PCA Clouds (ICET 2005)


Published on

Imran Sarwar Bajwa, S. Irfan Hyder [2005], "PCA Based Image Classification of Single-Layered Cloud Types", in 1st IEEE International Conference on Emerging Technologies (ICET 2005), Islamabad, Pakistan, Jan 2005, pp:365-369

  • Be the first to comment

  • Be the first to like this

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide

PCA Clouds (ICET 2005)

  1. 1. IEEE --- 2005 Interntational Conference on Emnerging Technologies September 17-18, Islamiiabad PCA based Image Classification of Single-layered Cloud Types Imran Sarwar Bajwa, Syed Irfan Hyder PAF- Karachi Institute of Economics and Technology imransbiawa(divahoo.com,nhyder(pafkiet. edu.pk Abstract [7]. PCA is a way of identifying patterns in data and The paper presents an automatic classification expressing the data in a way that highlights its similaritiessystem, which discriminates the diferent types ofsinigle- and differences. PCA strives to identify relatively fewerlayered clouds uising Principal Component Anialysis "features" or components that as a whole represent the(PCA) with enhanced accuracy as compared to other full object state and hence are appropriately termedtechniques. PCA is an image classification techniqute "Principal Components". Thus, principal componentstypically used for ace recognition. Principal extracted by PCA implicitly represent all the features.components are the distinctive or peculiar featuires of an However, these abstracted features may or may notimage. The approach described in this paper iuses this include a specific feature [5].PCA capability Jbr enhancing the accuracy of cloud Better accuracy in cloud classification meansimage analysis. To demonstrate this enhancement, a accurate categorization of clouds according to high, midsoJtware classifier system has been developed that and low levels. These high, mid and low-level clouds areincorporates PCA capability Jbr better discrimination of further classified in their particular sub classes illustratedclouid images. The system is first trained using clouid in Section 3.3. PCA can easily handle a large amount ofimages. In training phase, system reads mqjor principal data due to its capability of reducing data dimensionalityfeatures of the different cloud images to produce an and complexity, thus getting better results [4]. Theimage space. In testing phase, a new cloud image can be algorithm provides a more accurate cloud classificationclassified by comparing it with the specified image space that yield better and concise forecasting of rain.using the PCA algorithm. Keywords: PCA, single cloud type recognition, 1.1 PCA and Eigenvectorsprincipal components, eigenvectors, weather prediction, Training procedure in any classification system isimage recognition. significant and can be beneficial. Using various algorithms training can be performed. Algorithm used in current image classifier system for training is principal1. Introduction component analysis, whose major emphasis is to locate and depict the principal features of the given sample Weather forecasting applications use various pattern image [12].recognition techniques to analyze clouds information andother meteorological parameters. Neural Networks is an The Eigenvectors is a key capability used in PCAoften-used approach [3], [13] for image processing. Some analysis algorithm. Eigenvectors are defined to be astatistical methodologies like FDA [4], RBFNN [I] and related set of spatial characteristics [6] that computer usesSVM [12] are also being used for image analysis. These to recognize a specific cloud type. PCA technique usesmethodologies require more training time and have training and testing sets of images. Eigenvectors of thelimited accuracy of about 70% [ 11]. This level of covariance matrix is computed from the training set ofaccuracy often degrades classification of clouds, and images. These eigenvectors represent the principalhence the accuracy of rain and other weather predictions components of the training images [7]. Theseis reduced [15]. eigenvectors are often ortho-normal to each other. In the context of clouds classification, these eigenvectors wouldPrincipal Component Analysis (PCA) treats each image form the cloud space. They may not correspond directlyas an entity. A set of images is needed, which defines a to any cloud feature like height, width and density. Whenclass based on the core feature, derived from those the eigenvectors are displayed, they look like a ghostlyimages [6]. A single class can cover a certain number of cloud. They can be thought of as a set of features thatimages. The number of images in a class can be together characterize the variation between cloud images.ineffective but the image quality can invariantly impactthe overall image analysis results and consequences. PCA Cloud Detection consists in locating a cloud inis used to avoid computational intense calculations. Its complex scenery, by locating and cutting it out. Someuse results in fast and relatively more accurate inferences methods search elliptical and polygonal forms [2], others 0-7803-9247-7/05/$20.00 ©2005 IEEE 365 Authorized licensed use limited to: The Islamia University of Bahawalpur. Downloaded on July 27, 2009 at 04:01 from IEEE Xplore. Restrictions apply.
