Interpretable Machine Learning Methods for Landslide Analysis
Landslides have the power to alter terrain, reshape ecosystems, damage anthropogenic structures, and destroy entire communities. On March 31, 2017, Mocoa fell victim to a devastating landslide that claimed 254 lives and left hundreds more missing. Given the region’s high frequency of landslides, con...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Published: |
Massachusetts Institute of Technology
2022
|
| Online Access: | https://hdl.handle.net/1721.1/143311 |
| _version_ | 1826211763070173184 |
|---|---|
| author | Gupta, Deepankar |
| author2 | Fernandez, John E. |
| author_facet | Fernandez, John E. Gupta, Deepankar |
| author_sort | Gupta, Deepankar |
| collection | MIT |
| description | Landslides have the power to alter terrain, reshape ecosystems, damage anthropogenic structures, and destroy entire communities. On March 31, 2017, Mocoa fell victim to a devastating landslide that claimed 254 lives and left hundreds more missing. Given the region’s high frequency of landslides, conducting a formal investigation of its landslides and the factors that make the region susceptible to them is urgent. In this project, we use machine learning to computationally study landslide detection, the likelihood of past landslides occurrence, and landslide susceptibility, or risk, in the Mocoa region. The region’s geographical and climate features make it elusive to remote-sensing technologies and difficult to survey. We combat the resulting data sparsity by carefully designing learning tasks and data pipelines for detection and susceptibility. We meaningfully extract 20 features with scientific or computational basis. We then provide a comprehensive evaluation of four different machine learning models: logistic regression (LR), decision tree (DT), random forest (RF), and convolutional neural network (CNN) on both the landslide detection and the landslide susceptibility tasks. All four models displayed performance that was significantly better than random on both tasks. CNN models achieved the highest classification accuracy in the area of interest (AOI) for both tasks, earning 87.3% for detection and 92.5% for susceptibility.We also probed all four types of models using multiple techniques to determine features important to their decision making. Across the landslide detection models, slope, aspect, and the presence of claystones were the most consistent important features for inferring past landslide likelihood. For landslide susceptibility, we found slope, aspect, distance from fault lines, and the presence of claystones to be the most consistent important features for inferring landslide risk. The reliance of the models on slope and aspect does not surprise due to landslides involving mass movement from high to low points of elevation. Less surprising, the presence of claystones also appears to be important to both landslide tasks. The connection between claystones and Mocoan landslides merits further investigation. The importance of distance from fault lines to susceptibility models suggests that seismic activity at fault lines is a key trigger for Mocoan landslides. |
| first_indexed | 2024-09-23T15:11:03Z |
| format | Thesis |
| id | mit-1721.1/143311 |
| institution | Massachusetts Institute of Technology |
| last_indexed | 2024-09-23T15:11:03Z |
| publishDate | 2022 |
| publisher | Massachusetts Institute of Technology |
| record_format | dspace |
| spelling | mit-1721.1/1433112022-06-16T03:10:48Z Interpretable Machine Learning Methods for Landslide Analysis Gupta, Deepankar Fernandez, John E. Freeman, William Massachusetts Institute of Technology. Department of Electrical Engineering and Computer Science Landslides have the power to alter terrain, reshape ecosystems, damage anthropogenic structures, and destroy entire communities. On March 31, 2017, Mocoa fell victim to a devastating landslide that claimed 254 lives and left hundreds more missing. Given the region’s high frequency of landslides, conducting a formal investigation of its landslides and the factors that make the region susceptible to them is urgent. In this project, we use machine learning to computationally study landslide detection, the likelihood of past landslides occurrence, and landslide susceptibility, or risk, in the Mocoa region. The region’s geographical and climate features make it elusive to remote-sensing technologies and difficult to survey. We combat the resulting data sparsity by carefully designing learning tasks and data pipelines for detection and susceptibility. We meaningfully extract 20 features with scientific or computational basis. We then provide a comprehensive evaluation of four different machine learning models: logistic regression (LR), decision tree (DT), random forest (RF), and convolutional neural network (CNN) on both the landslide detection and the landslide susceptibility tasks. All four models displayed performance that was significantly better than random on both tasks. CNN models achieved the highest classification accuracy in the area of interest (AOI) for both tasks, earning 87.3% for detection and 92.5% for susceptibility.We also probed all four types of models using multiple techniques to determine features important to their decision making. Across the landslide detection models, slope, aspect, and the presence of claystones were the most consistent important features for inferring past landslide likelihood. For landslide susceptibility, we found slope, aspect, distance from fault lines, and the presence of claystones to be the most consistent important features for inferring landslide risk. The reliance of the models on slope and aspect does not surprise due to landslides involving mass movement from high to low points of elevation. Less surprising, the presence of claystones also appears to be important to both landslide tasks. The connection between claystones and Mocoan landslides merits further investigation. The importance of distance from fault lines to susceptibility models suggests that seismic activity at fault lines is a key trigger for Mocoan landslides. M.Eng. 2022-06-15T13:11:37Z 2022-06-15T13:11:37Z 2022-02 2022-02-22T18:32:05.467Z Thesis https://hdl.handle.net/1721.1/143311 In Copyright - Educational Use Permitted Copyright MIT http://rightsstatements.org/page/InC-EDU/1.0/ application/pdf Massachusetts Institute of Technology |
| spellingShingle | Gupta, Deepankar Interpretable Machine Learning Methods for Landslide Analysis |
| title | Interpretable Machine Learning Methods for Landslide Analysis |
| title_full | Interpretable Machine Learning Methods for Landslide Analysis |
| title_fullStr | Interpretable Machine Learning Methods for Landslide Analysis |
| title_full_unstemmed | Interpretable Machine Learning Methods for Landslide Analysis |
| title_short | Interpretable Machine Learning Methods for Landslide Analysis |
| title_sort | interpretable machine learning methods for landslide analysis |
| url | https://hdl.handle.net/1721.1/143311 |
| work_keys_str_mv | AT guptadeepankar interpretablemachinelearningmethodsforlandslideanalysis |