7.1.5. Image co-registration

Image co-registration is performed when the intention is to study two or more images in a series, typically to understand change. Images may come from the same or from different sensors, and have the same or different spatial resolutions. The rationale of co-registration is to ensure that the images become spatially aligned so that any feature in one image overlaps as well as possible its footprint in any other image in the series. An exact match of footprints is, however, seldom possible (Gomez-Chova et al., 2011). In agricultural applications, certainly in smallholder contexts, accurate image co-registration is a prerequisite to the accurate extraction of temporal profiles of farm fields. Misalignment between images may result in the extraction of incorrect temporal profiles (or crop signatures) and may subsequently give incorrect crop mapping results.

Co-registration is normally carried out by selecting one image as the reference (base/master) to which all the other images are aligned. The selection of the master image can be carried out by use of the image metadata, for instance taking into consideration off-nadir angle and cloud cover percentage, with low values for these two parameters being beneficial. The co-registration process requires identification of common features (e.g. road junctions, stream crossings, etc.) in the master and “warping” of the other images (i.e. those to be co-registered). The locations of the common features are called tie points, and these are used by the co-registration application. In most software applications, tie points are automatically generated on the basis of a few manually selected tie points. Once enough tie points have been generated, a polynomial function is used to align the warped image to the master.

In our thoroughly open-source processing chain, the image co-registration step was implemented using a scale-space detection scheme (Lindeberg, 1998) that automatically identifies tie objects in multiple images. In rural Africa/Asia cases, given that often little or no man-made infrastructure is present in images, these objects typically are tree crowns and wells or other natural (sometimes still man-made) structures that surround the fields under study. This scheme is implemented in R and is integrated in the overall workflow.