This paper introduces a multi-modal network that learns to retrieve by content vertical aerial images of French urban and rural territories taken about 15 years apart. This means it should be invariant against a big range of changes as the (natural) landscape evolves over time. It leverages the original images and semantically segmented and labeled regions. The core of the method is a Siamese network that learns to extract features from corresponding image pairs across time. These descriptors are discriminative enough, such that a simple kNN classifier on top, suffices as final geo-matching criteria. The method outperformed SOTA “off-the-shelf' image descriptors GEM and ResNet50 on the new aerial images dataset.

The resources published on the Web of data are often described by spatial references such as coordinates. The common data linking approaches are mainly based on the hypothesis that spatially close resources are more likely to represent the same thing. However, this assumption is valid only when the spatial references that are compared have been produced with the same positional accuracy, and when they actually represent the same spatial characteristic of the resources captured in an unambiguous way. Otherwise, spatial distance-based matching algorithms may produce erroneous links. In this article, we first suggest to formalize and acquire the knowledge about the spatial references, namely their positional accuracy, their geometric modeling, their level of detail, and the vagueness of the spatial entities they represent. We then propose an interlinking approach that dynamically adapts the way spatial references are compared, based on this knowledge.