Automatic cloud masking of Earth observation images is one of the first required steps in optical remote sensing data processing since the operational use and product generation from satellite image time series might be hampered by undetected clouds. The high temporal revisit of current and forthcoming missions and the scarcity of labeled data force us to cast cloud screening as an unsupervised change detection problem in the temporal domain. We introduce a cloud screening method based on detecting abrupt changes along the time dimension. The main assumption is that image time series follow smooth variations over land (background) and abrupt changes will be mainly due to the presence of clouds. The method estimates the background surface changes using the information in the time series. In particular, we propose linear and nonlinear least squares regression algorithms that minimize both the prediction and the estimation error simultaneously. Then, significant differences in the image of interest with respect to the estimated background are identified as clouds. The use of kernel methods allows the generalization of the algorithm to account for higher-order (nonlinear) feature relations. After the proposed cloud masking and cloud removal, cloud-free time series at high spatial resolution can be used to obtain a better monitoring of land cover dynamics and to generate more elaborated products. The method is tested in a dataset with 5-day revisit time series from SPOT-4 at high resolution and with Landsat-8 time series. Experimental results show that the proposed method yields more accurate cloud masks when confronted with state-of-the-art approaches typically used in operational settings. In addition, the algorithm has been implemented in the Google Earth Engine platform, which allows us to access the full Landsat-8 catalog and work in a parallel distributed platform to extend its applicability to a global planetary scale.
An automatic cloud masking is one of the first required processing steps since the operational use of satellite image time series might be hampered by undetected clouds. The high temporal revisit of current and forthcoming missions allows us to consider cloud screening as an unsupervised change detection problem in the temporal domain. Therefore, we propose a cloud screening method based on detecting abrupt changes in the temporal domain. The main assumption is that image time series follow smooth variations over land (background) and abrupt changes in certain spectral and spatial features will be mainly due to the presence of clouds. The method estimates the background and common surface changes using the full information in the time series. In particular, we propose linear and nonlinear least squares regression algorithms that minimize both the prediction and estimation error simultaneously. Then, significant differences in the image of interest with respect to the estimated background are identified as clouds. The use of kernel methods allows the generalization of the algorithm to account for higher-order (nonlinear) feature relations. After cloud detection, cloud-free time series at high spatial resolution can be used to obtain a better monitoring of the land cover dynamics and to generate more elaborated products. The proposed method is tested in a dataset with 5-day revisit time series from
SPOT-4 at high resolution and Landsat-8 time series.
Support Vector Machines (SVM) have been widely adopted by the remote sensing community in the last decade.
The standard algorithm has been mainly applied to image classication tasks. Many advanced developments
based on SVM have been introduced as well. This paper, nevertheless, revises the standard formulation of SVM.
An important part of the paper is about the intuition on the SVM parts: the cost, the regularizer and the free
parameters. Finally, the paper revises three interesting simple modications well suited to tackle remote sensing
image classication: constraining the margin, including invariances and the information of unlabeled samples.
Some examples are given to illustrate these concepts.
This paper presents a novel unsupervised clustering scheme to find changes in two or more coregistered remote
sensing images acquired at different times. This method is able to find nonlinear boundaries to the change
detection problem by exploiting a kernel-based clustering algorithm. The kernel k-means algorithm is used in
order to cluster the two groups of pixels belonging to the 'change' and 'no change' classes (binary mapping). In
this paper, we provide an effective way to solve the two main challenges of such approaches: i) the initialization
of the clustering scheme and ii) a way to estimate the kernel function hyperparameter(s) without an explicit
training set. The former is solved by initializing the algorithm on the basis of the Spectral Change Vector (SCV)
magnitude and the latter is optimized by minimizing a cost function inspired by the geometrical properties of
the clustering algorithm. Experiments on VHR optimal imagery prove the consistency of the proposed approach.
This paper evaluates the potential use of nonlinear retrieval methods to derive cloud, surface and atmospheric
properties from hyperspectral MetOp-IASI and MTG-IRS spectra. The methods are compared in terms of both
accuracy and speed with the current IASI and IRS L2 PPFP implementation, which consists of a principal component
extraction, typically referred as to Empirical Orthogonal Functions (EOF), and a subsequent canonical
linear regression. This research proposes the evaluation of some other methodological advances considering 1)
other linear feature extraction methods instead of EOF, such as (orthonormalized) partial least squares, and
2) the linear combination of nonlinear regression models in the form of committee of experts. The nonlinear
regression models considered in this work are artificial neural networks (NN) and kernel ridge regression (KRR)
as nonparametric multioutput powerful regression tools. Results show that, in general, nonlinear models outperform
the linear retrieval both in the presence of noise and noise-free settings, and for both IASI and IRS
synthetic and real data. The combination of models makes the retrieval more robust, improves the accuracy,
and decreases the estimated bias. These results confirm the validity of the proposed approach for retrieval of
Many remote sensing data processing problems are inherently constituted by several tasks that can be solved
either individually or jointly. For instance, each image in a multitemporal classification setting could be taken
as an individual task but relation to previous acquisitions should be properly considered. In such problems,
different modalities of the data (temporal, spatial, angular) gives rise to changes between the training and
test distributions, which constitutes a difficult learning problem known as covariate shift. Multitask learning
methods aim at jointly solving a set of prediction problems in an efficient way by sharing information across
tasks. This paper presents a novel kernel method for multitask learning in remote sensing data classification. The
proposed method alleviates the dataset shift problem by imposing cross-information in the classifiers through
matrix regularization. We consider the support vector machine (SVM) as core learner and two regularization
schemes are introduced: 1) the Euclidean distance of the predictors in the Hilbert space; and 2) the inclusion
of relational operators between tasks. Experiments are conducted in the challenging remote sensing problems of
cloud screening from multispectral MERIS images and for landmine detection.
