Deep Learning (DL)-based classification schemes for hyperspectral remotely sensed data have been introduced in the last few years with remarkable success due to their capability to learn the non-linear nature of the information that conforms hyperspectral images. In particular, Convolutional Neural Networks (CNNs) have been successfully used for solving problems requiring multi-class classification in the remote sensing field involving feature extraction. CNNs operate over small cubes of the dataset called patches centered in the pixels of the image instead of relying only on the spectral information corresponding to each pixel. These networks require a high number of observations to properly produce a generalized model. In these circumstances data augmentation techniques can help alleviate the problem by generating new, synthetic samples from existing data. Imputation is a statistical technique consisting in filling or replacing missing observations or values of a subset of observations by others obtained via inference from the original dataset. In this paper, a preliminary idea for a data augmentation technique based on the use of data imputation techniques for CNN classification is presented. Different hyperspectral images of the Earth surface widely used in the remote sensing field have been considered as test datasets. The results show the viability of the preliminary idea.
Image registration is an essential preprocessing task in many applications of hyperspectral images capturing the Earth surface. Maximally Stable Extremal Regions (MSER) is a feature–based method for image registration which extracts regions by thresholding the image at different grey levels. Its invariance to affine transformations makes it ideal for image registration. This method is usually employed in text detection and recognition as well as in the medical domain. Hyperspectral images contain spectral information that can be used for improving the image alignment. This article presents a first approach to a hyperspectral remote sensing image registration method based on MSER that efficiently exploits the information contained in the different spectral bands. The experimental results over four hyperspectral images show that the proposed method is promising as it achieves a higher number of correct registration cases than other feature–based methods.
In this paper the problem of studying the presence of different vegetation species and artificial structures in the riversides by using multispectral remote sensing information is studied. The information provided contributes to control the water resources in a region in northern Spain called Galicia. The problem is solved as a supervised classification computed over five-band multispectral images obtained by an Unmanned Aerial Vehicle (UAV). A classification scheme based on the extraction of spatial, spectral and textural features previous to a hierarchical classification by Support Vector Machine (SVM) is proposed. The scheme extracts the spatial-spectral information by means of a segmentation algorithm based on superpixels and by computing morphological operations over the bands of the image in order to generate an Extended Morphological Profile (EMP). The texture features extracted help in the classification of vegetation classes as the spatial-spectral features for these classes are not discriminant enough. The classification is computed over segments instead of pixels, thus reducing the computational cost. The experimental results over four real multispectral datasets from Galician riversides show that the proposed scheme improves over a standard classification method achieving very high accuracy results.
Change Detection (CD) techniques applied over multitemporal multispectral or hyperspectral remote sensing images allow monitoring changes in the land use or catastrophe tracking, among other applications. A multiclass CD technique for multidimensional images that is robust in the presence of noise is presented in this paper. The technique combines fusion at feature level to perform a first change/no change labeling (binary CD) and a later stage with fusion at decision level that performs a supervised multidate classification of the changed pixels (multiclass CD) obtaining the final from-to change map. The acquisition of multidimensional images usually corrupts the original signal by adding noise. This noise can be related with natural random processes or it can be produced during the sensor operation. Additive White Gaussian Noise (AWGN) and speckle noise simulate these effects. In this paper the robustness of the proposed CD technique in noisy scenarios for these two types of noise of varying intensity is evaluated. The experimental results show that the proposed technique is more robust than other alternatives, achieving accuracies close to those obtained in the absence of noise. The proposed technique is designed to be efficiently computed in GPU, thus dealing with the high computational cost of the processing of multidimensional images.
Supervised classification allows handling a wide range of remote sensing hyperspectral applications. Enhancing the spatial organization of the pixels over the image has proven to be beneficial for the interpretation of the image content, thus increasing the classification accuracy. Denoising in the spatial domain of the image has been shown as a technique that enhances the structures in the image. This paper proposes a multi-component denoising approach in order to increase the classification accuracy when a classification method is applied. It is computed on multicore CPUs and NVIDIA GPUs. The method combines feature extraction based on a 1Ddiscrete wavelet transform (DWT) applied in the spectral dimension followed by an Extended Morphological Profile (EMP) and a classifier (SVM or ELM). The multi-component noise reduction is applied to the EMP just before the classification. The denoising recursively applies a separable 2D DWT after which the number of wavelet coefficients is reduced by using a threshold. Finally, inverse 2D-DWT filters are applied to reconstruct the noise free original component. The computational cost of the classifiers as well as the cost of the whole classification chain is high but it is reduced achieving real-time behavior for some applications through their computation on NVIDIA multi-GPU platforms.
The classification of remote sensing hyperspectral images for land cover applications is a very intensive topic. In the case of supervised classification, Support Vector Machines (SVMs) play a dominant role. Recently, the Extreme Learning Machine algorithm (ELM) has been extensively used. The classification scheme previously published by the authors, and called WT-EMP, introduces spatial information in the classification process by means of an Extended Morphological Profile (EMP) that is created from features extracted by wavelets. In addition, the hyperspectral image is denoised in the 2-D spatial domain, also using wavelets and it is joined to the EMP via a stacked vector. In this paper, the scheme is improved achieving two goals. The first one is to reduce the classification time while preserving the accuracy of the classification by using ELM instead of SVM. The second one is to improve the accuracy results by performing not only a 2-D denoising for every spectral band, but also a previous additional 1-D spectral signature denoising applied to each pixel vector of the image. For each denoising the image is transformed by applying a 1-D or 2-D wavelet transform, and then a NeighShrink thresholding is applied. Improvements in terms of classification accuracy are obtained, especially for images with close regions in the classification reference map, because in these cases the accuracy of the classification in the edges between classes is more relevant.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.