Você está na página 1de 9

Application of genetic programming and Landsat multi-date imagery

for urban growth monitoring


Djerriri Khelifa
1 a,b
, Malki Mimoun
b

a
Division of Earth Observation, Centre for Spatial Techniques, Arzew 31200, Oran, Algeria
b
EEDIS Laboratory, University of Sidi Bel Abbes 22000, Algeria
ABSTRACT
Monitoring of earth surface changes from space by using multi-date satellite imagery was always a main concern to
researchers in the field of remotely sensed image processing. Thus, several techniques have been proposed to saving
technicians from interpreting and digitizing hundreds of areas by hand.
The exploiting of simple, easy to memorize and often comprehensible mathematical models such band-ratios and indices
are one of the widely used techniques in remote sensing for the extraction of particular land-cover/land-use like urban
and vegetation areas. The results of these models generally only need the definition of adequate threshold or using
simple unsupervised classification algorithms to discriminate between the class of interest and the background.
In our work a genetic programming based approach has been adopted to evolve simple mathematical expression to
extract urban areas from image series. The model is built from a single image by using a basic set of operators between
spectral bands and maximizing a fitness function, which is based on the using of the M-statistic criterion.
The model was constructed from the Landsat 5 TM image acquired in 2006 by using training samples extracted with the
help of a Quick-bird high spatial resolution satellite image acquired the same day as the Landsat image over the city of
Oran, Algeria. The model has been tested to extract urban areas from multi-date series of Landsat TM imagery

1. INTRODUCTION
The analysis of multi-date remotely sensed images is known as a powerful tool to a monitor of the spatial and temporal
processes of urban growth and land use change. This is likely due to the ability of earth observation satellites to provide
spatially consistent data sets that cover large areas with both high spatial detail and high temporal frequency.
Classification based approaches for change monitoring starts usually with image preprocessing, which is a step
introduced in an image analysis schema to minimize the influence of unwanted points and to maximize the influence of
desired ones. It composed generally of operations such as filtering, enhancing, transformation, combination, principal
component analysis, and spectral band indices or combinations of these processes applied to the original data
1
. After the
transformation, some features and spectral properties become more discernible when compared with the raw data.
The using of simple, easy to memorize and often comprehensible mathematical models such band-ratios and indices are
one of the widely used techniques in remote sensing for the extraction of particular land-cover/land-use like urban and
vegetation areas. Such indices can be used by personnel with no or limited image processing expertise during operational
situations. The results of these models generally only need the definition of adequate threshold to discriminate between
the class of interest and the background.
In our work a genetic programming based approach has been adopted to evolve simple mathematical expressions to
extract urban areas. The model is built from a single Landsat TM image by using a basic set of operators between multi-
spectral bands (addition, subtraction, multiplication and division) and maximizing a fitness function based on the using

1
kdjerriri@cts.asal.dz; phone +213 772615113; fax +213 772615113
*
*
of the M-statistic separability measure, which assesses the degree of discrimination between the urban and non-urban
pixel groups. A high value of M criterion denotes that the classes are well separated and that the two regions are
relatively easy to discriminate.
The model was constructed from the Landsat 5 TM image acquired in 2006 by using training samples collected with the
help of a Quick-bird high spatial resolution satellite image acquired the same day as the Landsat TM image over the city
of Oran, Algeria. The model has been tested to extract urban areas from multi-date series of Landsat TM imagery
acquired respectively in the years: 1984, 2003 and 2011. The generalization capacity of the obtained model has been
analyzed, evaluated and compared to results from GP-based models built with different fitness functions and to results
from other supervised classification techniques.
The article describes firs the study area and the used data, then presents a brief overview of GP and its main terminology
followed by a description of the proposed framework. Results from the application of the proposed methods on multi-
temporal datasets were showed. Finally the conclusions obtained from these experiments are discussed.

