Team Carso

https://www.youtube.com/watch?v=KkyAbJT5ZmI

https://sffornasari.github.io/UKLandslideTracker/

Landslides are occurred due to the loss of stability of a slope. Stability disturbance may occur due to heavy rains and vibration of the ground (eg. earthquake). In this study, we focused on the landslides that are triggered by rain in the UK. To that we developed a machine learning model to have a probability distribution of observing a landslide for a given location. To do that, we used several parameters that are land cover, elevation, slope, soil water index, and rainfall.

Landslide database is retrieved from NASA Center for Climate Simulation. Even though the database cover many years we choose years between 2015 and 2017. It is due to fact that, the common time span of rainfall data and soil water index covers these years. Land cover is the type of usage a given position. Land cover dataset that we used has 23 discrete classes of a given position such as wetland, reforestation, deforestation, water expansion. Elevation data is also used as an input since heavy rains may not occur after certain elevation threshold. Furthermore, locations on sea level cannot produce a landslide unless the average elevation of the given area is below sea level. Slope information is one of the major parameters that is used. Landslides occur when the stability of a slope is vanished. Soil water index and rainfall data are used as time series. We choose 30 days of data (30 data points) to present their effect on the landslides. Soil water index is the moisture condition of the ground. Copernicus' Soil Water Index database provide European level data in daily basis for various depths. We used the depth of 2 meters if presented. However, when the data is absent for the given position, we checked the depths for 5m, 10m, 15m, 20m, 40m, 60m, and 100m. We also used the daily rainfall data from UK Centre from Ecology & Hydrology and retrieve 30 days of data before each event.

In the end we have 146 landslide events with the above mentioned information. This data provides us the positive examples. A machine learning algorithm is developed to have the probability of observing a landslide for a given conditions. To do that, we randomly picked points of about the same number on the UK to have negative examples (non landslides). The model has been developed in Tensorflow keeping its architectural complexity low due the small dataset available. It takes as input elevation and slope of the location of interest , its landcover class and the timeseries relative to the water content index and precipitations of the last 30 days. The time series are converted to a single value using 2 fully connected layers (one for each timeseries) which, along with the other inputs, are then fed to a deep neural network with a single hidden fully connected layer. The output is a single value representing the probability of a landslide at the location of interest. The model has then just 87 trainable parameters. The dense layers use a ReLu activation function. Adam is used as optimizer. The cross-entropy is used as loss function but also the accuracy is also monitored during the training phase.

Earthquakes are excluded from our study since, first, earthquakes are relatively rare in the UK and, second, including earthquakes to a machine learning approach may requires deeper knowledge about the earthquake-landslide interaction. Submarine landslides are also out of our interest.

The output of our model is an iOS app which allows users to follow possible landslide locations nearby them. To do that, grid sizes are approximately 1km in latitudes and longitudes. In the app, one can see two different type of alerts. The ones with orange color show low level landslide risk which means the the probability of observing a landslide in the given area are between 40% to 70%. The ones with red color show the high level risk which is the probability exceeds 70%. Our model uses daily time series which means that the alerts do not have hourly resolution.

We used Copernicus to retrieve the Soil Water Index Data which helped us to see the water accumulation of an area.

We also used NASA to have the landslide database for the study area.

It was a nice opportunity to create an end product only in 48 hours. We haven't started seen any possible databases and started from the ground zero and during the days we check multiple sources to see what type of metric we should use in order to solve the problem. We had changed our machine learning approaches since during the project we realized some of the databases are not useful, not applicable etc. Thanks to our team members that are coming from different backgrounds we were able to approach to problem in different approaches. We are happy to developed a preliminary model to the problem of landslides which can damage to settlements and infrastructures.

Soil Water Index Data from Copernicus programme and Global Land Service

Bauer-Marschallinger, B. ; Paulik, C. ; Hochstöger, S. ; Mistelbauer, T. ; Modanesi, S. ; Ciabatta, L. ; Massari, C. ; Brocca, L. ; Wagner, W. Soil Moisture from Fusion of Scatterometer and SAR: Closing the Scale Gap with Temporal Filtering. Remote Sensing 2018, 1019, 1030. DOI 10.3390/rs10071030

Landcover Data from Copernicus programme and Global Land Service

• Buchhorn, M.; Smets, B.; Bertels, L.; Lesiv, M.; Tsendbazar, N.-E.; Masiliunas, D.; Linlin, L.; Herold, M.; Fritz, S. (2020). Copernicus Global Land Service: Land Cover 100m: Collection 3: epoch 2015: Globe (Version V3.0.1) [Data set]. Zenodo. DOI:10.5281/zenodo.3939038
• Buchhorn, M.; Smets, B.; Bertels, L.; Lesiv, M.; Tsendbazar, N.-E.; Masiliunas, D.; Linlin, L.; Herold, M.; Fritz, S. (2020). Copernicus Global Land Service: Land Cover 100m: Collection 3: epoch 2016: Globe (Version V3.0.1) [Data set]. Zenodo. DOI:10.5281/zenodo.3518026
• Buchhorn, M.; Smets, B.; Bertels, L.; Lesiv, M.; Tsendbazar, N.-E.; Masiliunas, D.; Linlin, L.; Herold, M.; Fritz, S. (2020). Copernicus Global Land Service: Land Cover 100m: Collection 3: epoch 2017: Globe (Version V3.0.1) [Data set]. Zenodo. DOI:10.5281/zenodo.3518036

Rainfall Data from UK Centre from Ecology & Hydrology

Tanguy, M.; Dixon, H.; Prosdocimi, I.; Morris, D.G.; Keller, V.D.J. (2019). Gridded estimates of daily and monthly areal rainfall for the United Kingdom (1890-2017) [CEH-GEAR]. NERC Environmental Information Data Centre. https://doi.org/10.5285/ee9ab43d-a4fe-4e73-afd5-cd4fc4c82556

Elevation and Slope Data from E.O. Wilson Biodiversity Foundation

Amatulli, G., Domisch, S., Tuanmu, M.-N., Parmentier, B., Ranipeta, A., Malczyk, J., and Jetz, W. (2018) A suite of global, cross-scale topographic variables for environmental and biodiversity modeling. Scientific Data volume 5, Article number: 180040. DOI: doi:10.1038/sdata.2018.40.

Landslide Database from Global Gridded Landslide Inventory of NASA Center for Climate Simulation

Stanley, T. A., D. B. Kirschbaum, G. Benz, et al. 2021. "Data-Driven Landslide Nowcasting at the Global Scale." Frontiers in Earth Science, 9: [10.3389/feart.2021.640043]

#landslide, #artificialintelligence, #machinelearning, #prediction, #hazard, #safety, #mobileapp

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