  2. 2. IEEE --- 2005 International Conference on Emnerging Technologies Septemtber 17-18, Islaminabadseek the texture and color of the clouds and still others in training set. After matching it infers that rather imageseek the patterns and boundaries of the cloud [3]. is recognized or not. PCA is used abundantly for image analysis and Clouid Tvpe Classificationclassification purposes [6] because it is a simple, non- This module finally detects and classifies the cloudparametric method of extracting relevant information type. Images are classified to their respective types on thefrom confusing data sets [1]. PCA extracts relevant basis of the matching inferences provided by previousfeatures from data by performing an orthogonal module.transformation. PCA also provides assistance to reduce acomplex data set to a lower dimension [7]. This paper 2.2 PCA based Extraction of the Attributesdemonstrates that PCA is relatively better than othertechniques in discriminating different types of single- The system presented in this work exemplifies thelayered cloud images. concept of Eigenvectors. These eigenvectors are a small group of characteristics extracted by the designed classifier system using PCA. PCA is a two-phase2. Methodology algorithm. * Training2.1 Systems General Architecture * Recognition The developed classifier system discriminates thesingle-layered cloud types. It carries out classification in 2.2.1 Trainingfive modules: image acquisition, detection, extraction of Training phase constructs an image-space, called arelated attributes, comparison of these attributes and cloud space, which is later required for classification infinally classification as shown in Fig. 1. testing phase. In training phase, the classifier system is trained by using sample data input. If it is required, output Cloud Extraction | pattem can be enhanced and improvised by retraining the Imag Detection of Attributes system by more refined and conspicuous data. Training is performed using n images in the following 6 steps: ass Cloudi Type Classificatio I1Inage Comparison < / STEP 1: Each sample image is converted into a row vector. A rowFig 1: General architecture of the cloud types classification vector can be constructed by concatenating each row withsystem first row in sequence. As in fig-2 a m x n matrix is converted into a single row 1 x mn vector X.Image acquirer This module helps to acquire new image. The imagecan be acquired through different sources e.g. digitalcamera. Images for testing and training phases areconverted to 256-bit Gray color image. Images are alsoscaled to 50 x 50 ratio. This ratio can vary from 30 x 30 X~~~~~~~~~~X x~~~~~~~~~~~xto 50 x 50. Fig 2: A row vector representation of a 2-D cloud imageCloued Detection STEP 2 This module detects the presence of the cloud The row vector matrix is constructed by combiningfragments in the images. In this module it is specified together the row vectors of n cloud images. Xi is a rowthat rather the images contain the clouds or not. vector of a sample image i, where i = 1 .. n.Extraction of Attributes This module identifies the various patterns in data.Cloud attributes are extracted from images using ~~~~~~~~~~~~~~...... 41J 1Principal Component Analysis algorithm. + . * Fi.:Woecoddsrbto.Rwvco fXlmage Comparison This module compares the principal features of thetest image with the image-space of already given images 366 Authorized licensed use limited to: The Islamia University of Bahawalpur. Downloaded on July 27, 2009 at 04:01 from IEEE Xplore. Restrictions apply.
  3. 3. IEEE --- 2005 International Conference on Emnerging Technologies Septemlber 17-18, IslamabadSTEP 3 4- Cloud Type Classification A mean cloud vector T of n row vectors is calculated 5- Evaluate Accuracyto extract required principal features. 6- Comparison with other Technologies 3.1 Data Collection VJ= (/)n) EX, Where i = I .. n Image data of clouds different types was obtainedSTEP 4 for training purposes. This data is available from differentA new matrix ( is constructed by subtracting mean cloud sources. Ground-based cloud images have been used invector ¶1 from each cloud image X of the training set. this experiment. These images of the general and sub- cloud types are available at different websites of worlds <P =XjI major weather forecasting organizations [13], [14] , [15].STEP 5 Overlapping sets of training images and testing imagesA data covariance matrix C is calculated by multiplying were used for the experiment. Global daytime cloudmatrix ( with its transpose matrix 0,. images are used in development and implementation aspects of a principal component analysis classification C= , ( - system. The designed image classifier system is used toSTEP 6 find the presence of clouds and classification of single- layer clouds in cloud images. Twenty highest valued eigenvectors are then pickedto make an image space from the resultant covariance 3.2 Normalizing Cloud Imagesmatrix C. Images for testing