Comparison of Frequency-based Contextual and Maximum Likelihood Methods for Land Cover Classification in Arid Environment
M.Z. Mat Jafri
The classification accuracy obtained from the classification of satellite images using pixel-by-pixel conventional methods can be improved if the contextual information is considered during the classification process. This study presents a comparison of frequency-based contextual and maximum likelihood approaches to identify the land cover patterns in arid environment of multi-spectral images collected by SPOT-2 satellite. In image classification, in order to obtain a good result, not only the image resolution is considered but the selection of the classifier to be used during decision making process is important as well. In present study, two classifiers have been experimented in order to evaluate their performances which is Maximum Likelihood classifier representing as conventional method whereas contextual approach representing as advanced method. Conventional classification methods commonly cannot handle the complex landscape environment in the image. The result of each method has often a salt and pepper appearances which is a main characteristic of misclassification. It seems clear that information from neighbouring pixels should increase the discrimination capabilities of the pixel based measured and thus, improve the classification accuracy and the interpretation efficiency. This information is referred to as spatial contextual information. The experimental results indicated that frequency-based contextual algorithm with 83.7% overall accuracy and 0.693 Kappa coefficient is more reliable than the maximum likelihood algorithm with 72.1% and 0.527 overall accuracy and Kappa coefficient, respectively. The high value of the frequency-based contextual classification is due to the fact that this algorithm could overcome the mixed pixel problem and reduce the speckle error in the image significantly.
May 31, 2011; Accepted: August 01, 2011;
Published: September 16, 2011
Image classification is an important part in many remote sensing applications
especially for detecting the land use/cover types. Remote sensing is a valuable
data source from which land cover information can be extracted efficiently (Chen
et al., 2008). In the past decades, there has been a growing trend
in the development of land cover map using remote sensing data. Land cover is
a fundamental variable that impacts on and links with many parts of the human
and physical environment (Foody, 2002). Effective classification
of remote sensing image data depends upon separating land cover types of interest
into sets of spectral classes that suited to the particular classifier algorithm
used. In present study, land cover for the selected date was estimated using
the supervised classification. To effectively derive reliable information from
satellite data, appropriate classification techniques are essential. A number
of classification approaches have been developed over the past decades. The
classifiers can be categorized as either common or advanced. Some of the common
classification algorithms include the K-Means, ISODATA, maximum likelihood classification
and minimum distance to means (Mather, 2004; Lillesand
and Kiefer, 1999). The advanced classification algorithms include the Artificial
Neural Network (ANN), contextual, decision trees, support vector machines and
object based image analysis (Lawrence et al., 2004;
Keuchel et al., 2003; Verbeke
et al., 2004; Lucieer, 2008; Hay
et al., 2003; Xu et al., 2003; Mustapha
et al., 2010a). But in present study, only contextual classification
and maximum likelihood classifier were highlighted. Spectral imagery has been
the primary tool for scene classification. In recent years, the progress of
computer capabilities makes spatial feature processing techniques practical
to implement in pursuit of improvement in classification accuracy (Olsen
et al., 2002). Contextual information is one kind of such spatial
relationship and has drawn our particular interest for remotely sensed imagery
interpretation shown in present study. In the last years, efforts have been
devoted to develop contextual classifier (Arai, 1993;
Kontoes and Rokos, 1996). Classifiers that incorporate
contextual information into classification have been reported in the literature
as well (Chica-Olmo and Abarca-Hernandez, 2000; Atkinson
and Lewis, 2000).
One of the advantages of classifying land cover types in arid environment is
that it free from cloud. Cloud covers are generally thicker in the equatorial
region (Sarkar and Kumar, 2002) and it causes the main
problem to the researcher when using remote sensing technique in their research.
Although, the cloud free condition in the arid area but there is still challenging
to do the classification in this area due to the similarity of spectral characteristics
among the land cover types. This situation would create the high possibility
of mixed pixels to be occurred. Essentially, mixed pixels problem mostly appeared
in urban class due to the large amount of spectral information are extracted
in that particular region. Mixed pixels problem would lead the difficulty in
decision making process. As stated by Lu and Weng (2006)
urban land use/cover classification is still challenge with medium or coarse
spatial resolution remotely sensed data due to the large number of mixed pixels
and the spectral confusions among different land use/cover types. The mixed
pixel problem would require additional physical information for a good classification.