2. STUDY AREA AND USED DATA
In our study the Landsat TM imagery has been chosen as data source to monitor urban growth from space. The Landsat
program offers the longest continuous scientifically valuable free datasets, which were widely used for studying land
cover change since the mid-eighties. The Program will continue to obtain valuable data and imagery through the recently
launched Landsat Data Continuity Mission (LDCM), which is the new satellite in the Landsat series.
The original imagery used in this project was a Landsat 5 scene of Oran city, Algeria taken in august 2006. A set of
training regions were collected by visual inspection of the image and with help of high spatial QuickBird image acquired
the same day as the 2006 Landsat image. The training set represents urban areas and other types of land-cover.
Three (03) other Landsat multispectral scenes were considered. The years selected were: 1984, 2003 and 2011 with
acquisition day near to the one of 2006.
Preprocessing methods such as atmospheric correction, and normalization of the images were not performed since the
main objective is to develop methods to be used by personnel with no or limited image processing expertise during
operational situations.

3. GENETIC PROGRAMMING (GP) FOR SYMBOLIC CLASSIFICATION
Evolutionary algorithms (EA) belong to Biological-Inspired Algorithms (BIAs), which have been successfully applied
recently in the fields of remote sensing
2,3
. The evolving, dynamic and meta-heuristics nature of EA makes them efficient
to find approximate solutions to optimization problems. Four (04) main branches of EA are algorithms, genetic
programming, evolutionary strategies, and evolutionary programming.
Most critics of using bio-inspired algorithms in general and EAs in particular, concern their expensive time cost.
However thanks to nowadays high performance computing machines, processing time can be reduced considerably, in
addition GP can easily parallelized through evaluating the population of solutions in parallel. Such parallelization can be
even done on personal computers through low-cost GPU-based parallel computing.
The main principle of the EA is that individual characteristics are transmitted from parents to children over generations,
and those with high fitness will survive and pass characteristics to their offspring. EA evolves iteratively a population of
individuals over generations. The individuals of the initial population are usually randomly generated or based on a prior
knowledge. In each generation, the fitness of each individual in the population is evaluated according to their capacity to
minimize or maximize the fitness function, depending on the nature of the problem. According to the fitness function
that indicates numerically how good an individual is as a solution to the problem; multiple individuals are stochastically
selected from the current population based on their fitness, and recombined by the crossover operator and possibly
randomly modified by mutation operators to form a new population. The new population is then used in the next
generation of the algorithm. Commonly, the algorithm terminates when either a maximum number of generations has
been produced, a satisfactory fitness level has been reached for the population, or when the best fitness did not change
continually throughout a fixed number of generations.
The goal of GP, as its name implies, is the evolution of computer programs or mathematical expressions instead of
sequence of values as for GA. GP individuals are usually seen as trees, where leaves correspond to terminal symbols
(variables and constants) and internal nodes correspond to non-terminals (operators and functions). The set of all the
non-terminal symbols allowed is called the function set, whereas the terminal symbols allowed constitute the terminal
set
4
. An example of tree that represents the expression for calculating normalized difference vegetation index is
illustrated in the figure 1.


Figure 1. Normalized difference vegetation index as a genetic programming tree solution

Five (05) major steps prior to the implementation of genetic programming application to solve a specific problem
5
:
1. Definition of the set of terminals: terminals can be interpreted as the set of arguments or inputs used by
the functions in the programs constituting the individuals of the population. The items in the terminal
set can be constants and variables.
2. Definition of the function set: a set of mathematical functions which will be used as operators in the
programs.
3. Definition of the fitness function: the fitness function is the driving force for the evolutionary process.
The fitness is a quantitative measure of the goodness of a candidate solution.
4. Definition of the parameters controlling the run such as: number of population, number of generations,
percentage of reproduction, percentage of crossover, percentage of mutation, maximum depth of tree,
and others.
5. Definition of the termination criteria: due to the iterative nature of this approach, it is required that a
termination criteria to be defined. The establishments of the maximum number of generations or the
stopping condition when a specific fitness is achieved are the most commons ones.


