Change detection for monitoring of man-made objects using time series of very high resolution spaceborne SAR images

Kurzfassung

Earth observation (EO) satellites with very high resolution (VHR) capabilities can provide regular and repeatable observations of many locations across the globe over long time periods, and are a cost-effective way to monitor remote and inaccessible areas. These satellites can be exploited to monitor different types of human activity, delivering insights which can be used to inform policy- and decision-making. Synthetic aperture radar (SAR) sensors are especially interesting because they can provide VHR images independently of sunlight and under all weather conditions, and are also very well suited for change detection. This thesis presents three novel methods for the monitoring of man-made objects using time series of VHR SAR images, which are then applied to address different practical applications. First, a method for the automatic estimation of all the relevant parameters of oil storage tanks is introduced, which can be applied for the monitoring of oil inventories. The dimensions of each storage tank, as well as the fill level in the case of tanks with a floating roof, are derived from its semicircular double reflections, which are detected using the coherent scatterers (CSs) in the SAR image. When a time series is available, the temporal information is exploited to provide more accurate and robust estimates. This method is application specific, but illustrates the monitoring of well-known, static objects, showing how prior knowledge on an object´s geometry and approximate location can be exploited. Secondly, an unsupervised change detection (CD) method is introduced. Changes are detected by the appearance and disappearance of CSs. These CSs are detected in each image and compared coherently across an image pair or a time series. This enables the detection of changes caused by the appearance, disappearance, or movement of man-made objects, as well as modifications to static objects, while ignoring changes to natural targets such as vegetation. The detected changes can be categorized according to their temporal behavior, and the corresponding changed objects can be segmented. As an example on the monitoring of unknown static objects in complex scenes, this method is applied to monitor construction activity in urban areas. Lastly, a method for the detection and classification of objects in individual SAR images is presented. This task is formulated as a template matching problem, and a fully convolutional Siamese network architecture is used. This approach is based on supervised learning but can be trained with relatively few labelled samples, and also takes into account the specific characteristics of SAR images, such as the significant effect that the imaging geometry has on an object´s appearance. As an example of the monitoring of moving objects, this method is applied to detect and classify different types of airplanes parked at airports. When applied to a time series in combination with the CD method, the arrival and departure dates of the detected airplanes can also be estimated, enabling a detailed monitoring of the level of activity at an airport. The proposed methods have been evaluated using time series of TerraSAR-X images with submeter resolution. The analysis of the obtained results has shown that all three methods perform well, indicating that they could be used in practice for the intended applications. While the method for the monitoring of oil storage is application specific, the methods for change detection and object recognition are more general and could be applied to monitor other types of manmade objects. Besides, as shown in this thesis, these two methods complement each other.