For that reason, the classification process needs the context information to
be considered instead of depending on spectral information alone. A contextual
classifier consistently produces higher classification accuracies than the per-pixel
classification (Kontoes and Rokos, 1996; Stuckens
et al., 2000). The purpose of present study is to highlight the potentiality
of using different classifiers to generate important land cover information
for the Mina city area using SPOT data.
MATERIALS AND METHODS
Description of study area: The study area is located in the Makkah province
which is in the western of the Saudi Arabia (Fig. 1) and was
conducted in the year of 2010. Mina city is located between Makkah and Arafah.
Mina is the city known as Tent City where the pilgrims travel to nearby mount
Arafah to spend the night camping in Mina valley during the Hajj season. Following
a series of fatal disasters culminating in a large fire in 1997, the Saudi government
decided to replace the existing cotton structure with fireproof Teflon coated
glass tents (Grundig, 2002). This selected area was
bounded by mountainous and desert terrain in every direction.
|| Location of the study area
The study area in the Arabian Peninsula is located on 21°24N to
21°26N latitude and 39°50E to 39°54E longitude.
Materials: The materials used for the classification were SPOT-2 satellite data of Mina City, Saudi Arabia and its surrounding areas which was acquired on January 2010 under clear weather conditions. It provides better spatial land-coverage maps and land-use classification maps for monitoring regional environments. The SPOT-2 image has three reflective bands with 20 m spatial resolution. In this research, all reflective bands were used in the processing and image analysis. This band contain the visible and near infrared band [Band 1: 0.50 to 0.59 μm (green band); Band 2: 0.61 to 0.68 μm (red band) and Band 3: 0.78 to 0.89 μm (near infrared band)].
Pixel based versus contextual classification: Digital image processing
is the numerical manipulation of digital image and includes pre-processing,
enhancement and classification. Image classification refers to the extraction
of differentiated classes or themes, usually land cover and land use categories,
from raw remotely sensed digital satellite data. The information contained in
a remotely sensed image and can be used to conduct image classification includes
spectral pattern, spatial pattern and temporal pattern. Spectral pattern is
the combination of Digital Numbers (DNs) for different feature types. Spatial
pattern refers to the spatial relationship of the pixels, such as image texture,
pixel proximity, features size and shape. Temporal pattern refers to temporal
characteristics of the features (Qian et al., 2010).
A wide range of classification algorithms has been developed to derive land
use and land cover information from remotely sensed images. Since remotely sensed
images consist of rows and columns of pixels, pixel-based classification becoming
the conventional method for land cover mapping. This classification method assigns
a pixel to a class fundamentally according to the spectral similarities (Gong
et al., 1992; Casals-Carrasco et al.,
2000). The unsupervised classification approach provides an automated platform
for image analysis, mainly based on surface reflectance and generally ignoring
basic land cover characteristics (i.e., shape and size) of landforms (Chust
et al., 2004). The supervised classification approach can preserve
the basic land cover characteristics through statistical classification techniques
using a number of well distributed training pixels. The Maximum Likelihood (ML)
classification method is well known for the analysis of satellite images. So
far, satellite image interpretation using the ML approach was mostly applied
for land cover classification (Huang et al., 2007)
and monitoring of land use changes (Shalaby and Tateishi,
2007). Although, the techniques are well developed and many successful applications
have been reported, it suffers from ignoring the spatial pattern in decision
making process. The ML classifier quantitatively evaluates both the variance
and covariance of the category spectral response patterns when classifying an
unknown pixel without considering contextual information. ML is a statistical
classifier so that it needs many training data for the classification process.
Therefore, with very large of training sets, it can generate the statistical
distribution and used this data for classifying the images (Mustapha
et al., 2010b). ML is a parametric classifier that assumes normal
spectral distribution for each feature of interest and an equal prior probability
among the classes is also assumed. This classifier is based on the probability
that a pixel belongs to a particular class.
Usually, the spectral signature is the main aspect of the classes used to classify
the pixels. Unlike traditional pixel-based methods, contextual technique considering
both spectral and spatial information in order to perform the classification
process instead of depending on spectral component alone. The integration of
spatial information in the image classification is expected to improve classification
accuracy. The contextual classifier uses pixel-centred window to estimate the
density function associated to a pixel. Selection of the window size is very
important in contextual classification. Pixel-window size determines the amount
of spatial information that can be included in the classification. Since optimal
pixel window varies with individual class and image resolution, it is usually
difficult to determine before image classification. Therefore, an appropriate
window size is usually determined empirically (Huang et
al., 2007). In term of the training data, contextual classifier do not
required large number of training set as this classifier is not a statistical
Methodology: Present study focuses on extracting land cover by adopting
ML and contextual approaches for SPOT-2 data. The reason for implementing this
mechanism is to perform comparison between the traditional and advanced approaches
in land cover classification. The results of these methods were evaluated using
the available field information on land cover and by visual interpretation.