4. EVOLUTIONARY FRAMEWORK
The proposed evolutionary framework (EF) consists of a system integrating genetic programming, separability measures,
and standard image classification algorithms. The main objective of the evolutionary framework is to develop, is a learn-
from a training set, feature specific spectral indices in the form of mathematical functions to map the original
multispectral image into a transformed image to facilitate subsequent operations to achieve the final goal of feature
extraction from multi-date images. Figure 2 shows the color composite of the original image and its associated
transformed image.

Figure 2. Original image and its associated transformed image.










Figure 3. The key components of the evolutionary framework
Figure 3 is a schematic describing the key components of the evolutionary framework. The input to the system consists
of the multispectral image to be classified, reference data in the form of training data, and the parameters controlling the
run. The reference data consists of examples where the feature is found and other examples the feature is not found. The
parameters controlling the run include the number of iterations, solutions population size, percentage of crossover,
percentage of mutation, and stopping criteria.
The system begins by randomly generating a set of candidate solutions for the problem. Each candidate solution is
applied to the multi-column original training dataset generating a new set of transformed single column dataset. The
transformed dataset is individually passed to the evaluation process, which uses the M-statistic to assess the separability
between the class of interest and other classes (presence or absence of the target feature). The set of candidate solutions
are then sorted by fitness values and the stopping criteria are evaluated. The stopping criteria are user-defined; however,
Image Data
Sample Data
Parameters
Input
Sample Data Genetic Programming
Fitness Function Function set
Stopping Criteria
Genetic Programming
Spectral Indice
Transformed
Image
Output
Classified Image
common choices are a maximum number of iterations and/or a fitness threshold. If one of the stopping criteria is
reached, the system stops and it outputs the mathematical expression of the candidate solution with the highest fitness. If
none of the stopping criteria are met, the system continues in an iterative fashion following the principles controlling the
biological evolution theory. A new set of candidate solutions is formed from the fittest individuals (candidate solutions
with the highest fitness values). Some candidate solutions are copied with no change (replication), while others are
carried forward to the next set of candidate solutions after genetic operations, such as crossover, mutation, and restarting
are performed. Crossover and mutation are standard operations in genetic algorithms and genetic programming. The
newly evolved set of candidate solutions are then applied to the original image and the entire process is repeated until
one of the stopping criteria is reached.
Fitness function
As fitness function to evaluate each candidate solution we used the M-statistic separability criteria, which measures the
separability between two classes c1 and c2, which are the class of objects of interest and the background (the remaining
objects). It can be calculated by normalizing the difference between the means of two classes c1 c2 by the sum of
their standard deviations c1 + c2. According to
6
a value of M<1 denotes that the histograms significantly overlap and
the ability to discriminate the two classes is poor (figure 4). A value of M>1 denotes that the histogram means are well
separated and that the two regions are relatively easy to discriminate.

Figure 4. Illustration of the M-statistc group separability measure
Another fitness function used in the experiments to evaluate candidate solutions, which is based on the using of K-means
unsupervised classification and Kappa statistic as accuracy measure
1
.
5. PRELIMINARY RESULTS AND DISCUSSION
Two types of experiments have been done. In the first one we used the proposed M-statistic based fitness function, while
in the second experiment the kappa coefficient accuracy measure has been utilized as fitness function.
Some results for the task of spectrally separating urban areas from the remaining image background are displayed in
table 1. The last column represents the best obtained mathematical expressions; the second column and the third columns
represent the tree depth and tree size respectively, which express the complexity of the discovered models. The fourth
column contains the fitness values (M-statistic value for M-statistic based fitness function and Kappa coefficient for K-
means based function).
These results were developed using samples from the 2006 Landsat image and then these mathematical formulas were
applied to other images: 1984, 2003, and 2011.