The most successful one-class classification methods are discriminative approaches aimed at separating the class of
interest from the outliers in a proper feature space. For instance, the support vector domain description (SVDD)
has been successfully introduced for solving one-class remote sensing classification problems when scarce and
uncertain labeled data is available. The success of this kernel method is due to that maximum margin nonlinear
separation boundaries are implicitly defined, thus avoiding the hard and ill-conditioned problem of estimating
probability density functions (PDFs). Certainly, PDF estimation is not an easy task, particularly in the case of
high-dimensional PDFs such as is the case of remote sensing data. In high-dimensional PDF estimation, linear
models assumed by widely used transforms are often quite restrictive to describe the PDF. As a result, additional
non-linear processing is typically needed to overcome the limitations of the models. In this work we focus on
the multivariate Gaussianization method for PDF estimation. The method is based on the Projection Pursuit
Density Estimation (PPDE) technique.1 The original PPDE procedure consists in iteratively project the data
in the most non-Gaussian directions (like in ICA algorithms) and Gaussianizing them marginally. However,
the extremely high computational cost associated to multiple ICA evaluations has prevented its practical use in
high-dimensional problems such as those encountered in image processing. Here, we propose a fast alternative
to iterative Gaussianization that makes it suitable for remote sensing applications while ensuring its theoretical
convergence. Method's performance is successfully illustrated in the challenging problem of urban monitoring.
This paper presents a semi-supervised one-class support vector machine classifier for remote sensing applications.
In <i>one-class</i> image classification, one tries to detect pixels belonging to one class and reject the others. When
few labeled target pixels and no labeled outlier pixels are available, the selection of the support vector machine
free parameters is very challenging. This problem can be alleviated by introducing the information of the wealth
of unlabeled samples present in the scene. The proposed algorithm deforms the training kernel by modelling
the data marginal distribution with the graph Laplacian built with labeled and unlabeled samples. The good
performance of the proposed method is illustrated in challenging remote sensing image classification scenarios
where information of only one class of interest is available. In particular, we present results in multispectral
cloud screening, hyperspectral crop detection, and multisource urban monitoring. Experimental results show the
suitability of the proposal, specially in cases with few or poorly representative labeled samples.
The Orthogonal Subspace Projection (OSP) algorithm is substantially a kind of matched filter that requires the
evaluation of a prototype for each class to be detected. The kernel OSP (KOSP) has recently demonstrated
improved results for target detection in hyperspectral images. The use of kernel methods helps to combat the
high dimensionality problem and makes the method robust to noise. This paper incorporates the contextual
information to KOSP with a family of composite kernels of tunable complexity. The good performance of the
proposed methods is illustrated in hyperspectral image target detection problems. The information contained in
the kernel and the induced kernel mappings is analyzed, and bounds on generalization performance are given.
Identification of land cover types is one of the most critical activities in remote sensing. Nowadays, managing
land resources by using remote sensing techniques is becoming a common procedure to speed up the process while
reducing costs. However, data analysis procedures should satisfy the accuracy figures demanded by institutions
and governments for further administrative actions.
This paper presents a methodological scheme to update the citrus Geographical Information Systems (GIS)
of the Comunidad Valenciana autonomous region (Spain). The proposed approach introduces a multi-stage
automatic scheme to reduce visual photointerpretation and ground validation tasks. First, an object-oriented
feature extraction process is carried out for each cadastral parcel from very high spatial resolution (VHR) images
(0.5m) acquired in the visible and near infrared. Next, several automatic classifiers (decision trees, multilayer
perceptron, and support vector machines) are trained and combined to improve the final accuracy of the results.
The proposed strategy fulfills the high accuracy demanded by policy makers by means of combining automatic
classification methods with visual photointerpretation available resources. A level of confidence based on the
agreement between classifiers allows us an effective management by fixing the quantity of parcels to be reviewed.
The proposed methodology can be applied to similar problems and applications.
