ANT COLONY OPTIMIZATION ALGORITHM FOR FEATURE SELECTION AND CLASSIFICATION OF MULTISPECTRAL REMOTE SENSING IMAGE Lintao Wen1, Qian Yin1*, Ping Guo1, 2 Image Proc. & Pattern Recognition Lab., Beijing Normal University, Beijing 100875, China 2 School of Computer Science and Technology, Beijing Institute of Technology, Beijing 100081, China Email:
[email protected],
[email protected] 1
1.
INTRODUCTION
Remote sensing has proved a powerful technology for monitoring the earth’s surface and atmosphere at a global, regional, and even local scale in recent years. With the development of high techniques, large amount of data is acquired by different types of sensors, which provide repeated coverage of the planet on a regular basis. Nowadays, an increasing quantity of multisource remote sensing data acquired in many geographical areas is available. To investigate these data, we need to develop effective data processing techniques to take the advantage of such multisource characteristics. In particular, in the context of classification problems, if we combine more data features to construct input vector, it may provide an improvement in accuracy, which may be of important significance in real applications. In general, the spectral and texture information of multisource remote sensing image plays an important role in the classification process. Spectral feature is regarded as one of the important pieces of information for remote sensing image interpretation. This kind of feature can be used to characterize most important contents for various types of the remote sensing images. On the other hand, the texture feature describes the attribution between a pixel and the other pixels around it. Texture features represent the spatial information of an image, which can be regarded as an important visual primitive to search visually similar patterns in the image. However, in classification, it has the shortcomings if only adopting the texture analysis method, such as the edge between different classes may be incorrectly classified, because texture feature extraction had to be considered based on a small region, not a single pixel. Spectral feature such as gray value can be extracted based on a single pixel, but it has limitation as the representative information of an image. From literatures, we can find that most existing classification studies for remote sensing image adopt only simple spectral or texture feature, or investigate with independent manner [1]. However, in classification of multispectral remote sensing image, usually it is difficult to obtain the higher classification accuracy if only consider image’s spectral feature or texture feature alone [2], besides it does not take the advantage of multisource characteristics. In the previous work, we proposed the method which composes the spectral feature and the texture feature together to form a new feature vector in classifier’s input space the most effective features of the given remote sensing image can be represented in this way. Consequence, we can overcome shortcomings of only using the single spectral feature or texture feature, and raise the classification accuracy. However, how to find the optimal combination of spectral and textural features to form a multi-feature vector is a difficult problem. Ant Colony Optimization (ACO) is a recently proposed metaheuristic approach for solving hard combinatorial optimization problems [3]. In order to get a better result in the classification, in this paper ,we presents a new approach by applying ACO algorithm to find a multi-feature vector composed of spectral and texture features. 2.
ALGORITHM AND METHODS
Before applying ACO algorithm, we should extract features from multisource remote sensing image first 2.1. Feature extraction In this study, we use five feature extraction (dimension reduction) methods to acquire the most available spectral features from multispectral images, including the Euclid distance measurement method (EDM), the discrete measurement criteria function method (DMCF), the minimum differentiated entropy method (MDE), the probability distance criterion method (PDC), and the principle component analysis method (PCA); We adopt four approaches for texture feature modeling problem, these are the gray level co-occurrence matrix (GLCM), the histogram measures (HM), the texture spectrum (TS), and the gray difference statistical quantity (GDSQ).
2.2. Applying ACO to find an optimal combination of features To reformulate the feature combination task into an ACO-suitable problem, it requires the problem to be represented as a graph. In the graph, nodes represent features, while the edges between nodes denote the choice of the next feature. The search for the optimal feature subset is then described as an ant traversal through the graph where a minimum number of nodes are visited that satisfies the traversal stopping criterion. ACO is a metaheuristic approach, which used to guide other heuristics in order to obtain better solutions than those that are generated by local optimization methods. We chose a subset evaluation function based on entropy to evaluate the heuristic desirability of traversing between features. In image classification, subset minimality and Āgoodnessā are two key factors so the pheromone update should be proportional to Āgoodnessā and inversely proportional to size. C-means classifier has been used to evaluate the “goodness” for the subset. During ants’ graph traversal, this is the probabilistic transition rule, denoting the probability of an ant k at feature i choosing to move to feature j at time t: k ij
p (t ) Where
k
Ji
¦
[W ij (t )]D [Kij ]E lJ ik
[W il (t )]D [Kil ]E
is the set of ant k’s unvisited features, K ij is the heuristic desirability of choosing feature j when at feature i
and W ij (t ) is the amount of virtual pheromone on edge (i, j).
D
and
E
are parameters to control impact of pheromone and
heuristic information to the probability. 2.3. Image classification In principle, we can construct a high dimensional vector using as many spectral and textural features as possible. However, it may not be suitable in improvement of accuracy because redundant information may degrade it. In the experiments, the combination of various spectral features with the different texture features together is investigated intensively with the ACO algorithm. Comparative studies also conducted with previous research work. In our experiment, we set a max dimension of the final vector each time, and use this parameter as a terminal condition. These are steps of our experiment: 1. Set initial parameters. 2. Place ants randomly on the graph (i.e. each ant starts with one random feature) and make them construct the feature vector. Use vectors ant constructed in classification and updates the pheromone. 3. Rerun step 2 until the stopping criterion satisfied. Then get the best vector and classify result. 3.
CONCLUSIONS
By analyzing the experimental results, we can draw the conclusion that ACO algorithm is helpful in searching optimal combination of the features used in multispectral remote sense image’s classification. Using the combination of the spectral and texture features obtained by ACO in classification always get a higher accuracy. And it is also possible to select feature in a high dimension feature space. 4.
REFERENCES
[1] Jiang Li, and R. M. Narayanan, “Integrated Spectral and Spatial Information Mining in Remote Sensing, Imagery”, IEEE Trans. on Geosciences and Remote Sensing, Vol. 42, No. 3, pp.673~684, 2004. [2] Qian Yin and Ping Guo, “Multispectral Remote Sensing Image Classification with Multiple Features”, Proceedings of the Sixth International Conference on Machine Learning and Cybernetics, pp.360-365, 2007. [3] Marco Dorigo and Thomas Stützle, Ant Colony Optimization, The MIT Press, Cambridge, MA, USA, 2004.