and training phase are of 256-bit *(StepS 6 are performed using Matlab 5.6) Gray color image and are overlapping. If the acquired image is not in specified bitmap format then it is2.2.2 Recognition converted into required format. The system obtains the In testing phase, each new image is analyzed and its image in the form of BMP of JPEG format.principal features are located. Then these principalfeatures are compared with the principal features of Acquired Image was of size 50 x 50 pixels forimage-space. If some match is found there, then the image processing in the designed system. But this ratio can beis classified according to the previously defined rules. tuned from 30 x 30 to 50 x 50. This module gets theRecognition or testing phase is performed in the image in integer or short co-ordinate i.e. perform scalingfollowing two steps. at 50 x 50 scale.STEP I 3.3 Define Classes A new cloud image is categorized by calculating An efficient and effective image classifier system oftenprojection Q on image-space by consists of a defined set of classes. These precisely defined classes are well separated by a set of features that n= U * (Z - ) are typically derived from the multi-dimensional Where Ui is image-space and Z is the new Image radiometric image data. The selection of classes is often influenced by desired application and classes may beSTEP 2 complicated. In this research, there are four defined If threshold (P matches with one of the thresholds in general classes and these general classes are furtherimage space then cloud recognition occurs and the divided into sub classes and they areparticular cloud type is specified. 1- Clear sky ,(i = I/k max (Qi -2) where (i,j= ..... n) 2- Low-level clouds i)- cumulus ii)- Stratocumulus3. Experiments iii)- Stratus 3- Mid-level clouds A series of experiments were done using the i)- Altocumulusdeveloped classified system to evaluate its accuracy. ii)- NimbostratusExperiments were performed using following steps: iii)- Altostratus 1- Data Collection 4- High-level clouds 2- Normalizing Cloud Images i)- Cirrus 3- Define Classes ii)- Cirrostratus iii)- Cirrocumulus 367 Authorized licensed use limited to: The Islamia University of Bahawalpur. Downloaded on July 27, 2009 at 04:01 from IEEE Xplore. Restrictions apply.
  4. 4. IEEE --- 2005 International Conference on Emiierging Technologies Septemnber 17-18, Islanmabad3.4 Cloud Type Classification 3.6 Comparison with other Techniques Two types of satellite images have been used, first as Various classification techniques and algorithms aretraining image and other images with clouds for testing. used for image classification. Each technique has its ownComparing the individual pixel values within 50 x 50 respective accuracy level. The derived results usingarray with a clear sky images depicts cloud fragments in a principal component analysis are compared with thesample image. Often the array of 32 x 32 array is used in results of other technologies used for cloud classification.conventional image recognition applications. As the Different technologies provide with different levelgreater number of pixels can immensely affect the accuracies.memory usage so array of smaller range is preferred. Butthis procedure also affects the overall image processing Results show that PCA, relative to other statisticalaccuracy. But PCA handles images so conveniently that techniques, is more accurate [table. 3]. Other statisticalan array of greater range may be used to get still higher techniques include Fuzzy Logic based systems that giveaccuracy. 84% accuracy [5] but Fuzzy systems are dependent on the appropriateness of the initial categories defined i.e. much If the cloud matches with the existing collected data effort is needed for domain knowledge and efficiencythen the program will display as match is found. It issues. Neural networks demand intense domaindisplays clouds general type as low, mid or high. knowledge and intuition for representation otherwiseProgram has also the capacity of prescribing the clouds suffer from divergent training sessions and inaccuratesub type and also describing the properties of each cloud results [11].sub type. PCA is the image classification technique, which3.5 Evaluate Accuracy provides higher accuracy up to 90%. Statistics show that PCA based image classifier system is a better classifier To test the accuracy of the designed system images than other used techniques. A comparison of PCA withwith clear sky and images with all types of single-layered other techniques is also given below.clouds are used. 17 Clear sky images were used fortesting and all images were successfully categorized. 