Table 1. Some results for the task of spectrally separating urban areas from the remaining image background
DEPTH SIZE FITNESS MODEL
M-STAT.
4 9 1.4194
b
7
(b
1
b
7
)
b
4
+b
1

5 19 1.97194
b
7
2
(b
1
b
7
)(b
1
b
5
+b
2
)
b
4
(b
1
b
2
)

5 23 2.47029
b
7
2
(b
3
+b
1
b
5
)(b
2
+b
1
b
5
)
b
1
4

KMEANS
9 23 0.945
LOG(ABS(B1-(((B2-B1)*((LOG(LOG(B5))/123.816)-((B1-200.829)/(B5-
200.829))))-LOG(B5))))
8 39 0.952 2*B2*B1+B2-B7*B5+B7*B1-B4-2*B5+B1+B7-B4*B2
5 16 0.894 ((B1-SQRT(B5))-SQRT(B5+B3))-SQRT(B1-B5-B3)

The results expressed as M-statistic values, which operate by evaluating the separation between the histograms that are
produced by plotting the frequency of all the pixel values within the two classes (urban and non-urban areas). This
criterion definition suggests that M statistic greater than 1 (M >1) denotes that the histogram means are well separated
and that the two regions are relatively easy to discriminate. Based on this categorization, the majority of results obtained
using the evolutionary frameworks are good.
The same thing can be said about the kappa coefficient results of GP based on K-means fitness function. The high Kappa
values obtained ( > 0.9) indicate good classification of urban areas versus other land cover types. The Kappa coefficient
was used in our study instead of the overall accuracy because it allows correction of argument due to chance.
The developed preprocessing functions were then applied to the full size original 2006 Landsat scene of Oran, Algeria
producing new single band images. The spectral values of this image vary considerably to the raw image, as shown in
the in Figure 5. After applying the transformation to the original image, the class of interest (urban areas in blue color)
becomes more discernible when compared with the raw data.
As the GP was evolved using only samples and not the whole original spectral bands, this may lead to differences in
results when apply the discovered models to full size images (k-means classification results). Thus, we have selected
only those models that provide good generalization to full data.

Figure 5. Separability between urban area and other class before and after the transformation
Band1 Band2 Band3 Band4
Band5
Band7
Transformed data
The generalization capability of the obtained models from 2006 image to the remaining images has been studied using
overall accuracy and kappa coefficient statistics. The results are shown in the tables below.
Table 2. Results of generalization of K-means based fitness models.
Date 1984 2003 2006 2011
Model 1: LOG(ABS(B1-(((B2-B1)*((LOG(LOG(B5))/123.816)-((B1-200.829)/(B5-200.829))))-LOG(B5))))
OA 85.4% 45.4% 97.4% 85.5%
KC 0.70 0.770 0.945 0.715
Model 2 : (13.982-SQRT(B5))*((B3*195.509)-ABS(B7))
OA 86.2% 61.2% 96.8% 88.4%
KC 0.718 0.295 0.934 0.767
Model 3 : 2*B2*B1+B2-B7*B5+B7*B1-B4-2*B5+B1+B7-B4*B2
OA 87.5% 94.1% 97.7% 92.2%
KC 0.738 0.879 0.952 0.841

Table 3. Results of generalization of M-stat. based fitness models.
Date 1984 2003 2006 2011
Model 1: ((ABS((B5-159.112))*B2)*LOG(B1/44.1427))*(((118.094/(B4+B1))*(139.184-B4))*((139.184-B4)*ABS(B3))), M = 3.26669
OA 87.4% 50.3% 99.5% 65.6%
KC 0.737 0.039 0.989 0.359
Model 2: ABS(SQRT(ABS(B3))/(B5+111.284)), M = 1.92654
OA 85.8% 75.1% 92.8% 0.90%
KC 0.713 0.528 0.852 0.814
Model 3 :
b

