Image Fusion in Hyperspectral Image Classification using Genetic Algorithm

Due to the high computational complexity of traditional convolutional neural networks, the execution time is long and the computational cost is too high. In this paper, we propose a deep separable convolutional neural network with attention mechanism added to improve the classification accuracy and generalization ability of hyperspectral images. The network uses separable convolution combined with residual connections to construct residual units with fewer parameters and adds an attention mechanism layer at the end of the network, which helps to improve the overall performance of the model. So this model has stronger generalization ability now, shorter computation time, and stronger network performance. Finally, the overall accuracy of the model in this paper is 98.48%, 99.1% and 97.40% on the Salinas dataset and the more newly proposed Wuhan Longkou and Wuhan Hanchuan datasets, respectively. It proves that the model has better generalization ability and can complete the calculation in a shorter time. Improving the classification accuracy of hyperspectral images like the Wuhan Longkou dataset is important for agricultural development.KeywordsResidual NetworkSeparable convolutionConvolutional neural networkAttention mechanismAgricultural hyperspectrum

In this paper discrete wavelet transform (DWT) and empirical mode decomposition (EMD) are employed as a preprocessing stage in a multiclassifier and decision fusion system. The proposed method consists of three steps. In the first step, 2D-EMD is performed on each hyperspectral image band in order to obtain useful spatial information. Then, useful spectral information is obtained by applying the 1D-DWT to each signature of 2D-EMD performed bands. A novel feature set is generated using both spectral and spatial information. In the second step, each feature is independently classified by support vector machines (SVM), creating a multiclassifier system. In the last step, classification results are fused using a decision fusion criterion to produce one final classification. The proposed method improves overall classification accuracy over independent classifiers when reduced number of features are employed.

Envelope modified versions of the empirical mode decomposition (EMD) method such as the B‐spline interpolation‐based EMD (B‐EMD) method and cardinal spline interpolation‐based EMD (C‐EMD) method have been proposed recently for purpose of improving its effectiveness. To shed further light on their performance, the behaviours of these EMD‐type methods in the presence of white Gaussian noises are investigated in this study based on extensive numerical experiments. Similarly to the EMD method, it turns out that the envelope modified EMD methods also act as filter banks essentially. However, the spectra among the first several intrinsic mode functions of the B‐EMD method have fewer overlaps than those of the EMD and C‐EMD methods, which indicate that the B‐EMD method has a better ability to alleviate the mode mixing problem for signals with higher frequencies. On the other hand, the C‐EMD method is shown to perform better than the EMD and B‐EMD methods on separating tones with lower frequencies.

In this paper, a novel method based on wavelet transformation of functional data for accurate classification of hyperspectral images is proposed. The motivation of the proposed method is that in the hyperspectral images, the spectral curve of each pixel can be viewed as a function, mathematically. And it can make the best of the abundant spectral information of the original hyperspectral images. Based on this, we have proposed an effective hyperspectral image classification method which mainly used functional principal component analysis (FPCA) before. However, FPCA method focuses on the global features analysis of hyperspectral image data and may not well extract local structural features. In this paper, in order to better reflect the local structural features of hyperspectral images, we make use of the time-frequency localization characteristics of wavelet transformation to decompose hyperspectral images. Then, the sparse wavelet decomposition coefficients are classified by using extreme learning machine (ELM). Experimental results on two popular hyperspectral images show that our method outperforms some state-of-the-art hyperspectral image classification methods.

The signature of hyperspectral image (HSI) pixels and their decomposition into empirical modes (EM) and low-frequency residuals are investigated. On the basis of estimates related to the EM-decomposition method, the possibility of switching from a 2-byte representation of the values of the HIS-signature to a 1-byte one is examined using the example of the Moffett Field from the AVIRIS spectrometer. It is revealed that the localization of the minimum window sizes for the first EM is correlated with the localization of the significant influence of the atmosphere; the first low-frequency residues have a fairly high correlation coefficient with the signature and the first 2 of them and their EM are most interesting for use; 50 of the 224 HIS-channels are noisy and can be excluded from consideration; EM with practically no loss of accuracy can be reduced to a 1-byte representation. The management of the classification capabilities of signatures by changing the threshold value of the correlation coefficient with the sample, as well as the application of the 1st and 2nd low-frequency residues in place of the signature, was studied. Classification capabilities of signatures in a 1-byte representation are almost equivalent to a 2-byte one, which makes it possible to put a signature with 1-byte representation as the object of compression. For the wavelet decomposition of the HSI data array, in combination with a 1-byte representation, a nearlossless compression ratio of 6.65 is obtained.

