landslide susceptibility maps. By integrating remote sensing-derived spatial features (elevation, slope, aspect, TWI, SPI, ERG, distances to fault/river/road, land use, AAP) with temporal hydrological variables (rainfall, reservoir-level fluctuations), the model comprehensively characterized environmental controls on landslide occurrence.
The resulting maps identified 24.7% of the area as high susceptibility zones, primarily associated with steep slopes, high cumulative rainfall, and significant reservoir-level fluctuations. Another 41.3% was classified as medium susceptibility, and 34.0% as low. This spatial agreement with historical landslide occurrences demonstrates the model's effectiveness in identifying high-risk areas, providing crucial support for land-use planning and disaster risk reduction strategies in this vulnerable region.
Advanced ROI Calculator
Estimate your potential gains from integrating advanced AI for predictive analytics, tailored to your operational scale.
Estimated Annual Savings
$0
Annual Hours Reclaimed
0
Implementation Roadmap
A structured approach to integrating advanced AI, designed for seamless adoption and measurable results within your enterprise.
Phase 1: Data Integration & Preprocessing (Weeks 1-4)
Establish a comprehensive geospatial-temporal database by collecting and harmonizing high-resolution satellite imagery, DEMs, geological data, and meteorological time series. Implement automated data cleaning, feature extraction (e.g., NDVI, slope), and Min-Max normalization. Structure temporal data into fixed-length sequences suitable for LSTM input.
Phase 2: Model Adaptation & Training (Weeks 5-8)
Customize the hybrid CNN-LSTM architecture based on regional characteristics, balancing feature extraction and overfitting risks. Train the model using the prepared dataset, focusing on optimizing hyperparameters (learning rate, batch size) and applying regularization techniques (dropout, L2) to ensure stable convergence and generalization on Kerman's unique terrain.
Phase 3: Validation & Comparative Analysis (Weeks 9-10)
Rigorously validate the trained model using an independent test dataset, computing standard classification metrics (accuracy, precision, recall, F1-score, AUC). Conduct comparative evaluation against traditional machine learning and standalone deep learning models to quantitatively demonstrate the superior performance and benefits of the integrated spatiotemporal approach.
Phase 4: Susceptibility Mapping & Deployment (Weeks 11-12)
Apply the validated CNN-LSTM model to generate high-resolution landslide susceptibility maps for the entire Kerman Province. Reclassify continuous probabilities into categorical susceptibility levels (low, moderate, high). Prepare these maps for integration into existing GIS platforms, supporting evidence-based land-use planning and regional disaster risk management strategies.
Ready to Transform Your Operations?
Book a personalized consultation to explore how these advanced AI strategies can be customized for your enterprise.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat