Theoretical Background and State of the Art - Historical Land Use/Land Cover Classification Using Remote Sensing - page 28

The temporal information dimension in used remotely sensed data is the most
useful factor in natural vegetation and agricultural applications for identifying crop
types (Wardlow et al. 2007 ). This is because agricultural features have great
(within-class and within-season) spectral flexibility, that is based on several
complex natural and biophysical factors (e.g., crop type/s, soil, water and geo-
graphical location). The observation and understanding of these various spectral
responses of crops, and comparison with the physical characteristics of remotely
sensed data recorded in various dates in the year (building a crop-specific temporal
record), would give us the appropriate date(s) during the growing stages in which
the crops of interest are spectrally separable. Also, by observing the physical
derived spectral indices from remotely sensed data that are sensitive to natural
vegetation cover over time, it is possible to discriminate crops (Ozdogan 2010 ).
Discrimination of crops using remote sensing imagery is generally achieved
with supervised or unsupervised classification algorithms (Jensen 2007 ). Recently,
nonparametric algorithms, expert knowledge and ancillary data have been used in
the process of cropland classification, improving the overall classification accu-
racy. One example of this is the establishment of neural networks for crop type
identification, which is the most important development in information extraction
from remotely sensed data in the last 15 years (Del Frate et al. 2003 ). Multi-sensor
data fusion and classification of time series data are being applied in cropland
classification more and more (Chen et al. 2008 ). The most simple method of
distinguishing crops is the classification of images into large-scale classification
categories including all agricultural features (Level 1 in LULC-classification)
(Campbell 2002 ). From this level of classification, agricultural features can be
classified into cropping and non-cropping regions.
The interaction between crop field scale and pixel size is a significant factor,
especially in heterogeneous cropping areas. For instance, large pixel dimensions
allow an increasing chance of recording mixed reflectance values. This resulted
mixed spectral response is confused by traditional local agricultural management
practices, such as found in most areas of the Euphrates River Basin, where crops
are sometimes planted in almost 30 m strips (see Fig. 5.29 ). This is alternated with
un-cropped areas (bare soil, stubble, dirt roads, etc.) of similar size to the cropped
strips. So, pixels that are not entirely homogeneous (e.g., solely forest, vegetation,
wheat crop, etc.), have mean reflectance values (composite spectral response that
might match neither feature's spectral response) as a result of more than one
feature within the pixel area. Such pixels are known as mixels and are an ever-
present problem in cropland classification, reducing their discriminating power
(Chen et al. 2008 ). Spectral Mixture Analysis techniques (SMA) have been
developed and used to solve the mixel-problem in remotely sensed data (Fitzgerald
et al. 2005 ). Confusion between natural vegetation and cropland is also another
major source of error in crop classification using low spatial and/or spectral res-
olution remotely sensed data. Sometimes this is also true of high-resolution
imagery. This type of confusion is especially common in areas with very com-
plicated traditional local agricultural management practices, which are controlled
by natural

topography or from land ownership (Loveland et al. 1999 ). An