In the Mina city, 4 classes were selected to represent and classify the image
namely Urban, Mountain, Mina Tent and Shadow. Although the shadow appears mostly
in the mountainous area, the authors decided to separate them as a new class
due to their obviously different in their spectral information against Mountain.
The description for each of the land cover category is stated in Table
1. The ground resolution for this satellite is 20 m and the image was acquired
using High Resolution Visible (HRV) sensors carried on the SPOT satellite. It
has three bands ranging from visible to near infrared portion. The total work
for present study has been executed using PCI Geomatica 10.3 image processing
software packages. The size of data set used in the image processing is 450x350
pixels and is presented in Fig. 2.
|| Description of each land cover category
|| The raw data of the study area (SPOT-2)
Prior to the image analysis, image pre-processing step was applied to make
the images comparable before classification. A composite colour has been done
on this set. The aim of this pre-processing is to have a better visual interpretation
of the scene and to be able to identify representative areas which will constitute
a training base for the supervised process. Two sets of training data were used
in order to perform the classification. One set of training data was used for
ML approach and the other set for Frequency-based Contextual (FBC) classifier.
The use of different set of training data is due to the ML approach needs more
training samples as it is a statistical classifier so that it can generate the
statistical distribution of each class for classification. Unlike ML approach,
contextual do not required to have many training data as this classifier is
not a statistical approach. The supervised classification of digital data is
carried out after creating training sites and the task was performed and the
classification was carried out using ML and FBC classification methods (Gong
and Howarth, 1992). Basically, there are many algorithms (e.g., minimum
distance, neural network, decision tree etc.) can be used in image classification
but in this study only these two classifiers were chosen for the classification
process. The selection of the classifier is very important in order to obtain
good classification accuracy. The classification accuracy is most important
aspect to assess the reliability of maps, especially when comparing different
techniques. During present study the accuracy has been estimated on the training
set samples and the results of all the classification techniques is summarized
in the table. A total of 190 points were selected randomly in order to test
the accuracy assessment of the classification result. A statistical classification
assessing is carried out by means of error matrix establish between truth ground
and obtained classifications. Performance of the obtained classifications is
evaluated by calculating kappa parameters which is a usual indicator in land
cover classification analysis. The error matrices related to the classification
of Fig. 4a and b are given in Table
2 and 3. The methodology used in this project is summarized
in Fig. 3.
An accuracy assessment: The accuracy assessment sites were used to provide
a statistically sound assessment of the accuracy produced by each of the automated
mapping approaches tested for this project. The accuracy assessment sites were
set aside until the map was completed and accuracy assessment was performed.
This process insured that the accuracy data were completely independent of the
training data (Thomas et al., 2003).
|| Procedures of the image classification
A common method for classification accuracy assessment is through the use of
an error matrix (also called confusion matrix). For a classified image, an error
matrix can be made by comparing the classification results with reference data.
In this matrix, the reference data are represented by the columns of the matrix
while the classified data are represented by the rows. The major diagonal of
the confusion matrix indicates the agreement between these two data sets. It
is typical to extract several statistics from the error matrix: overall accuracy,
producers accuracy and users accuracy. Congalton
(1991) used the terms Producer Accuracy (PA) and User Accuracy (UA) to describe
within class measures of classification accuracy and thus provide a breakdown
of the figure for overall accuracy in error matrix. PA gives the analyst an
estimate of how successful the classification procedure is in the different
classes. UA gives the user an estimate of how reliable the thematic map is as
a predictive device in the different classes.
RESULTS AND DISCUSSION
From the matrices, the kappa coefficient, overall classification accuracy and
accuracy of each class is easily calculated. The resulted classified imagery
using context is finding to reveal meaningful patterns. Visual interpretation
between classified images reveals that image produced by contextual approach
is better than ML. The major improvement of the overall accuracy between the
conventional ML classifier and the FBC classification method is the speckle
error. The high content of noise in Fig. 4a produces a less
precise of the classification accuracy than in Fig. 4b. Overall
accuracy of contextual classification was around 11 percent better compared
to the ML classification. It was shown that the involvement of the spatial information,
speckle error (salt and pepper appearances) can be reduced significantly. This
visual interpretation is confirmed by the statistical information given on Table
2 and 3.
