Using satellite imagery to track the impact of climate change on housing insurance

Company background

The user of this use case is an organization interested in better understanding and preventing the damages done on houses due to the issue of shrinkage-swelling of clay. They have varied activities linked to the monitoring of natural hazards. Since the project is still ongoing, they will remain anonymous. 

‍

Challenge

Over the last few years, the issue of Shrinkage-swelling of clay has increased significantly in France, due to changes in the climate. This phenomenon refers to the alternating movements (and often repeated over time) of shrinkage and swelling of the soil, respectively associated with the phases of drought and rehydration of soils. Soil with a high shrink-swell capacity, like clay, is problematic and is known as shrink-swell soil or expansive soil. The houses or buildings built in this area suffer heavy damage as this phenomenon deteriorates the foundations of houses. The repairs for these damages are very expensive and represent more than one billion euros of annual compensation for the insurers in France only.

The objective of this project was to see if the satellite imagery could replace their current solution based on weather stations in order to identify areas beforehand and be able to assess the risk at the individual house level, as their current solution was functioning at the city level. To achieve this goal, PREVENT partnered with SpaceSense to build a new solution. 

Solution

To solve this challenge, PREVENT wanted to see if there was a relation between the surface soil moisture level (that can be measured from satellite imagery) of an area and the risk of house foundation damages. They used a machine learning (ML) based method over a period of time to measure the soil moisture of an area and used this to construct monthly anomaly maps for 2019. If the map indicated an anomaly, then it meant that the anomalous area was significantly drier as compared to previous years. They then checked if these anomalous areas were correlated with the claims they measured during that year.

1. Building the soil moisture machine learning model

Sentinel 1’s Synthetic Aperture Radar (SAR) data was used for predicting the surface soil moisture. This radar imagery can be linked to soil water content near the surface because of the highly different dielectric properties of water and soil, with estimations of soil moisture directly linked to the reflected radar signal strength. Sentinel 1 data for this project was obtained using SpaceSense’s ARD module.

‍

To build the model they created a training dataset based on ground truth soil moisture data, obtained from a public network of sensors called International Soil Moisture Network (ISMN) and fused with satellite images obtained from Sentinel 1 data using the SpaceSense data fusion module. The fused data was then preprocessed, checked for data quality, resampled to 50m, and converted into a ML ingestible format using features from our data preparation for AI toolkit. A decision tree model was then trained on that dataset to measure soil moisture from Sentinel 1 images at a spatial resolution of 50m. 

This is how they built their own soil moisture monitoring system from satellite imagery.

‍

2. Creating the anomaly maps

Four cities of France were selected as areas of Interest (AOIs) with their corresponding historical claim locations for the purpose of testing and validating of the solution. They selected the year 2019 as the test year, as they had significant claims during that year.

Here is how they built the maps for each AOI:

1. Generate the soil moisture maps
2. Generate the anomaly maps
3. Correlation validation

‍

Here are the steps which they followed:

1. Generating soil moisture maps: They calculated the soil moisture levels for a time period of 4 years (the year of interest + 3 years before). They had a new soil moisture map every three days on average. Then they created monthly soil moisture values by simply averaging the individual soil moisture values, this is done to account for events such as rainfall.
2. Generate the anomaly maps: For each pixel of the AOI, the monthly averaged soil moisture pixel value is compared to the same months for the three previous years. Through a proprietary algorithm they generate an anomaly value for each pixel. The drier it is, the lower the value is going to be. The map is then normalized between -2 and 2. The red areas in the image below are the areas significantly drier than the previous years.
3. ‍Correlation validation: The third and last step was to see if the claims made in the AOI in 2019 (July) were located in “high-risk” red areas. You can see below in blue the locations of the claims and their location compared to the “high-risk” areas.

‍

The study showed that more than 80% of the claims made in July 2019 were situated in an area which had a minimum anomaly value of -0.25.

Result

Through its use of satellite imagery and the SpaceSense solution, PREVENT was able to identify a means to more accurately detect the high-risk areas for house foundations claims due to the  shrinkage-swelling phenomenon.

PREVENT is now considering implementing this system on a more regular basis to identify high-risk areas as the events are happening, and provide a more granular way to their members to assess the risk for their insurees.