Input, ---+ W, a Tu
oW net,
SE > 0 .=f{(net,A) ----20,
A 3)
| Outputs
zd (activation)
Input, ---> W, |
weights
Figure 1. An Artificial Neuron Computational Structure.
where net-2W, O; (Schalkoff, 1992).
j
The programs use a back-propagation network that learns
using the Generalized Delta Rule:
AW, =n5,09 +CAW,
where n = learning rate, & = momentum, à, = error at the
kth-layer, o; is the output of layer j and W, is the
connection weight between the jth-layer node and the kth-
layer node (Li and Si, 1992).
The term "back-propagation" refers to the training method
by which the connection weights of the network are
adjusted. The training of the network is similar to any
supervised classification procedure, i.e. calibration blocks
have to be selected and used to adapt the classifier. In this
case network weights were adapted. The back-propagation
learning procedure is simple and will not be detailed here.
A multi-layer feed forward neural network using back-
propagation was evaluated in this research. Specifically,
EASI/PACE software NNCREAT, NNTRAIN and NNCLASS
(PCI, 1994) were used.
RESULTS AND DISCUSSION
Per-pixel Classification
Although three-date SAR combination had an 4%
improvement of classification accuracy over the best single
date classification alone, the overall validation accuracies
for both single-date SAR and multi-temporal SAR were very
low (see Table 1). The first reason for the poor accuracies
was that MLC was not an effective classifier for SAR data
classifications due to speckle. The second reason for the
poor performances was that ERS-1 SAR data do not provide
enough differences for eight crop classes due to its high
incidence angle. Satellite SAR systems with multi-incidence
angle, multi-resolution, and multi-wavelength, such as the
Canadian RADARSAT, are very desirable to improve the
performance of satellite SAR data for crop classification.
The third reason for the poor accuracies was that the
calibration and validation blocks were selected based on the
August 5 field data, but the change of the crops over the
50
growing season could cause confusions. For example, a
pasture/cut-hay class in August was an alfalfa class in June.
TM3,4,5 alone produced a 89.895 classification accuracy
(Table 1). Combinations of SAR and TM data improved the
classification accuracies in general. The best overall
accuracy (91.85) was the combination of all three dates SAR
and TM3,4,5. It represents a 2% increase over the TM3,4,5
classification alone.
Per-field Classification
Per-field classification with an ANN proved to be very
effective. The crop classification accuracies improved by
almost 20% using the combination of June, July and August
SAR data. The best single-date (Aug.5) SAR classification
with ANN improved the overall accuracy by about 26%
(Table 2). The accuracies (<60%), however, were not high
enough for operational crop inventory and analysis.
The best per-field classification of 96.8% with an ANN
classifier was achieved using the combination of TM3,4,5
and Aug. 5 SAR data (Table 2). 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. Another possible explanation is that the
SAR data and TM data were acquired one day apart and the
ground condition was the same. In this classification, all
crops achieved 100% accuracies except alfalfa and
pasture/cut-hay-alfalfa classes (Table 3). Alfalfa had 16.3%
commission error to first corn class (i.e., good growth), and
pasture/cut-hay-alfalfa had a 15.1% commission error with
second class of soybeans (i.e., poor growth).
The second best classification accuracy was achieved using
the combination of TM3,4,5 and July 24 SAR data (Table 2).
The classification accuracy of TM3,4,5 using an ANN,
however, is lower than that of a MLC. This is possibly due
to the second corn class (ie., poor growth) was not well
trained. It resulted in poor accuracy of second corn class
(only 25.9%) with a commission error to barley/oat class
62.1%, while all other classes were 100% correctly
classified.
The best ANN classification improved 5% in accuracy than a
MD classification using the same dataset. ANN produced
better accuracies in general than those derived from a MD
classifier (Tables 3&4). This is because the post-
segmentation classifier based on the MD classification of
field means used calibration data obtained as in a per-pixel
classification. Such a procedure fails to exploit the full
range of information that segmentation offers.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996
e E es. all
N
S(