Abstract. Hyperspectral image enhancement has been a concern for the remote sensing society for detailed end member detection. Hyperspectral remote sensor collects images in hundreds of narrow, continuous spectral channels, whereas multispectral remote sensor collects images in relatively broader wavelength bands. However, the spatial resolution of the hyperspectral sensor image is comparatively lower than that of the multispectral. As a result, spectral signatures from different end members originate within a pixel, known as mixed pixels. This paper presents an approach for obtaining an image which has the spatial resolution of the multispectral image and spectral resolution of the hyperspectral image, by fusion of hyperspectral and multispectral image. The proposed methodology also addresses the band remapping problem, which arises due to different regions of spectral coverage by multispectral and hyperspectral images. Therefore we apply algorithms to restore the spatial information of the hyperspectral image by fusing hyperspectral bands with only those bands which come under each multispectral band range. The proposed methodology is applied over Henry Island, of the Sunderban eco-geographic province. The data is collected by the Hyperion hyperspectral sensor and LISS IV multispectral sensor.

Hybrid pixel based method for multimodal image fusion based on Integration of Pulse Coupled Neural Network (PCNN) and Genetic Algorithm (GA) using Empirical Mode Decomposition (EMD)

Hyperspectral image classification using multi-feature fusion

Power grid as an important infrastructure which ensures the healthy development of economy and society and accurate and reasonable prediction of the power grid investment demand has always been the focus problem of the power planning department and the power grid enterprises. In view of the complex nonlinear and nonstationary characteristics of the power grid investment demand sequence, a novel hybrid EMD-GASVM-RBFNN forecasting model based on empirical mode decomposition (EMD) method, support vector machines optimized by genetic algorithm (GA-SVM) model, and radial basis function neural network (RBFNN) model is proposed. Firstly, the EMD method is used to decompose the original power grid investment data sequence into a series of IMF components and a residual component which have stronger regularity compared with the original data. Then, according to the different characteristics of each subsequence, the GA-SVM and RBFNN model will be used to forecast different subsequences, respectively. Next, the prediction results of different subsequences are aggregated to obtain the final prediction results of the power grid investment. Finally, this paper dynamically simulates China’s power grid investment from 2018 to 2020 based on the EMD-GASVM-RBFNN hybrid forecasting model and Monte Carlo method.

Hyperspectral image classification based on combining texture and multiscale guided filter feature of principal component

The key to the fault diagnosis of rotating machinery is to extract the fault characteristics from the vibration signal of rotating machinery. Signal analysis and processing are the most commonly used methods for feature extraction. In this paper, a method combining the empirical mode decomposition (EMD) and spectral kurtosis is presented for fault detection of rotating machinery based on vibrational signal. According to the modulation characteristic of the rolling element bearing fault signal and the disadvantage of depending on the experience to select resonance high frequency band, an improved method integrating EMD with spectrum kurtosis for rolling bearing fault diagnosis is put forward. Firstly, the bearing fault signal is decomposed into a series of intrinsic mode function (IMF) through EMD method. Secondly, the false IMF components are eliminated through mutual information, kurtosis and cross-correlation, then the useful IMFs are selected to construct the fault signal. Finally, by designing the optimal band pass filter with the spectral kurtosis, it can be obtained that, the envelope demodulation spectrum of the filtered signal, and the fault feature of rolling bearing are extracted. The experimental result of the rolling bearing fault diagnosis shows that the proposed method can effectively extract the weak fault feature of rolling bearing, which is superior to envelope analysis.

A comparative study between Empirical Wavelet Transforms and Empirical Mode Decomposition Methods: Application to bearing defect diagnosis

Empirical Mode Decomposition (EMD) and Wavelet method have been widely used in the vibration signal analysis, mainly because of the powerful ability of blind modal separation. In the flutter test, modal identification and modal trajectories tracking online can help to ensure the test safety. However, traditional EMD method is only suitable for offline data batching, which limits its applications to the real-time or near real-time signal analysis. For this reason, according to the online processing requirements of flutter test under the turbulent excitation conditions, combining with wavelet spectrum analysis, an online EMD method based on a sliding window for structural modal identification and flutter boundary prediction (FBP) is proposed in this paper. The corresponding mathematical principles and implementation algorithms are given as well, which decompose the measured signal into the Intrinsic Mode Function (IMF) firstly, and then estimate the modal parameters by using the wavelet power spectrum method. The FBP result is typically obtained online during the test process by using curve fitting and extrapolation of the modal damping. The numerical simulations and test data from a low-speed wind tunnel flutter test are used to verify the effectiveness and engineering feasibility of the proposed method.

Object-based classification methods can improve the accuracy of hyperspectral image classification due to the fact that they incorporate spatial information into the classification procedure by assigning neighboring pixels into the same class. In this paper, a new object-based feature extraction method is proposed which makes use of information theory to reduce the Bayes error. In this way, the proposed method exploits higher order statistics for feature extraction which are very effective for non Gaussian data such as hyperspectral images. The criterion to be minimized is composed of three mutual information terms. The first and second terms, consider the maximal relevance and minimal redundancy, respectively, while the third term takes into account the segmentation map containing disjoint spatial regions. To obtain the segmentation map, we apply the firefly clustering algorithm whose fitness function simultaneously considers the intra-distance between samples and their cluster centroids, and inter-distance between centroids of any two clusters. Our experimental results, performed using a variety of hyperspectral scenes, indicate that the proposed framework gives better classification results than some state-of the-art spectral–spatial feature extraction methods.

Hyperspectral remote sensing image classification based on residual generative Adversarial Neural Networks