Applications of multi-image remote sensing

This thesis studies the problem of 3D reconstruction from a collection of high-resolution satellite images. Satellite multi-view 3D reconstruction requires a very fine control of the acquisition geometry, in order to guarantee the consistency of altitude estimates obtained from different views. The first part of the thesis is therefore devoted to the optimization of the mathematical representation of the acquisition geometry, which usually takes the form of RPC camera models. We propose a bundle adjustment methodology that maximizes the geometric consistency between a set of satellite views and the associated RPC cameras. This methodology incorporates an RPC estimation algorithm that allows the direct composition of the original unrefined models with corrective transformations, without using approximate intermediate representations. The second part of the thesis presents different practical applications of multi-image remote sensing, most of which benefit from the consistency control of the acquisition geometry. The different methods concern the following topics: the detection of volume changes on the Earth's surface across different dates; the geometrically consistent generation of large-scale mosaics built from smaller satellite images; a neural rendering network (NeRF) capable of learning the geometry of a satellite scene in a self-supervised manner and also of synthesizing new realistic views, with the ability to distinguish shadows and transient objects from permanent structures; and a comparison between classic algorithms and supervised deep learning networks for dense stereo matching. As a result, this thesis describes a variety of cutting-edge ideas on the exploitation of optical satellite images that have the potential to improve activities related to large-scale land surface knowledge, such as surveillance, urban planning or natural resource management. The presented methods are evaluated with high-resolution images from the WorldView-3 and SkySat constellations. The implementation of most methods is also released as open-source Python code.