Aniati Murni

Integrated visual vocabulary in latent Dirichlet allocation–based scene classification for IKONOS image

Abstract Scene classification based on latent Dirichlet allocation (LDA) is a more general modeling method known as a bag of visual words, in which the construction of a visual vocabulary is a crucial quantization process to ensure success of the classification. A framework is developed using the following new aspects: Gaussian mixture clustering for the quantization process, the use of an integrated visual vocabulary (IVV), which is built as the union of all centroids obtained from the separate quantization process of each class, and the usage of some features, including edge orientation histogram, CIELab color moments, and gray-level co-occurrence matrix (GLCM). The experiments are conducted on IKONOS images with six semantic classes (tree, grassland, residential, commercial/industrial, road, and water). The results show that the use of an IVV increases the overall accuracy (OA) by 11 to 12% and 6% when it is implemented on the selected and all features, respectively. The selected features of CIELab color moments and GLCM provide a better OA than the implementation over CIELab color moment or GLCM as individuals. The latter increases the OA by only ∼ 2 to 3%. Moreover, the results show that the OA of LDA outperforms the OA of C4.5 and naive Bayes tree by ∼ 20 % .

Comparison of hybrid neural systems of KSOM-BP learning in artificial odor recognition system

This report proposes an adaptive recognition system, which is based on Kohonen self-organization network (KSOM). As the goals in the research on artificial neural network are to improve the recognition capability of the network and at the same time minimize the time needed for learning the patterns, these goals could be achieved by combining two types of learning, i.e. supervised learning and unsupervised learning. We have developed a new kind of hybrid neural learning system, combining unsupervised KSOM and supervised back-propagation learning rules. This hybrid neural system will henceforth be referred to as hybrid adaptive SOM with winning probability function and supervised BP or KSOM(WPF)-BP. This hybrid neural system could estimate the cluster distribution of given data, and directed it into predefined number of cluster neurons through creation and deletion mechanism. Comparison with other developed hybrid neural system is done for determination of various odors from Martha Tilaar Cosmetics product in an artificial odor recognition system. The performance of our developed learning system in term of its recognition ability and its learning time is explored in this report.

Evaluation of five feature selection methods for remote sensing data

This paper evaluates five potential feature selection methods in the application of remote sensing. The five methods include the sequential forward floating selection, the joint pair approach, band selection based on variance, the principal component transform, and the visual-based selection. Optical-sensor image and synthetic aperture radar image are used for experiments. Several recommendations are made based on this study. For optical-sensor images, the five feature selection methods: sequential forward floating selection, joint pair, band selection, principal component transform, and visual-based selection could have about the same classification accuracy using two to five selected features. The case study has shown that the sequential forward floating selection is the best feature selection method for both optical and synthetic aperture radar feature image selection, followed by the joint pair (for two-feature selection), visual-based selection, band selection, and principal component transform. For band L and band X synthetic aperture radar feature images, entropy, homogeneity, inverse difference moment, and maximum probability, East to West and West to East semivariogram, the local mean value, maximum, and minimum are the best features of the co-occurrence matrix model, semivariogram model, and local statistic model. For Landsat TM images band 7, 4, 5, 3, 1, and 2 are significant feature images. Applying the sequential forward floating selection to select two to five features from the potential features can obtain classification accuracy greater than 90 percent.

Geospatial data extrapolation using data mining techniques and cellular automata

This paper describes a geospatial knowledge discovery model of historical maps data set with relative geographic referenced. The knowledge about spatiotemporal dynamic is represented by the transition rules of cellular automata model. Set of transition rules obtained by applying three data mining techniques on large amount of data grid. First, multiple linear regression analysis applied on each subsequent pair of N data grid to obtained (N-1) rules. Second, by applying clustering analysis, then they extracted into a small number of rules, which is represented all of the rules, and they associated with the first data grid of the related pair. Finally, the selected rules used in determining the next value of the given data using classification analysis. Selection of the rule applied to the data based on the distance between the data and the associated data grid of the selected rule. The model had been evaluated on ordinal data type from fire danger rating and nominal data from land use and land cover status. Model accuration measured and visualized by comparing actual data and the simulated data. The accuration ranges between 80%-95% in the first case and 90,5%-95,2% in the second. In the first case, by the segmentation of the model, the performance can be improved significantly, especially for von Neumann scheme.

Proposal for multispectral image compression methods

This paper has proposed two image compression and decompression schemes for multispectral images. Two issues were considered in the proposed methods. The first issue is the possibility of applying the compression process directly to a set of multispectral images, where the standard JPEG should be applied to each individual image. Considering this issue, a compression and decompression method is proposed based on a hybrid of lower bit suppression and Karhunen-Loeve transform and named as KLT Hybrid. The second issue is the possibility of obtaining a general codebook for a bulky of typical data such as a set of hyperspectral images. Considering this issue, another compression and decompression method is proposed based on vector quantization (VQ) where the general codebook is obtained by a proposed fair-share amount method. Four performance indicators were used to evaluate the results. The indicators include compression ratio, root mean square error, maximum absolute error, and signal to noise ratio. The experimental results have shown good performance indication of both methods.

Identification of malignant skin cancer using back-propagation learning with Kanhunen-Loeve transformation

Malignant melanoma is the deadliest form of cancer, fortunately, if it is detected early, even this type of cancer may be treated successfully. In this paper, we present a novel network approach for the automated separation of melanoma from benign categories of cancer, which exhibit melanoma-like characteristics. To reduce the computational complexities, while increasing the possibility of not being trapped in local minima of the back-propagation neural network, we applied Karhunen-Loeve transformation technique to the originally training patterns. We also utilized a cross entropy error function between the output and the target patterns. Using this approach, for reasonably balance of training/testing set, about 94% of correct classification of malignant and benign cancers could be obtained.