MULTI-TEMPORAL ERS-1 SAR AND LANDSAT TM DATA FOR AGRICULTURAL CROP
CLASSIFICATION: AN ARTIFICIAL NEURAL NETWORK APPROACH
Yifang Ban, Ph.D. Candidate
Earth-Observations Laboratory, Institute for Space and Terrestrial Science
Department of Geography, University of Waterloo
Waterloo, Ontario, Canada N2L 3G1
Tel: (519) 888-4567 x5267
Fax: (519) 888-6768
yban@cousteau.uwaterloo.ca
Commission VII, Working Group 2
KEYWORDS: Multitemporal, SAR, Landsat, Agriculture, Classification, Artificial Neural Networks
ABSTRACT:
Multi-temporal ERS-1 synthetic aperture radar (SAR) and Landsat TM data were used to evaluate an artificial neural network
approach. for crop classification. Six major crops, i.e., winter wheat, corn (good growth & poor growth), soybeans (good
growth & poor growth), barley/oats, alfalfa, and pasture/cut-hay-alfalfa, were classified into eight classes. The results show
that both a single-date and multi-temporal SAR data yielded poor classification accuracies using a maximum likelihood
classifier (MLC). With per-field approach using a feed forward artificial neural network (ANN), the overall classification
accuracy of three-date SAR data improved almost 2096, and the best classification of a single-date (Aug. 5) SAR data improved
the overall accuracy by about 26%. These accuracies (<60%), however, were not high enough for operational crop inventory
and analysis. Using the combination of TM3,4,5 and Aug. 5 SAR data, the best per-field ANN classification of 96.8% was
achieved. It represents a 8.5% improvement over a single TM3,4,5 classification alone. It also represents a 5% increase over
the best per-pixel classification. This indicates that a combination of mid-season SAR and VIR data was best suited for crop
classification. The results also show that the best ANN classification had a 5% higher accuracy than a minimum distance (MD)
classification using the same dataset.
INTRODUCTION
Radar remote sensing has the potential to play an important
role in agricultural crop monitoring due to its independence
from solar illumination and cloud cover. With the launch of
the European Remote Sensing Satellite (ERS-1), the first
long-duration spaceborne imaging SAR system became
available. This and other spaceborne SAR systems, such as
JERS-1, ERS-2 and the Canadian RADARSAT, provide
researchers with an excellent opportunity for developing
multi-temporal SAR agricultural applications.
The synergistic effect of integrating SAR data and imagery
acquired in the visible and infrared (VIR) portions of the
spectrum has also been recognized as important for two main
reasons. First, timeliness of SAR fills information gaps
during overcast or hazy periods at the critical stages of the
growing season, and second, the combination of data from
different parts of the spectrum often leads to increased
classification accuracy. Previous studies have shown that
combining airborne SAR and satellite VIR data improves
crop classification accuracies (Brown et al., 1984; Guindon
et al.,, 1984; Hirose et al, 1984; Brisco et al, 1989;
Fiumara and Pierdicca, 1989; Dixon and Mack, 1990; Brisco
and Brown, 1995). Very little research, however, has been
done to improve crop classification accuracies using data _
from two satellite sensors (Kohl et al., 1993; Fog et al.,
1993). Thus, the potential of Satellite SAR and VIR
synergism still needs further investigation.
Conventional statistical classifiers, such as MLC, make a
number of untenable assumptions about the dataset to be
classified (Foody et al., 1995). For example, this parametric
approach requires the data to have a Gaussian distribution.
SAR data, however, are not normally distributed due to
speckles. Therefore, the accuracies of SAR crop
classification using conventional statistical classifiers are
often not high enough for crop inventory and analysis. In
order to improve classification accuracy, it is necessary to
explore robust classifiers using non-parametric and non-
statistical approaches.
ANN classifier presents a distribution-free approach to
image classification. It also has the special advantages of
simple local computations and parallel processing
(Schalkoff, 1992). In the past few years, studies have shown
that neural networks compared well to statistical
classification methods in classification of multi-date, multi-
source remote sensing/geographic data, very high
dimensional data and classification with high number of
classes (e.g., Benediktsson et al., 1990a; Benediktsson et
al., 1990b; Kanellopoulos et al, 1991). Foody et al. have
(1994; 1995) also found that ANN produced higher
classification accuracies in general than those derived from
statistical classifiers when they were applied to airborne
SAR data for classifications of agricultural crops. Therefore,
it is desirable to investigate the effectiveness of ANN for
crop classifications using satellite SAR and VIR data.
The objective of this study was to evaluate the synergy of
multi-temporal ERS-1 SAR and Landsat TM data for crop
classification using an artificial neural network approach.
The specific objectives were:
* to evaluate the crop classification accuracies using a
single-date SAR data alone and multi-temporal SAR data,
* toevaluate the synergism of multi-temporal ERS-1 SAR
and Landsat TM data for improving crop classification,
and
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996