Environmental Engineering Reference
In-Depth Information
k-Nearest Neighbour (KNN). The results refer to
average percentage accuracy in a 10-fold cross-
validation experiment where the algorithms are
repeated trained on a sample of
10
th of the data and
then tested on the remaining
10
th.
where the ground truth for the training set corre-
sponded to the class labels allocated by the rota-
tional algorithm proposed by Rico-Ramirez (2004)
and Rico-Ramirez and Cluckie (2007). In addition
to providing an easily understandable set of robust
classification rules, LID3 also has the advantage
that it can classify an image in real time, pixel by
pixel, in contrast to Rico-Ramirez's algorithm,
which requires preprocessing of each image in its
entirety before any pixel level classifications can
be made. The features used by LID3 were as fol-
lows: reflectivity factor (Z
h
), the differential refec-
tivity (Z
dr
), the linear depolarization ratio (L
dr
) and
the height measurement (H
0
).
The data for the experiments were generated
from 1354 images resulting in 191,235 labelled
data vectors. Examples of rules drawn from the
resulting linguistic decision tree are as follows:
.
IF L
dr
is between low and med., AND H is be-
tween med. and high, AND Z
h
is only high, AND
Z
dr
is betweenmed. and highTHENthe pixel class
is Rain: 0.998, Snow: 0.002, Bright band:0
.
IF L
dr
is only med., AND H is between med and
high, ANDZ
h
is betweenmed. and high, ANDZ
dr
is only low THEN the pixel class is Rain: 0.03,
Snow: 0.97, Bright band:0
.
IF L
dr
is only high, ANDH is onlymed., ANDZ
h
is only high, ANDZ
dr
is only high THEN the pixel
class is Rain: 0.02, Snow:0,Bright band: 0.98
Table 8.1 shows a comparison of the results of
LID3 with a number of other machine learning
algorithms including Naive Bayes, Neural Net-
works, Support Vector Machines (SVM) and
Time Series Modelling
In this section we discuss the application of AI
techniques to time series forecasting problems
relating to two different river catchments.
Bird Creek catchment
The Bird Creek river basin is in Oklahoma (USA)
close to the northern state border with Kansas.
The catchment area covers 2344 km
2
with the
outlet of the basin near Sperry about 10 km north
of Tulsa. The area itself is 175 to 390m above the
mean sea level and has no mountainous regions or
large water surfaces to influence local climate
condition. Some 20% of the catchment surface is
covered by forest and the main vegetative cover is
grassland, with the soil storage capacity being
described as very high (see Georgakakos and
Smith 1990). Figure 8.7 shows the river basin
describing the Bird Creek area - see Georgakakos
et al. (1988) for more information.
The database considered here was collected to
form part of a real-time hydrological model inter-
comparison exercise conducted in Vancouver,
Canada, in 1987 and reported by the World Mete-
orological Organization (1992). The database con-
tains information on average rainfall derived from
12 raingauges situated in or near the catchment
area (U
t
average rainfall at time t) and on stream-
flow measured using a continuous state recorder
(Y
t
flow at time t). Randon et al. (2004) applied the
Fuzzy Bayesian learning algorithm based on a
Semi-Naive Bayes assumption. In this study the
data were split into a training set consisting of
2090 examples from November 1972 to April
1974, and a test set of 1030 examples from
November 1974 to December 1974. The objective
of the experiment was to predict flow 36 hours in
advance, corresponding to Y
t
รพ
6
. It was assumed
Table 8.1
Comparison of resultswith LID3 andMachine
Learning Algorithms
Accuracy
Bright
Band
Algorithm
Rain
Snow
Average
Naive Bayes
75.3%
68.5%
98.6%
80.8%
Neural Network
85.2%
75.8%
99.0%
86.7%
SVM
85.1%
84.0%
98.4%
89.2%
KNN
85.1%
84.0%
98.4%
89.2%
LID3
93.5%
89.0%
99.4%
94.0%
KNN, k-Nearest Neighbour; SVM, Support Vector Machines;
