Classification of clouds in satellite imagery using over-complete dictionary via sparse representation☆
Introduction
Meteorological satellite imagery not only plays a crucial role in weather forecast, but also have promising application in monitoring some natural disasters, such as typhoon, flood, snowstorms and forest fire [7], [14]. Cloud type recognition using satellite imagery is helpful to obtain information about atmosphere, ocean and cloud conditions in all-weather; moreover, it also has great significance for weather monitoring and disaster prevention and mitigation [3], [6]. However, visual/manual interpretation is one of the major cloud recognition methods at present, the results are unreliable and operator dependent (influenced by the physiological status, psychological status, cognition, thinking and judging experience, etc.). This condition prevents the development of scientific, automation and quantitative trend for weather forecasting [12].
There are two key problems in automatic recognition of cloud type in satellite cloud image, feature extraction and classifier design, and the research of these two aspects are active. Based on the BP neural network, Perez et al. [14] proposed a way to identify stratocumulus and determine the macro and micro physical properties (cloud thickness, effective particle size and cloud temperature). Georgiev and Kozinarova [6] applied the multi-scale method to analyze the weather system and combined the potential vorticity theory of dynamic meteorology with satellite water vapor imagery to sketch the development trend of strong convective clouds. In 2005, based on the fractal theory, Junjie et al. [8] extracted the texture feature of cloud imagery and conduct the recognition experiment of cloud and clear sky with the neural networks. Liu et al. [11] through analyzing the texture pattern of the high-level cirrus, put forward a new method to forecast the laws of the typhoon intensity. However, under the circumstances of high-dimensional features and small sample, the artificial neural network used to recognize cloud type can hardly get a good result, which shows at least one of the following defects: poor generalization ability, slow learning speed, low convergence ability [2], [17].
To solve the problem of cloud recognition fundamentally, it is essential to study from the lowest level of image processing (that is, the image representation). For feature extraction, its essence is applying a few sparse coefficients to describe the major of image features [18]. Furthermore, due to the complexity and multiplicity of spectral properties coming from various clouds and their underlying surface, a single pixel is usually the comprehensive reflection of different clouds and surface features, we can believe that it is a linear combination of several components of cloud system in picture, which exactly coincide with the idea of sparse representation, that treats the image as a linear combination of multiple atoms [13], [5]. Regarding to the difficulty of cloud image classification, this paper introduces sparse representation in two aspects: feature extraction and classification algorithm. Initially, the grayscale and spectral characteristics of the samples from different cloud type were used to construct an adaptive over-complete dictionary in order to represent the cloud samples sparsely and regard the sparse coefficients as the dictionary feature of samples. Then, the sparse representation coefficients matrix of training sample, resulted from representing sparsely the various training sample using the adaptive dictionary in turn, was served as projection subspace of different cloud patterns. Finally, via orthonormal process for the projection axis of the projection subspace, we design an effective sparse classifier and achieve the cloud type recognition according to the similarity between the test sample and specific subspace. By processing real satellite data, this paper shows a scheme to recognize cloud type based on the sparse representation theory.
Section snippets
Cloud classification system for satellite imagery
Different cloud types have different microphysical properties and they run different dynamic processes, so the research on cloud classification is essential in understanding the rule of atmospheric evolution to improve the service of meteorological, and it can guide people to intervene atmospheric operation such as the artificial rainfall. At present, the meteorological satellite has become an important means of cloud detection. The significant task of the satellite image analysis is extracting
The sparse representation of cloud spectral feature and the extraction of dictionary feature
Due to each pixel in cloud image is generally a comprehensive reflection of different cloud type and underlying surface, if the original spectral features are directly used in cloud classifying, it is hard to achieve pleasing classification accuracy, so a new method of feature extraction is needed urgently. In this case, the goal of cloud classification should be indentifying every pixel of cloud image into appropriate cloud type which playing the dominant role of the pixel. In recent years,
Sparse classifier based on reconstruction residual
Lei et al. [10] reported a very interesting work by using sparse representation for pattern recognition based on reconstructed residual as follows.
Assuming that training samples yi ∊ RM, i = 1, 2, … , N belong to n different types, respectively, and denoting mi as the number of the ith class training samples (number of total training samples is ). These sample vectors are arranged in turn of class to construct a sample matrix Ψ = [D1, D2, … , Dn], Ψ ∊ RM × N, and each column vector of the sub-matrix
Simulation results
This section presents the simulations conducted on real satellite data. The accuracy of CCSI-ODSR is evaluated; comparisons among CCSI-ODSR and some other traditional approaches are made. As mentioned in Section 2, six different classes representing background and clouds were defined. 10 daytime FY2D satellite data containing IR1, IR2, IR3, IR4, VIS channel imagery with all types of predefined classes are used in our simulations. Three meteorologists examine the selected cloud images and
Conclusions
This paper studies the sparse representation based solutions to cloud classification for satellite imagery and propose a new cloud classification scheme named as CCSI-ODSR in detail. The problem of the cloud classification in terms of similarity assessment between the test sample and specific cloud type subspace was formulated. Different cloud classifiers were benchmarked on their classification accuracy by data labeled by meteorologists, the results show that the proposed CCSI-ODSR is a better
Acknowledgments
This work is supported in part by the National Natural Science Foundation of China under Grants 61271399; Science and Technology Project of Ningbo under Grants 2011A610192, 2013D10011; K.C. Wong Magna Fund in Ningbo University.
References (19)
- et al.
Usefulness of satellite water vapour imagery in forecasting strong convection: a flash-flood case study
Atmos. Res.
(2009) - et al.
Nighttime cloud properties retrieval using MODIS and artificial neural networks
Adv. Space Res.
(2009) - et al.
Physical and statistical approaches for cloud identification using meteosat second generation-spinning enhanced visible and infrared imager data
Remote Sens. Environ.
(2008) - et al.
The K-SVD: an algorithm for designing of over complete dictionaries for sparse representation
IEEE Trans. Signal Process.
(2006) - et al.
An automatic cloud-masking system using backpro neural nets for AVHRR scenes
IEEE Trans. Geosci. Remote Sens.
(2003) - et al.
Cloud type classification algorithm for CloudSat satellite based on support vector machine
Trans. Atmos. Sci.
(2011) Compressive sensing
IEEE Trans. Inf. Theory
(2006)- et al.
Image denoising via sparse and redundant representations over learned dictionaries
IEEE Trans. Image Process.
(2006) - et al.
Automatic determination of center of tropical cyclone in satellite-generated IR images
IEEE Geosci. Remote Sens. Lett.
(2011)
Cited by (17)
A nonlinear dimensionality reduction framework using smooth geodesics
2019, Pattern RecognitionCitation Excerpt :Advanced data collection techniques in today’s world require researchers to work with large volumes of nonlinear data, such as global climate patterns [1,2], satellite signals [3,4], social and mobile networks [5,6], the human genome [7,8], and patterns in collective motion [9,10].
A novel fuzzy based deep neural network for rain fall prediction using cloud images
2023, Concurrency and Computation: Practice and ExperienceMLK-SVD, the new approach in deep dictionary learning
2021, Visual ComputerCloud classification: Principles and applications
2021, International Journal of Hydrology Science and Technology
- ☆
This paper has been recommended for acceptance by Jie Zou.