Slideland 5

https://www.youtube.com/watch?v=S-CyE0I_KDA

https://www.slideland.earth/

https://github.com/slideland

We have developed an interactive tool for both researchers and affected communities. It consists of both an interactive website, with useful APIs, as wells as a repository with the source code and scripts which for gathering satellite data from historic landslide locations, which can be used to train machine learning models.

The site shows an interactive 3D map of the world where risk zone for landslides are highlighted. Users can search for locations to find the areas they live near or plan to travel by. They can thus, minimize their risk by choosing safer paths or by preparing their homes for what could happen. If the worst comes to the worst, users can report landslide locations. This will be published on the site for others to see, but also for the validation of connected models.

The sites incorporate an API so researchers can connect their landslide susceptibility models. In this way at-risk communities have a single place for all information. This also allows researchers to compare predictions with their previous models or to others models. The site includes a leader board of the best models. The accuracy is determined by the true positive rate of predictions of landslides which occurred after the model was connected, as not to validate on training data. Then scoring is weighted based on log diagnostic odds ratio with bayesian updating, to be able to compare models with different number of validation points. It also weighs in the precision of models as not to give a high score to a model that predicts the whole globe to be a risk zone.

In the repository we also publish a script to download satellite data, from Google Earth Engine [6], of historic landslide locations. With minimal changes it can be configured to download data from any of the many maps provided in the Earth Engine Data Catalog. This can be used to create datasets for machine learning models. We propose one such model, but did not have the time to finish it. As of now, the script train_autoencoder.py, trains an autoencoder to compress elevation images of historic landslide locations. The method proposed in [4] could be used to find similar sites, prone to landslides, by feeding in new map images and looking at reconstruction probabilities. This could be done for the whole globe and by adding a rain accumulation risk function this could become one of the models to be evaluated by our site. If successful the model could also be expanded to incorporate other types of maps such as land use or lithology. According to [5] these are the most important parameters apart from rainfall and landform.

We used the NASA landslide susceptibility map [1][2][3] as a default data source for our generic landslide warning map. Although it already comes with an editor, we felt that the ability to see other models, compare them and easily add new ones were a worthy addition. Still the NASA landslide susceptibility model has a great true positive rate, so we choose to use it as the standard model.

We wanted to choose a challenge where we believed we could have a real impact on the world. After some research we realized that the state of current landslide research is far from perfect and that this affects the everyday life of many people around the world. We felt that our team had the right background to tackle this problem: a mathematician, a designer, two programmers, and a machine learning engineer. 

As we are all located in different cities, this became our first virtual hackathon. The experience was wonderful, we mixed a long period of intense work shifts with periods of hysterical laughter. In the end, we did not finish the machine learning prediction model, but this was just a small part of our project and we are really proud of what we have accomplished. But of course, we are just standing on the shoulders of others. To create such a tool in just two days would never be possible without the open source research, software and datasets. We therefore want to give a huge thank you to Nasa and everyone else who contributes to making research open.

[1] Emberson, R., D. Kirschbaum, and T. Stanley. 2020. "New global characterisation of landslide exposure." Natural Hazards and Earth System Sciences, 20 (12): 3413-3424. doi:10.5194/nhess-20-3413-2020

[2] Stanley, T., and D. B. Kirschbaum (2017), A heuristic approach to global landslide susceptibility mapping, Nat. Hazards, 1–20, doi:10.1007/s11069-017-2757-y

[3] Kirschbaum, D. and Stanley, T. (2018), Satellite‐Based Assessment of Rainfall‐Triggered Landslide Hazard for Situational Awareness. Earth's Future. . doi:10.1002/2017EF000715

[4] An, Jinwon, and Sungzoon Cho. "Variational autoencoder based anomaly detection using reconstruction probability." Special Lecture on IE 2.1 (2015): 1-18.

[5] Dou, J., Yunus, A.P., Bui, D.T. et al. Improved landslide assessment using support vector machine with bagging, boosting, and stacking ensemble machine learning framework in a mountainous watershed, Japan. Landslides 17, 641–658 (2020). https://doi.org/10.1007/s10346-019-01286-5

[6] Gorelick, N., Hancher, M., Dixon, M., Ilyushchenko, S., Thau, D., & Moore, R. (2017). Google Earth Engine: Planetary-scale geospatial analysis for everyone. Remote Sensing of Environment.

Development resources

Github

Python

MongoDB

TypeScript

Heroku

Vercel

#opensource #visualization #machinelearning #API #landslide

This project has been submitted for consideration during the Judging process.