Extracting Features from Remotely Sensed Images
Introduction
The aim of this project is to develop image analysis techniques to help
with the problem of identifying and digitising certain types of features
from remotely sensed images. The research is conducted by Dr Paul Lewis, Dr Mark
Dobie and Dr Mark Nixon as an EPSRC funded project.
The industrial collaborators for the project were Scott Wilson
Kirkpatrick Ltd, a firm of consulting engineers. They use satellite and
aerial imagery to survey sites for engineering projects. This is useful
for surveying large areas and for remote areas where reliable, up to
date maps may not be available.
Their primary interest was digitising thin, curvilinear features such as
rivers, roads and railways. Currently these features are manually
digitised which is time consuming and labour intensive. Other features
of interest included wide features, such as estuaries and lakes, and
features with well defined shapes, such as oil and gas tanks.
Developments
Two main methods were developed to help with these problems. D'esopo's algorithm allows the rapid, interactive
digitisation of single, narrow features. A minimum
spanning tree method allows a whole network of similar features to
be extracted in one go. For further details, see some publications associated with this work. This pair
of figures shows the results of applying the
minimum spanning tree method.
The boundaries of wide features can be digitised if the image is
preprocessed with an edge detector. This makes the boundaries show up as
narrow lines, which can be digitised using the two methods developed.
These methods, together with a range of image preprocessing tools, are
implemented in the extract image processing
system. This package provides a flexible environment for experimenting
with image processing algorithms.
Future Work
These two methods allow narrow features and boundaries of wide features
to be digitised semi-automatically. More sophisticated methods could be
applied to wide features to provide area and centreline or centroid
information. For identifying similarly shaped features, a generalised Hough transform could be used, given
one example of the feature.