An Introduction to Remotely Sensed Data Analysis - Sampling Spatial Units for Agricultural Surveys - page 83

4.6.2 Supervised Classification
Supervised classification is an essential tool for extracting quantitative information
from remotely sensed image data. Supervised classification is much more accurate
for mapping classes, but heavily depends on the cognition and skills of the image
specialist. Supervised classification is a very complex topic, and involves different
steps.
Some pre-processing is required before pixels can be classified into known
classes. One of the most important pre-processing operations is to find similar
regions in the image (from a spectral point of view), and collect the main spectral
characteristics of the classes we are interested in. These homogenous regions are
then used to determine the training sites in the following classification. This topic is
linked to the choice of statistical unit (see Chap. 5 ) .
A variety of algorithms are available for supervised classification, ranging from
those based upon probability distribution models for the classes of interest, to those
in which the multispectral space is partitioned into class-specific regions using
optimally located surfaces.
In supervised classification, the set of labels to be assigned to the regions is a
priori known. In agricultural surveys, the classical situation is that a sample of
spatial units is selected from a frame covering the domain. These units can be
points, regular or irregular polygons (Sects. 1.4 and 3.2 ), or geocoded data mapped
through an agricultural household survey. These sources of information constitute
the basic set of labels that are defined in the questionnaire (see Sect. 9.2 ), and their
geographical positions are available from randomly selected portions of land.
The general procedure for using these data is as follows:
• Step 1. Define the list of labels for classifying the RS image, by aggregating or
eliminating codes defined in the ground survey. The main reason for modifying
the list of labels is that not all the land cover codes can be discriminated using
remotely sensed data. Besides, some can only be discriminated (particularly in
some period of the year) if they are aggregated (for example, barley, soft wheat,
and durum wheat).
• Step 2. Choose some pixels that are representative of each class of interest, to
form the so-called training areas. From a technical point of view, this operation
is usually performed by overlapping the raster satellite or airborne image with a
vector map observed using a ground survey. This phase, denoted as feature
extraction, can be either automatic or manual, depending on the quality of the
ground data and the geometric accuracy with the overlapping image.
• Step 3. Use the training areas to estimate the parameters of the chosen classifi-
cation algorithm.
• Step 4. The labels are assigned to pixels, and the image is classified.
The term training area is used because the spectral characteristics of these areas
are a priori known and are used to train the classification algorithms to assign labels