5:15 PM - 7:15 PM
[HRE13-P02] Application of using CNN for Evaluating gentle slope collapse Risk
1. Introduction
In recent years, the frequent occurrence of heavy rainfall due to climate change has led to slope failures that cause significant social and economic damage. Traditionally, for deep-seated landslide risk evaluation, a model employing CNN with eight types of terrain characteristic images has been proven effective (Kikuchi et al., 2023). However, because there is relatively abundant training data for shallow landslides, it is expected that this method can be transferred to evaluate shallow landslide risks. Furthermore, from the perspectives of data acquisition and computational cost, comparing it with a simpler CI map-only input model is an important issue. This study aims to compare the performance of the two models and establish a risk evaluation method for shallow landslides that is suitable for practical applications.
2. Analysis
Method In this study, two types of CNN models were constructed, differing only in the input images. One model combines eight types of terrain characteristic indices calculated from the DEM, while the other model uses only the CI map provided by Chuo Development. Each input image was divided into tiles of 100 × 100 pixels, and 100 models were explored using the structural automatic search function of the Neural Network Console (NNC).
3. Results
Evaluation on both validation and test areas revealed that the eight-indicator model exhibited a high Recall (detection rate), but this came with a trade-off in Accuracy due to an increase in False Positives. In contrast, although the CI map model showed slightly lower Recall, its Accuracy was comparable to that of the eight-indicator model (0.73), clearly demonstrating the operational advantages of simplified input data. Furthermore, because the CI map model greatly reduces both computational cost and data acquisition effort, it is suggested as a method well suited for large-scale landslide risk evaluation. On the other hand, when Recall is prioritized, the eight-indicator model tends to detect shallow landslide hazardous areas at a higher probability, which can be advantageous in scenarios where a conservative evaluation is required. A comparison between the two models shows that while the eight-indicator model offers high-precision judgments, it faces challenges related to data preparation and computational load. In contrast, the CI map model is easier to implement, although it suffers from a slight reduction in prediction accuracy; hence, selecting the appropriate model based on the application scenario is crucial. Additionally, the results of this study indicate that the CNN-based approach—originally focused on deep-seated landslides— is also applicable to the risk evaluation of shallow landslides, suggesting that the utility of CNNs can be expanded to the analysis of a broader range of landslide phenomena.
4. Conclusion
This study applied a CNN model, originally developed for deep-seated landslide risk evaluation, to the risk evaluation of shallow landslides by comparing a model using eight types of terrain characteristic indices with a CI map-only model. The results confirmed that the eight-indicator model, while showing high Recall, produced a significant number of false detections, whereas the CI map model, despite its simpler input, achieved comparable Accuracy. These findings suggest that, for practical applications, a CI map-only approach can provide a viable method for risk evaluation. Moreover, the study demonstrates that a CNN-based method is applicable to the evaluation of shallow landslides, highlighting the potential to extend CNN techniques to analyze a wider range of landslide phenomena. Future work will focus on expanding the target areas, adjusting the tile size, and reducing misclassifications through techniques such as ensemble learning to develop models with higher accuracy and broader applicability.
In recent years, the frequent occurrence of heavy rainfall due to climate change has led to slope failures that cause significant social and economic damage. Traditionally, for deep-seated landslide risk evaluation, a model employing CNN with eight types of terrain characteristic images has been proven effective (Kikuchi et al., 2023). However, because there is relatively abundant training data for shallow landslides, it is expected that this method can be transferred to evaluate shallow landslide risks. Furthermore, from the perspectives of data acquisition and computational cost, comparing it with a simpler CI map-only input model is an important issue. This study aims to compare the performance of the two models and establish a risk evaluation method for shallow landslides that is suitable for practical applications.
2. Analysis
Method In this study, two types of CNN models were constructed, differing only in the input images. One model combines eight types of terrain characteristic indices calculated from the DEM, while the other model uses only the CI map provided by Chuo Development. Each input image was divided into tiles of 100 × 100 pixels, and 100 models were explored using the structural automatic search function of the Neural Network Console (NNC).
3. Results
Evaluation on both validation and test areas revealed that the eight-indicator model exhibited a high Recall (detection rate), but this came with a trade-off in Accuracy due to an increase in False Positives. In contrast, although the CI map model showed slightly lower Recall, its Accuracy was comparable to that of the eight-indicator model (0.73), clearly demonstrating the operational advantages of simplified input data. Furthermore, because the CI map model greatly reduces both computational cost and data acquisition effort, it is suggested as a method well suited for large-scale landslide risk evaluation. On the other hand, when Recall is prioritized, the eight-indicator model tends to detect shallow landslide hazardous areas at a higher probability, which can be advantageous in scenarios where a conservative evaluation is required. A comparison between the two models shows that while the eight-indicator model offers high-precision judgments, it faces challenges related to data preparation and computational load. In contrast, the CI map model is easier to implement, although it suffers from a slight reduction in prediction accuracy; hence, selecting the appropriate model based on the application scenario is crucial. Additionally, the results of this study indicate that the CNN-based approach—originally focused on deep-seated landslides— is also applicable to the risk evaluation of shallow landslides, suggesting that the utility of CNNs can be expanded to the analysis of a broader range of landslide phenomena.
4. Conclusion
This study applied a CNN model, originally developed for deep-seated landslide risk evaluation, to the risk evaluation of shallow landslides by comparing a model using eight types of terrain characteristic indices with a CI map-only model. The results confirmed that the eight-indicator model, while showing high Recall, produced a significant number of false detections, whereas the CI map model, despite its simpler input, achieved comparable Accuracy. These findings suggest that, for practical applications, a CI map-only approach can provide a viable method for risk evaluation. Moreover, the study demonstrates that a CNN-based method is applicable to the evaluation of shallow landslides, highlighting the potential to extend CNN techniques to analyze a wider range of landslide phenomena. Future work will focus on expanding the target areas, adjusting the tile size, and reducing misclassifications through techniques such as ensemble learning to develop models with higher accuracy and broader applicability.