||Confusion matrix table derived from Maximum Likelihood classifier
|Overall accuracy: 72.1%, Kappa coefficient: 0.527
||Confusion matrix table derived from Frequency-based Contextual
|Overall accuracy: 83.7%, Kappa coefficient: 0.693, * PA: Producer
accuracy, UA: User accuracy
Overall accuracy of FBC classification with the integration of spatial information
leads to impressively improved results, up to 83% of accuracy with 0.693 kappa
value is achieved in comparison with the output derived from traditional ML
classifier where only around 72% of accuracy with 0.527 kappa is obtained. This
result confirms the experience that the incorporation of contextual information
tends to become more accurate than statistical ML approach (Gong
and Howarth, 1992; Stuckens et al., 2000;
De Jong et al., 2001; Jackson
and Landgrebe, 2002). For instance, Gong and Howarth
(1992) reported that an improvement of classification accuracy of 15-20%
can be achieved using the frequency-based classifier over the ML approach. In
the meantime, the incorporation of contextual information in the classification
process improved accuracy by 5.8% as stated by Stuckens
et al. (2000) and Jackson and Landgrebe (2002),
they could increase 13% of accuracy when using contextual technique compared
to ML approach. In addition, De Jong et al. (2001)
resulted 80.1 and 92.7% of classification accuracy for ML and contextual, respectively,
hence, the contextual outperform ML classifier by approximately 12%.
In addition, the accuracies for each class were presented in the error matrix
tables and can be found by evaluating the user and producer accuracies column.
For ML approach, the user accuracy varied between 60.8% for Urban and 85.3%
for Mountain. Mina Tent and Shadow has an accuracy of 75.0 and 83.3%, respectively
(Table 2). In the meantime, the user accuracy for the FBC
method varied between 73.9% for Urban and reached the highest value of 100%
for Mina Tent, 85.7 and 89.2% were obtained from Shadow and Mountain classes
|Fig. 4 (a-b):
||Classified land cover map over Mina area using (a) Maximum
Likelihood classification and (b) Frequency-based Contextual classification
This statistical evaluation shows that all classes have better accuracy when
performed by FBC method compared to ML approach. For producer accuracy, the
accuracy for each class using ML approach was as follow: 90.0% for Mina Tent,
62.7% for Mountain, 81.9% for Urban and 83.3% for Shadow (Table
2) whereas the results obtained from FBC method were 37.5, 87.6, 82.3 and
85.7% for the same sequence (Table 3). Again FBC has a better
result compared to ML approach in all classes except the Mina Tent class. For
this class, ML classifier performed better than contextual approach as shown
by the result from visual interpretation as well as from error matrix table.
In FBC method, most of the Mina Tent class was classified as urban class due
to the high pixel confusion among the classes. Overall, the classification result
performance of the ML method is poor, especially in terms of delineating built
up area from surrounding features in the arid environment. In arid zone of Saudi
Arabia, the urban areas are often surrounded by mountain. However, they may
also confuse with nearby bare soil and stony desert which present very similar
spectral characteristics as construction materials such as concrete. The urban
environment represents one of the most challenging areas for remote sensing
analysis due to the high spatial and spectral diversity of surface materials.
The detail explanation of the each class can be explained by analyzing the error matrix table. Table 2 presents the error matrix for the ML classifier. The matrix shows that 64 out of 75 points of the Mountain class had been correctly classified. Only 11 points had been classified wrongly which is all points were misclassified as Urban. For the Urban category, it is shows that out of 97 pixels that had been tested, 59 pixels were correctly classified and most of the remaining pixels were misclassified as Mountain. In addition, the same pattern was obtained for FBC method where Mountain and Urban classes were confusion among each other. For the Mountain, 99 out of 111 observations had been correctly classified and 11 points were misclassified as Urban. A total of 51 out of 69 observations had been correctly classified for Urban and 18 points were wrongly classified with 13 points were misclassified as Mountain (Table 3). Although, FBC has a better result but both of the classifiers have the difficulty to classify these classes due to the spectral confusion among them is high. This situation happens due to the fact that the mountainous area in the arid environment is not cover by tree but it is filled by stones and rocks which is has a similar spectral characteristic of urban area. On the other hand, both classifiers perform well for the Mina Tent and Shadow classes as their spectral characteristics significantly different against other features. For Mina Tent, 9 out of 12 observations were correctly classified for ML (Table 2) whereas all points were correctly classified for FBC (Table 3). From 6 points that have been tested for Shadow class on ML approach, 5 points were correctly classified and only a point was misclassified for FBC method (6 out of 7 points) for this class (Table 3).