2
(b
3
+b
1
-b
S
)(b
2
+b
1
-b
S
)
b
1
4

OA 89.0% 97.1% 98.5% 89.5%
KC 0.772 0.9416 0.968 0.791

Results from table 2 and table 3 show that the models developed perform well for the 2006 image, where overall (OA)
accuracies and Kappa statistics (KC) are high (OA > 90% and KC > 0.8). The generalization capability to other images
differs from model to another, some of them present good generalization and others overfit the data for which is trained,
for example the model1 for M-statistic based fitness function, where kappa = 0.98 for 2006 data and 0.039 for 2003 data.
Figure 6 shows the transformed image by using the best obtained expression and its k-means classification.

Figure 6. Transformed image by using the first obtained expression and its k-means classification


Figure 7. Results of extraction of urban area from 1984, 2003 and 2011 images.
Finally, the areas of urban fabric have been calculated for each image by using the support vector machine classifier
(SVM) supervised classification results as reference to compared with the obtained unsupervised classification results.
Results for the application of best obtained models to other multi-date images are displayed in table 4. These results
represent the estimated area in km
2
of the urban class.
Table 4. Estimated area of urban class in km2
IMAGES LANDSA 5 - 1984 LANDSAT 5 - 2003 LANDSAT 5 - 2006 LANDSAT 5 - 2011
E
S
T
I
M
A
T
E
D

U
R
B
A
N

A
R
E
A

M-STAT. 57.6 71.8 74.6 77.1
K-MEANS 50.0 72.4 73.0 75.7
SVM 55.6 72.9 72.9 77.0
1984 2003
2011
6. CONCLUSION
In this study an urban growth monitoring approach has been proposed, where a GP was adopted to evolve simple
mathematical expressions to extract urban areas from Landsat image series. The model is built from a single image by
using a basic set of operators between spectral bands and maximizing a fitness function, which is based on the using of
the M-statistic separability criterion and k-means unsupervised classification.
The model was constructed from the Landsat 5 TM image acquired in 2006 by using training samples extracted with the
help of a Quick-bird high spatial resolution satellite image acquired the same day as the Landsat image over the city of
Oran, Algeria.
The obtained results show well the ability of the proposed GP based framework to generate spectral indices capable of
discriminate between urban area and the remaining objects in the image. Such indices are easy to implement in
commonly used remote sensing image processing software.
Finally the proposed multispectral image preprocessing and urban growth monitoring method requires further work. It
can be implemented using parallel architecture to speed up the process, improved by adding new functions (PC
transformation, filtering and texture analysis), or applied to other data (hyperspectral and high spatial imagery).
REFERENCES
[1] Momm, H. G., Easson, G., and Wilkins, D., Genetic programming as a preprocessing tool to aid multi-temporal
imagery classification, In Proceedings of the ASPRS 2006 Annual Conference. 1-11 (2006).
[2] Easson, G., and Momm, H. G., Evolutionary computation for remote sensing applications. Geography Compass,
4(3), 172-192 (2010).
[3] Rauss, P. J., Daida, J. M., and Chaudhary, S., Classification of spectral imagery using genetic programming,. Ann
Arbor, 1001, 48109 (2000).
[4] Espejo, P. G., Ventura, S., and Herrera, F., A survey on the application of genetic programming to classification,
Systems, Man, and Cybernetics, Part C: Applications and Reviews, IEEE Transactions on, 40(2), 121-144 (2010).
[5] Momm, H., and Easson, G., Evolving spectral transformations for multitemporal information extraction using
evolutionary computation,. Journal of Applied Remote Sensing, 5(1), 053564-053564 (2011).
[6] Kaufman, Y. J., and Remer, L. A., Detection of forests using mid-IR reflectance: an application for aerosol studies.
Geoscience and Remote Sensing, IEEE Transactions on, 32(3), 672-683 (1994).

Você também pode gostar