Action I Development of 3D calculation methodologies applied to archival aerial images.
Denis Feurer (denis.feurer@ird.fr)
Old aerial images, with a resolution finer than one meter and dating back to the beginning of the last century, have the potential to document past events. Furthermore, the recent computer vision algorithms used in photogrammetry (“SfM” approaches, Structure from Motion) open up, because they are “all-image” approaches, new possibilities for using these images. Indeed, SIFT (Scale Invariant Feature Transform) type algorithms make it possible to automatically detect points in images by associating them with a descriptor invariant to linear transformations (change of scale and rotation). The objective of the action thus aims to assess the potential of such data and algorithms for the documentation of past 3D changes.
The idea of this work starts from the assumption that, for certain spatio-temporal scales, these descriptors can also present an invariance in time. The proposed method, called Time-SIFT, simply consists of carrying out the first steps of the calculation with a set of images from different dates in order to constitute a single geometric reference. Once all the data is in a single geometry, it becomes possible to calculate 3D differences between dates. A multi-factor analysis was carried out to assess the interest of the Time-SIFT method compared to other treatment strategies.
Studies carried out on images covering an area of 170km² over a period of more than 40 years have demonstrated that the use of the Time-SIFT method was the first factor of success for the detection of 3D changes. The application of this method has made it possible to highlight 3D changes linked to urban developments, road works but also samples and embankments.