The central task of detecting and identifying disruptions to critical infrastructure requires automated and robust methods to distinguish between the desired state and abnormal conditions. The methods used must be unaffected by seasonal or weather-related changes (fallen leaves, rain, snow) to ensure operation regardless of the geographical zone or the central task of detecting and identifying disruptions to critical infrastructure requires automated and robust methods to distinguish between the desired state and abnormal conditions. The methods used must be unaffected by seasonal or weather-related changes (fallen leaves, rain, snow) to ensure operation regardless of the geographical zone or time of year and to reduce the rate of false positives. At the same time, it is important that the methods are sensitive enough to minimise the rate of false negatives, i.e. failures to detect a disruption.
One of the methods explored by IIMEO is change detection. The basic principle is that (SAR) images of certain features in an area of interest (e.g. railway lines) are acquired at different times and compared to each other. Depending on the amount of data to be processed, the available resources and the urgency of the information request, change detection can take place directly on the acquisition platform or on the ground.
In the framework of IIMEO, two kinds of artificial-intelligence-based approaches to change detection are being evaluated. The first is based on thresholding techniques with low computational complexity but potentially lower accuracy; the second is more computationally intensive and utilises deep learning neural network architectures to achieve high accuracies. These types of change detection methods are state of the art but have been developed for different applications and SAR sensors.
The second method being investigated as part of IIMEO is anomaly detection. Contrasting to change detection, anomaly detection identifies data with features not explainable by the probability distribution characterising the desired state, i.e. an undisrupted infrastructure. While such outliers are easy to find in real low-dimensional data, their reliable identification in images is a challenge. This is because they are not necessarily outliers related to the immediate properties of the image (e.g. histogram data), but can also be outliers related to the image context (a tree is not abnormal in an image per se, but it is abnormal when it is lying on railway tracks).
Deep learning methods excel at extracting and generalising such features from data and are therefore ideally suited for anomaly detection. Various approaches facilitate this, including semi-supervised methods that train a neural network on an anomaly-free dataset by reconstructing samples from abstracted features. Anomalous samples are identified as those that cannot be reconstructed correctly.