Landslide Prediction and Mapping Using Geospatial and Neural Networks: A Game-Changer in Disaster Management

Geologically, Sikkim is distinguished by notable mountain structures, including the Main Central Thrust and Main Boundary Thrust zones, indicating a complex geological evolution. Additionally, the region is classified as Zone IV on India’s Seismic Zoning Map, underscoring the imperative for earthquake preparedness and sustainable urban planning.

To conduct the analysis the authors identified nine critical factors that contribute to the triggering of landslides, which are essential for understanding and mitigating these natural events are as follows:

1. Aspect (Slope direction) – Determines the influence of sun exposure and rainfall on slope stability.
2. Elevation – Higher altitudes may increase or decrease landslide risks depending on soil composition.
3. Land Use/Land Cover (LULC) – Deforestation, urbanization, and agriculture impact slope stability.
4. NDVI (Vegetation Index) – Measures how much vegetation is present (higher vegetation = more stability).
5. Rainfall – Heavy rains weaken soil structure, making slopes prone to landslides.
6. Slope Gradient – Steeper slopes are more vulnerable to landslides.
7. Soil Type – Different soil compositions have varying abilities to absorb water and resist erosion.
8. Earthquake Activity – Tremors can trigger landslides in already unstable slopes.
9. Lithology (Rock Composition) – Determines how well the ground resists weathering and erosion.

To create a dataset, the authors gathered a landslide inventory comprising 693 landslide samples and 695 non-landslide samples from Bhukosh, Geological Survey of India (GSI). This dataset was then divided into 70% for training and 30% for testing purposes. The analysis examined various activation functions, specifically ReLU (Rectified Linear Unit), Logistic (Sigmoid Function), Tanh (Hyperbolic Tangent), and Identity (Linear Activation). A comparative analysis was conducted to evaluate how these functions affect model performance, using metrics such as kappa, sensitivity, specificity, and AUC. The results deepen understanding of machine learning models and offer insights for enhancing methodologies in this domain. The hyperparameter settings used in the analysis are detailed as follows:

• solver = adam
• activation = relu; logistic; tanh; identity
• hidden layer = [50,50,50,50,50]
• max_iter = 500
• learning_rate = adaptive
• learning_rate = 0.01

The performance of various machine learning models was rigorously compared using multiple metrics to assess their effectiveness in distinguishing between landslide-prone and stable areas. One of the key metrics analyzed was the Area Under the Curve (AUC), which provides insight into the models’ ability to correctly classify instances. The AUC values for the models were as follows: MLPNN_logistic achieved an AUC of 0.757, making it the most effective model in this regard. The MLPNN_relu model followed closely with an AUC of 0.751, while MLPNN_tanh recorded an AUC of 0.750, indicating its similar utility. The least successful in terms of AUC was the MLPNN_identity, with a value of 0.736.

In terms of performance metrics, the MLPNN_relu model demonstrated a commendable balance between its training and testing scores, which were 0.742 and 0.740, respectively. Additionally, it exhibited high sensitivity, with a value of 0.908, reflecting its strong capability to identify true positive instances (i.e., landslide-prone areas). However, this model had lower specificity at 0.594, suggesting that it struggled to accurately identify true negatives (i.e., stable areas).

Conversely, the MLPNN_logistic model distinguished itself by attaining the highest kappa value of 0.504, an indication of good agreement between predicted and observed classifications. This model recorded training and testing scores of 0.760 and 0.748, respectively. It also demonstrated substantial sensitivity at 0.883, alongside a specificity of 0.631, underscoring its robustness in categorizing the data accurately.

The MLPNN_tanh model displayed consistent performance with training and testing scores of 0.745 and 0.741, respectively. It achieved a sensitivity of 0.875, indicating a strong capacity to identify landslide-prone areas, but had a slightly lower specificity at 0.624.

Lastly, the MLPNN_identity model, while achieving the highest specificity at 0.673, recorded the lowest sensitivity at 0.798. This resulted in a kappa value of 0.466, highlighting its diminished effectiveness in accurately identifying landslide-prone areas compared to the other models. Overall, the comparison of these models elucidates their varying strengths and weaknesses in identifying landslide susceptibility.

Furthermore, the study classifies the landslide susceptibility maps into distinct categories, thereby facilitating the identification of high-risk areas for the implementation of effective mitigation strategies. The spatial distribution of these classifications—very low, low, moderate, high, and very high—differs across various machine learning models. The MLPNN_relu model indicates that 26.79% of the study area is classified as having “very high” susceptibility, while only 5.74% is deemed “very low.” In the MLPNN_logistic model, the “high” category is the most prevalent at 25.96%, whereas “very low” is the least prevalent at 6.27%, suggesting a balanced focus in classification. The MLPNN_tanh model reveals that 27.46% of the area is categorized as having “low” susceptibility, with “very low” at 6.42%.

Conversely, the MLPNN_identity model tends to classify more terrain as having lower susceptibility, with the “moderate” category representing 46.81% and the “very low” category making up only 3.07%. This analysis underscores the influence of activation functions and model parameters on landslide susceptibility maps. Among the models examined, MLPNN_logistic and MLPNN_relu are identified as the most effective, with logistic activation exhibiting a slight overall advantage over ReLU. Nevertheless, ReLU is preferred due to its heightened sensitivity in mitigating false negatives when identifying landslide-prone areas.

The integration of geospatial technology and neural networks is fundamentally transforming the field of landslide prediction. This research demonstrates that the Multi-Layer Perceptron Neural Network (MLPNN) employing logistic activation functions yields the highest accuracy and reliability in landslide susceptibility mapping, thereby assisting communities in their preparedness for disasters prior to their occurrence. As technological advancements persist, the amalgamation of artificial intelligence-driven predictive modelling with real-time geospatial monitoring is poised to significantly mitigate the catastrophic effects of landslides on a global scale.