36 Technology Name Accuracy ErrorImages of single-layered cloud types were used and _______ ______ ______ ______ ______ ______ Per. R atioshowed results with high accuracy. A matrix of results of PCA (Principal Component Analysis) 92.30% 0.9%testing images is shown below. NMF (Non-Negative Matrix Factoriz.) [8] 69.94% S.1°/%Classes Clear Low- Mid- High- BPNN (Back Propagation NN) [9] 71.80%/0 sky level level levelClear sky 17 0 0 0 RBFNN (Radial Basis Function NN) [1] 73.20% 7.3%Low-level 0 34 1 1 SVM (Super Vector Machinie) [ 12] 84.11%Mid-level 0 1 32 3 Fisher Discriminant Analysis [4] 64.00%/oHigh- 1 1 3 31 Table 1. Testing results of different cloud type images FLNN (Fuzzy Logic Neural Networks) [5] 81.00% 3.4% A matrix representing classification accuracy test (%) Wavelet Transforms [6] 78.30% 3.9%for cloud free and single-layered cloud types is Probabilistic Neural Networks [2] 86.01%constructed. Classification inferences of PrincipalComponent Analysis for different cloud types are shown K-SOM (K-Self Organizing Maps) NN [I0] 80.00%in the matrix. Overall classification accuracy for single- Table 3. Accuracy comparison in different techniqueslayered clouds is determined by dividing number ofcorrectly classified samples by the total number ofsamples. An accuracy test (%) table is shown here. 4. Conclusion & Future Work Clear Low- Mid- High- Classes Classes ~sky level level level PCA is an efficient identifier in terms of time and provides better accuracy in cloud image recognition. A Clear sky 100 --- --- 0.5 PCA-based system provides high speed processing with Low-level --- 94.1 0.5 0.2 relatively better accuracy. PCA also easily handles a large Mid-level --- 0.2 88.9 0.8 amount of data due to its capability of reducing data High- 0.3 0.3 0.8 86.2 dimensionality and complexity. PCA algorithm provides a more accurate cloud classification that infers better and Table 2. Total Accuracy 92.3% = 368 Authorized licensed use limited to: The Islamia University of Bahawalpur. Downloaded on July 27, 2009 at 04:01 from IEEE Xplore. Restrictions apply.
  5. 5. IEEE --- 2005 International Conference on Emnerging Technologies September 17-18, Islamabadconcise forecasting of rain. Probably, the more long-term [5] Bryan A. Baum, Vasanth Tovinkere & Jay Titlow, Ronaldweather forecasting is also possible. M. Welch, 1997. Automated Cloud Classification of Global AVHRR Data Using a Fuzzy Logic Approach. Journal of In this report only one type of cloud has been Applied Meteorology. pp 1519-1539.addressed and that is single-layered clouds. Other type is [6] Chi-Fa Chen, Yu-Shan Tseng and Chia-Yen Chen, 2003.Multi-layered clouds. Multi-layered clouds are 60 % to Combination of PCA and Wavelet Transforms for Face65% of total clouds. So their identification and Recognition on 2.5D Images. Conf of Image and Visionclassification is also a significant task. Computing03 26-28 November 2003 Future goal of the research is to analyze all cloud [7] Cristina Conde ,Antonio Ruiz and Enrique Cabello, 2003. PCA vs Low Resolution Images in Face Verification.types using satellite image. Satellite image contains Proceedings of the 12th International Conference on Imageclouds of all types and there is very complex infonnation. Analysis and Processing (ICIAP03).For simple cloud recognition, co-variance matrix has been [8] David Guillamet, Bemt Schiele, and Jordi Vitri. Analyzingused. To cover the whole scenario (all cloud types) Non-Negative Matrix Factorization for Image Classification.another type of matrix named, co-exist matrix can be Conference on Pattern Recognition ICPR 2002, Quebec,used. There is also need of generalizing the algorithm. Canada, August 2002This process may need much more effort. Furthermore, [9] Kishor Saitwal, r. Azimi Sadjadi and Donald Rinki, 1997. Athe accuracy estimations can be further improved using a multi-channel temporarily adaptive system for continuousmore carefully selected disjoint sets of image data for the cloud classification from Satellite Imagery.training set and the testing set. [10] Bin Tian, Mamood R. AzimiSadjadi, Thomas H. Vonder and Donald Rienki, 1999. Neural Network based cloud classification on satellite Imagery using textural features5. References [11] Simon Haykin, McMaster University, Ontario Canada 1998, Neural Networks a Comprehensive Foundation,[1 ] Su Hongtao, David Dagan Fetig, Zhao Rong-chun, 1997. 2/E, Prentice Hall publishers Face Recognition Using Multi-feature and Radial Basis Function Network. [12] D. Cao, 0. Masoud, D. Boley, N. Papanikolopoulos, 2004. Online Motion Classification using Support Vector Machine,[2] Bankert, R. L., 1994: Cloud classification of AVHRR IEEE 2004 International Conference on Robotics and imagery in maritime regions using a probabilistic neural Automation, April 26 - May 1. network. Jotirnal ofApplied Meteorology., 33, 909-918. Boulder, CO 80307. [13] VanderZwaag, B.J., Slump, C.H. and Spaanenburg, L. (2002) Analysis of neural networks for edge detection,[3] Barsi, A., Heipke, C., Willrich, F., 2002, Detecting road Proceedings. Pro-Risc02 (Veldhoven) pp. 580-586 junctions by Artificial Neural Networks - JEANS, International Archives oqfPhotogrammetrv. Remote Sensinsg [14] http://www.noaa.gov/ and Spatial Information Science (34) 3B, pp. 18-21 [15] http://www.eumetsat.de/en/area2/cgms[41 Seong-Wook Joo, December 2003. Face Recognition using [ 16] http://www.nottingham.ac.uk/meteosat/ PCA and FDA with intensity normalization. 369 Authorized licensed use limited to: The Islamia University of Bahawalpur. Downloaded on July 27, 2009 at 04:01 from IEEE Xplore. Restrictions apply.