Developing Machine Learning tools for the automatic interpretation of InSAR data

Hundreds of large cities worldwide are sinking; and this will get worse as by 2050 when almost 70% of the world’s population is set to live in megalopolises, the majority of these are in low lying coastal areas. At the same time sea levels are rising. According to the World Economic Forum, several of the globe’s cities, including New York, Dhaka, London and Bordeaux, could be partially or totally submerged by 2050-2100. As a city grows the environment is put under additional pressure and this often leads to subsidence. land less suitable for building upon is developed, in low lying coastal regions these areas are often poorly consolidated recent superficial deposits. Loading of such deposits causes consolidation which adds to subsidence resulting from increased groundwater abstraction required for industry and to support a growing population.

In order to mitigate against the effects of subsidence it is imperative to understand the subsidence; its location, magnitude, timing and crucially the underlying cause. InSAR offers the ability to understand the spatial extent, magnitude and timing and when integrated with in-situ data the cause can be determined. However, this interpretation process can take a significant amount of time. With the advent of continental scale InSAR data, such as the European Ground Motion Service, and automated online processing facilities such as COMETS LICSBAS system InSAR data is becoming far easier to access. This means huge volumes of data are generated and therefore automated methods are required to extract not only the areas of ground motion but also to indicate the underlying cause of the motion.

To this end we have been using integrated time series of optical and InSAR data for areas of rapid urban growth to understand the cause of subsidence. Combination of interpretations with expected patterns of subsidence derived from models of groundwater abstraction and ground loading allow us to separate subsidence signals from these causational factors. In turn this enables the generation of characteristic time series of subsidence that we would expect to see as a result of each process. Such characteristic time series form libraries that will be the basis for machine learning to automatically interpret the InSAR data.

We will present the creation of such subsidence libraries and illustrate the process with examples from Hanoi, Kuala Lumpur and Bandung. We will also present the machine learning method where a fully automated approach using Seasonal and Trend decomposition allows trends (such as stable, linear subsidence, non-linear subsidence and seasonal) within the InSAR time series to be identified and grouped into common trend behaviours. Metrics based on derived trends also allow the ‘strength’ of certain components (such as seasonal signals) to be automatically assessed.