The traditional pixel-based classification typically yielded large uncertainty in the classification results as it cannot handle the complex landscape environment. Aim to this problem, the contextual image classification was processed. In contextual classification method, the analyst can adjust the classification process by changing the growth of the kernel window. In the meantime, 11x11 windows were used to classify the images for this assessment. Selection of the window size is performed by trial and error basis and the 11x11 window was chosen as the best window to perform the classification. Essentially, selection of the window size is depending on the complexity of the image. If the image is more complex, hence, FBC needs the larger window for classification process. Otherwise, smaller window is enough for the smooth images.
The analysis presented in present study also shows that mixed pixels cause spectral confusion and that such pixels are sometimes assigned to the wrong thematic land cover class. The contextual method makes it possible to identify mixed pixels in an image and assign these mixed pixels to the proper land cover class. Basically, mixed pixels occur at the border of the classes and in the complex landscape environment such as in the urban/built up area. In a per pixel classification, these pixels are often wrongly labelled due to the mixed spectral signature of the pixels. By using ML classifier, some of the Urban is classified as Mountain as their spectral signature resembles. Results of the classified image by using contextual approach show that image classification can significantly improve by capturing spatial information in the classification procedure for the same area.
Present study underlines the performances of two widely used classification
techniques for classification of Mina area using SPOT-2 satellite image. The
performance of FBC classification technique is reliable assessing high accuracy
followed by ML. The landscape of Mina city is diverse and complex, comprising
both homogeneous and heterogeneous surface features, causing problems of spectral
variability in the satellite image data. In general, the use of contextual information
indicated larger improvements in the classification accuracy. For present study,
the use of the contextual information could be improved the classification accuracy
up to approximately 11%. The accuracy assessment report showed that FBC algorithm
can predict land use and land cover of the complex urban environment more accurately.
It also means this classifier shows great potential for dealing with heterogeneous
surface features in urban areas. In conclusion, overall performance of FBC is
better than ML but in certain homogeneous surface such as Mina Tent, ML performed
better than FBC.
Arai, K., 1993. A classification method with a spatial-spectral variability. Int. J. Remote Sensing, 14: 699-709.
Atkinson, P.M. and P. Lewis, 2000. Geostatistical classification for remote sensing: An introduction. Comput. Geosci., 26: 361-371.
Casals-Carrasco, P., S. Kubo and B.B. Madhavan, 2000. Application of spectral mixture analysis for terrain evaluation studies. Int. J. Remote Sensing, 21: 3039-3055.
Chen, X., J. Chen, M. Shen and W. Yang, 2008. Land use/land cover change detection using change-vector analysis in posterior probability space.. Proceedings of the Geoinformatics 2008 and Joint Conference on GIS and Built Environment: The Built Environment and its Dynamic, June 28-29, Guangzhou, China, pp: 714405.1-714405.8.
Chica-Olmo, M. and F. Abarca-Hernandez, 2000. Computing geostatistical image texture for remotely sensed data classification. Comput. Geosci., 26: 373-383.
Chust, G., D. Ducrot and J.L.L. Pretus, 2004. Land cover mapping with patch-derived landscape indices. Landscape Urban Planning, 69: 437-449.
Direct Link |
Congalton, R.G., 1991. A review of assessing the accuracy of classifications of remotely sensed data. Remote Sens. Environ., 37: 35-46.
CrossRef | Direct Link |
De Jong, S.M., T. Hornstra and H.G. Maas, 2001. An integrated spatial and spectral approach to the classification of Mediterranean land cover types: The SSC method. Int. J. Applied Earth Observ. Geoinform., 3: 176-183.
Foody, G.M., 2002. Status of land cover classification accuracy assessment. Remote Sens. Environ., 80: 185-201.
CrossRef | Direct Link |
Gong, P. and P.J. Howarth, 1992. Frequency-based contextual classification and gray-level vector reduction for land use identification. Photogr. Eng. Remote Sens., 58: 423-437.
Direct Link |
Gong, P., D. Marceau and P.J. Howarth, 1992. A comparison of spatial feature extraction algorithms for land use classification with SPOT HRV data. Remote Sensing Environ., 40: 137-151.
Direct Link |
Grundig, L., 2002. High performance cutting pattern generation of architectural textile structures. Proceedings of Fourth International Colloqium on Computation of Shell and Spatial Structures, June 5-7, Chania-Crete, Greece, pp: 1-15.
