Detection of Sangeang Api Volcanic Ash Cloud Based on VBICA-SVM method

Abstract

Volcanic ash cloud can threaten the safety of aviation transportation. The current detection of the volcanic ash cloud mainly depends on the remote sensing images. Independent component analysis (ICA) and support vector machine (SVM) have very strong complementary to each other in the volcanic ash cloud detection. Taking the moderate resolution imaging spectroradiometer (MODIS) images of Indonesia Sangeang Api volcanic ash clouds as data source on May 30, 2014, this paper propose the variable Bayesian ICA and SVM (VBICA-SVM) method for detection of the volcanic ash cloud, which is based on the analysis of the physical properties of volcanic ash cloud and the absorption spectral characteristics in the scope of thermal infrared bands. The result shows that the VBICA-SVM method can effectively detect the spatial distribution of the volcanic ash clouds, and has great potential applications in the ash cloud monitoring and disaster mitigation. Introduction Remote sensing has the advantages of wide coverage, large amount of information and short revisit cycle, and widely used in the resource investigation, environmental monitoring, etc. [1, 2] Compared with the traditional digital image, the remote sensing images contain more diverse terrain types, therefore, the objects distribution in remote sensing image is more complex. The ash cloud monitoring using remote sensing technique is a huge system project, and the key is how to accurately identify the spatial distribution of the ash cloud [3, 4]. Although the researchers at home and abroad used all kinds of remote sensing data and methods for identification study on some large-scale volcanic ash cloud [5], most of these methods are based on the mathematical model, and can’t completely meet the requirements of actual ash cloud detection. As a new signal processing technology, ICA has close association with the problem of blind source separation, and it has the ability of separating source signals from the mixed ones without the need of prior knowledge [6, 7]. The SVM can realize the non-linear classification of objects in remote sensing image. Both have a great advantage in remote sensing images processing. At present, the application field of the ICA and SVM has expanded from solution of mixed blind signals to signal detection, image processing and pattern recognition, etc. But the ICA model has problems containing the independence and invariance of components, no noise assumption as well as the uncertainty of the final solution [8], and SVM has problem such as small sample properties. Given this, this paper proposes VBICA-SVM method used for detection of volcanic ash cloud from remote sensing images. In the proposed method, the Bayesian network is introduced into the ICA model, we used Bayesian inference to complete the study of unknown hidden variables (independent components), and we optimized it by combination with the variational approximation algorithm to make the isolated independent components as consistent as the real situation of the surface. And then the obtained independent component of VBICA make as the inputted feature vector of SVM. The first section of our paper is the introduction. The second section is the theoretical basis of ICA and SVM method. The third section is the detection of Indonesia Sangeang Api volcanic ash cloud based on VBICA-SVM method. The fourth section is precision evaluation of Indonesia International Conference on Advances in Mechanical Engineering and Industrial Informatics (AMEII 2015) © 2015. The authors Published by Atlantis Press 782 Sangeang Api volcanic ash cloud, and the last section is our conclusion. Theoretical basis Add noise information in the standard ICA model, making it the linear mixed ICA model with noise [9, 10], and the formula is shown below: ( ) ( ) ( ) x t As t t e = + (1) Where the x(t) refers to a mixed signal with M dimension, s(t) refers to a source signal (hidden variable) with L dimension, mixed matrix A is M × L, e(t) refers to Gaussian noise, which usually is a diagonal matrix with the inverse variance for Λ and the mean for zero. In linear mixed ICA model containing noises, the probability calculation formula for mixed signal x(t) is: 1 2 1 ( | , , ) | det( ) | exp[ ] 2 D p x s A E p Λ = Λ − (2) Where 1 ( ) ( ) 2 T D E x As x As = − Λ − , det(·) refers to the determinant value. In the model we assume the source signals are independent of each other, so the probability distribution formula for s can be expressed as 1 ( ) ( ) L