Every day new images covering earth regions are available on Internet - i.e. as it
happens on Google Earth (TM) -. Public institutions that deals with land and
environment monitoring are strongly interested in automatic evaluation of land
changes both for planning reasons and administrative tasks (i.e. illegal building
discovering).

This kind of activity could be easily performed with instruments able to compare
aerial representation of land picked in different times. In order to perform this
evaluations our research group has been studying and developing procedures for
automatic comparison of aerial photos.

The procedure illustrated in this paper, operating on GRASS GIS is a first tentative
to register a new image comparing it with an existing one, searching subsequently
homologue points (in order to georeferencing automatically the second image) and
hence looking for corresponding building. The procedure is based on algorithms of
image segmentation and image regions comparison on samples found by region growing
calculations. In the general context, methodologies for correspondences searching are
classified on the base of the entity involved: area based matching (ABM) by image
gray levels, feature based matching (FBM) by edges and regions, relational by
symbolic descriptions. In the present work, we propose a FBM approach that could be
also a support for high level description of the objects detected on the images.

In particular, the implemented methodology is composed by the following computational
phases:

   * RGB colour thesholding in order to retain interesting objects;
   * segmentation by region growing algorithm to find connected regions
     ( it is used an automatic procedure to place seeds on image);
   * description of each detected region (centroid coordinates, shape
     factors by using Fourier descriptors, and area);

These computational steps are performed on various images related approximately to
same land area taken at different time, position, or resolution. Given two images to
process, in order to find corresponding areas, simple patterns of displacement of
relevant regions are searched: triangles having as vertexes three near regions are
described by normalized distances (considering areas) and angles. A matrix counts all
the correspondences found between the triangles from each image in order to use them
to compute the transformation. In this way, the two images could be registered with
an acceptable error, and after a refinement based on the whole common area on two
images, they could be georeferenced. Moreover, previous described methodology
individuates regions relevant for successive evaluation of land changes or anomalies
by simple difference, and it is suitable for symbolic description and analysis. In
fact, different preliminary colour segmentation allows to insulate various land
objects (i.e. roofs, streets, and so on), distinguishable by texture and hue.

Preliminary experimentations show good results also when the system deals with all
the possible planar affine transformation (rotation, translation and scale), using
both aerial and Internet images with different resolution.

The papers also deal with the construction of a library that is based both on
functions already present in GRASS standard required libraries as on other open
source libraries.

The presented research keep going in order to improve georeferencing results: new
functions are going to be added with the aim to reduce computational times searching
correspondence not only for objects with the largest area but also for objects with
distinctive shape factors.