FEATURE SELECTION AND CLASSIFICATION
Image interpretation consists of three stages:
(a) selection of object signatures (feature selection),
(b) decision-making process (classification procedure),
(c) accuracy estimate.
There are two basic types of object signatures utilized in image
interpretation: spatial and spectral. Temporal changes, sometimes called
temporal signatures, are used for interpretation of multidate images. The
importance of spatial signatures for interpretation of forest patterns on
aerial photographs is well established (Sayn-Wittgenstein, 1970). Correlation
of ground truth with spatial patterns in the ERTS images is more difficult due
to the averaging effect of large pixels on the spatial variation within
classes (Sayn-Wittgenstein and Kalensky, 1974). As a consequence, most of the
ERTS image classifiers are based on spectral signatures only. However,
combination of the spatial and spectral signatures in a single ERTS classifier
yielded a significantly increased classification accuracy (Kirvida, 1973).
signatures only because of the irregular distribution and small size of
coniferous stands in the test area. A complete procedure of digital image
processing adopted for this study is described in the system flowchart (Fig.
1). Classification of single-date imagery was based on the 4-dimensional
feature space defined by pixel values in one ERTS scene consisting of the four
ERTS multispectral images. Classification of the multidate imagery was based
on the 12-dimensional feature space defined by pixel values in the three
multidate ERTS scenes each consisting of the four multispectral images.
The Larose Forest classification algorithm was based on spectral
Each ERTS scene defined a matrix S n of four-dimensional spectral
D
as recorded by the multispectral scanner:
G
1,3 220
S
D
where
S
D
ERTS-1/MSS scene recorded at date D