Hay, G.J., T. Blaschke, D.J. Marceau and A. Bouchard, 2003. A comparison of three Image-object methods for the multiscale analysis of landscape structure. ISPRS J. Photogrammetry Remote Sensing, 57: 327-345.
Direct Link |
Huang, H., J. Legarsky and M. Othman, 2007. Land cover classification using Radarsat and Landsat imagery for St. Louis, Missouri. Photogrammetric Eng. Remote Sens., 73: 37-43.
Direct Link |
Jackson, Q. and D.A. Landgrebe, 2002. Adaptive bayesian contextual classification based on markov random fields. IEEE Trans. Geosci. Remote Sens., 40: 2454-2463.
Keuchel, J., S. Naumann, M. Heiler and A. Siegmund, 2003. Automatic land cover analysis for Tenerife by supervised classification using support vector machines. Remote Sensing Environ., 86: 530-541.
Kontoes, C.C. and D. Rokos, 1996. The integration of spatial context information in an experimental knowledge-based system and the supervised relaxation algorithm-two successful approaches to improving Spot-xs classification. Int. J. Remote Sensing, 17: 3093-3106.
Lawrence, R., A. Bunn, S. Powell and M. Zambon, 2004. Classification of remotely sensed imagery using stochastic gradient boosting as a refinement of classification tree analysis. Remote Sensing Environ., 90: 331-336.
Direct Link |
Lillesand, T.M. and R.W. Kiefer, 1999. Remote Sensing and Image Interpretation. John Willey and Sons Ltd., New York.
Lu, D. and Q. Weng, 2006. Use of impervious surface in urban land use classification. Remote Sensing Environ., 102: 146-160.
Direct Link |
Lucieer, V.L., 2008. Object-oriented classification of side scan sonar for mapping Benthic marine. Int. J. Remote Sensing, 29: 905-921.
Direct Link |
Mather, P.M., 2004. Computer Processing of Remotely Sensed Images: An Introduction. John Willey and Sons Ltd., West Susses, England.
Mustapha, M.R., H.S. Lim and M.Z. Mat Jafri 2010. Comparison of neural network and maximum likelihood approaches in image classification. J. Applied Sci., 10: 2847-2854.
Mustapha, M.R., H.S. Lim, M.Z. Mat Jafri and F.M. Hassan, 2010. Performances of frequency-based contextual classifier in land use/cover classification using high resolution satellite images. Proceedings of the Multispectral, Hyperspectral and Ultraspectral Remote Sensing Technology, Techniques and Application III Conference, Oct. 13-14, Incheon, Korea, pp: 78570-78570.
Olsen, R.C., J. Garner and E.V. Dyke, 2002. Terrain classification in urban wetlands with high-spatial resolution multi-spectral imagery. Proc. SPIE, 4881: 686-691.
Qian, J., Q. Zhou and X. Chen, 2010. Improvement of urban land use and land cover classification approach in arid areas. Proceedings of the Image and Signal Processing for Remote Sensing XVI Conference, Sept. 20-22, Toulouse, France, pp: 78301-78301.
Sarkar, S.K. and R. Kumar, 2002. Cloud characteristics and cloud attenuation in millimeter wave and microwave frequency bands for satellite and remote sensing applications over a northern Indian tropical station. Int. J. Infrared Millimeter Waves, 23: 937-944.
Shalaby, A. and R. Tateishi, 2007. Remote sensing and GIS for mapping and monitoring land cover and land use changes in the Northwestern coastal zone of Egypt. Applied Geography, 27: 28-41.
Stuckens, J., P.R. Coppin and M.E. Bauer, 2000. Integrating contextual information with per-pixel classification for improved land cover classification. Remote Sensing Environ., 71: 282-296.
Direct Link |
Thomas, N., C. Hendrix and R.G. Congalton, 2003. A comparison of urban mapping methods using high-resolution digital imagery. Photogram. Eng. Remote Sens., 69: 963-972.
CrossRef | Direct Link |
Verbeke, L.P.C., F.M.B. Vabcoillie and R.R. De Wulf, 2004. Re-using back propagating artificial neural network for land cover classification in tropical savannahs. Int. J. Remote Sensing, 35: 2747-2771.
Direct Link |
Xu, B., P. Gong, E. Seto and R. Spear, 2003. Comparison of gray-level reduction and different texture spectrum encoding methods for land use classification using a panchromatic Ikonos image. Photogrammetric Eng. Remote Sensing, 69: 529-536.
Direct Link |