Uncertainty of interpretability in Landslide Susceptibility Mapping: A Comparative Analysis of Statistical, Machine Learning, and Deep Learning Models

Overview

• This study investigates the interpretability of statistical, machine learning (ML), and deep learning (DL) models in predicting landslide susceptibility.
• It incorporates various interpretation methods and two types of input factors: a comprehensive set of 19 contributing factors and a dedicated set of 9 triggering factors.
• The convolutional neural network model achieved the highest accuracy, while Extreme Gradient Boosting and Support Vector Machine also demonstrated strong predictive capabilities.
• The interpretability of predictions varied among different models, particularly when using the broader set of 19 contributing factors.

Plain English Explanation

Landslides can cause significant damage and loss of life, so being able to predict where they are likely to occur is crucial. This study looked at different types of machine learning models, including deep learning, to see how well they could predict the likelihood of landslides. They used two sets of factors as inputs: a comprehensive set of 19 factors that are statistically related to landslides, and a more focused set of 9 factors that are directly linked to triggering landslides.

The deep learning model, specifically a convolutional neural network, was the most accurate at predicting landslide susceptibility. Other advanced machine learning models, like Extreme Gradient Boosting and Support Vector Machines, also performed well and outperformed more traditional statistical models. This suggests that these complex models are able to capture the intricate relationships between the input factors and where landslides are likely to occur.

However, the researchers found that the interpretability of the model predictions varied, especially when using the broader set of 19 factors. Interpretability refers to how well we can understand and explain the reasoning behind the model's predictions. Some interpretation methods, like SHAP, LIME, and DeepLIFT, led to different results. Using the more focused set of 9 triggering factors improved the interpretability of the models, as the key factors identified were more consistent across different models and aligned better with field investigations.

Technical Explanation

This study compared the performance and interpretability of statistical, machine learning (ML), and deep learning (DL) models in predicting landslide susceptibility. The researchers incorporated various interpretation methods, such as SHAP, LIME, and DeepLIFT, and used two types of input factors: a comprehensive set of 19 contributing factors and a dedicated set of 9 triggering factors.

The models evaluated included logistic regression, random forest, Extreme Gradient Boosting, support vector machine, and convolutional neural network. The convolutional neural network achieved the highest accuracy, with 0.8447 using the 19 factors and 0.8048 using the 9 factors. Extreme Gradient Boosting and Support Vector Machine also demonstrated strong predictive capabilities, outperforming the conventional statistical models.

The interpretability of the model predictions varied, particularly when using the broader set of 19 contributing factors. The interpretation methods led to different results in terms of the key factors identified as influential. Using the more focused set of 9 triggering factors improved the interpretability, as the key factors identified were more consistent across different models and aligned better with the findings of field investigation reports.

Critical Analysis

The study provides valuable insights into the performance and interpretability of different machine learning and deep learning models for landslide susceptibility mapping. The use of comprehensive and focused sets of input factors, as well as the incorporation of various interpretation methods, is a strength of the research.

However, the study does not delve into the potential limitations or caveats of the research. For example, the study is limited to a specific geographical region, and the generalizability of the findings to other areas may need further investigation. Additionally, the study does not discuss the potential biases or uncertainties inherent in the input data or the interpretation methods used.

Furthermore, the study could have explored the tradeoffs between model accuracy and interpretability in more depth. While the deep learning model achieved the highest accuracy, the interpretability of its predictions was more variable compared to the other models. This raises questions about the balance between model performance and the ability to explain the reasoning behind the predictions, which is an important consideration in real-world applications.

Future research could also investigate the impact of feature engineering and selection on the interpretability of the models, as well as explore alternative interpretation methods or combinations of techniques to enhance the consistency and reliability of the interpretations.

Conclusion

This study demonstrates the potential of advanced machine learning and deep learning models in predicting landslide susceptibility, with the convolutional neural network achieving the highest accuracy. However, the interpretability of the model predictions varied, particularly when using a comprehensive set of input factors. Focusing on a dedicated set of triggering factors improved the interpretability, suggesting a tradeoff between model performance and the ability to explain the reasoning behind the predictions.

The findings highlight the importance of considering both model performance and interpretability in real-world applications, such as enhancing traffic incident management with large language models or improving streamflow forecasting. By understanding the strengths and limitations of different modeling approaches, researchers and practitioners can develop more robust and transparent landslide susceptibility mapping tools to support disaster risk reduction and mitigation efforts.