This paper formulates the problem of distinguishing changed from unchanged pixels in remote sensing images as a
minimum enclosing ball (MEB) problem with changed pixels as target class. The definition of the sphere shaped
decision boundary with minimal volume that embraces changed pixels is approached in the context the support vector
formalism adopting a support vector domain description (SVDD) one-class classifier. The SVDD maps the data into a
high dimensional feature space where the spherical support of the high dimensional distribution of changed pixels is
computed. The proposed formulation of the SVDD uses both target and outlier samples for defining the MEB, and is
included here in an unsupervised system for change detection. For this purpose, nearly certain examples for the classes of
both targets (i.e., changed pixels) and outliers (i.e., unchanged pixels) for training are identified based on thresholding
the magnitude of spectral change vectors. Experimental results obtained on two different multitemporal and multispectral
remote sensing images pointed out the effectiveness of the proposed method.
In this paper, we focus on different kinds of regularization for Linear Discriminant Analysis (LDA) in the
context of ill-posed remote sensing image classification problems. Several LDA-based classifiers are studied
theoretically and tested on various remote sensing datasets. In addition, we introduce an efficient version of
the standard regularized LDA recently presented in Ref. 1 to cope with high-dimensional small sample size
(ill-posed) problems. Experimental results demonstrate the suitability of the proposal.
We explicitly formulate a family of kernel-based methods for (supervised and partially supervised) multitemporal classification and change detection. The novel composite kernels developed account for the static and temporal cross-information between pixels of subsequent images simultaneously. The methodology also takes into account spectral, spatial, and temporal information, and contains the familiar difference and ratioing methods in the kernel space as a particular cases. The methodology also permits straightforward fusion of multisource information. Several scenarios are considered in which partial or complete labeled information at the prediction time is available. The developed methods are then tested under different classification frameworks: (1) inductive support vector machines (SVM), and (2) one-class support vector data description (SVDD) classifier, in which only samples of a class of interest are used for training. The proposed methods are tested in a challenging real problem for urban monitoring. The composite kernel approach is additionally used as a fusion methodology to combine synthetic aperture radar (SAR) and multispectral data, and to integrate the spatial and textural information at different scales and orientations through Gabor filters. Good results are observed in almost all scenarios; the SVDD classifier demonstrates robust multitemporal classification and adaptation capabilities when few labeled information is available, and SVMs show improved performance in the change detection approach.
In addition to typical random noise, remote sensing hyperspectral images are generally affected by non-periodic partially deterministic disturbance patterns due to the image formation process and characterized by a high degree of spatial and spectral coherence. This paper presents a new technique that faces the problem of removing the spatial coherent noise known as vertical stripping (VS) usually found in images acquired by push-broom sensors, in particular for the Compact High Resolution Imaging Spectrometer (CHRIS). The correction is based on the hypothesis that the vertical disturbance presents higher spatial frequencies than the surface radiance. The proposed method introduces a way to exclude the contribution of the spatial high frequencies of the surface from the destripping process that is based on the information contained in the spectral domain. Performance of the proposed algorithm is tested on sites of different nature, several acquisition modes (different spatial and spectral resolutions) and covering the full range of possible sensor temperatures. In addition, synthetic realistic scenes have been created, adding modeled noise for validation purposes. Results show an excellent rejection of the noise pattern with respect to the original CHRIS images. The analysis shows that high frequency VS is successfully removed, although some low frequency components remain. In addition, the dependency of the noise patterns with the sensor temperature has been found to agree with the theoretical one, which confirms the robustness of the presented approach. The approach has proven to be robust, stable in VS removal, and a tool for noise modeling. The general nature of the procedure allows it to be applied for destripping images from other spectral sensors.
Accurate and automatic detection of clouds in satellite scenes is a key issue for a wide range of remote sensing applications. With no accurate cloud masking, undetected clouds are one of the most significant source of error in both sea and land cover biophysical parameter retrieval. Sensors with spectral channels beyond 1 um have demonstrated good capabilities to perform cloud masking. This spectral range can not be exploited by recently developed hyperspectral sensors that work in the spectral range between 400- 1000 nm. However, one can take advantage of their high number of channels and spectral resolution to increase the cloud detection accuracy, and to describe properly the detected clouds (cloud type, height, subpixel coverage, could shadows, etc.) In this paper, we present a methodology for cloud detection that could be used by sensors working in the VNIR range. First, physically-inspired features are extracted (TOA reflectance and their spectral derivatives, atmospheric oxygen and water vapour absorptions, etc). Second, growing maps are built from cloud-like pixels to select regions which potentially could contain clouds. Then, an unsupervised clustering algorithm is applied in these regions using all extracted features. The obtained clusters are labeled into geo-physical classes taking into account the spectral signature of the cluster centers. Finally, an spectral unmixing algorithm is applied to the segmented image in order to obtain an abundance map of the cloud content in the cloud pixels. As a direct consequence of the detection scheme, the proposed system is capable to yield probabilistic outputs on cloud detected pixels in the image, rather than flags. Performance of the proposed algorithm is tested on six CHRIS/Proba Mode 1 images, which presents a spatial resolution of 32 m, 62 spectral bands with 6-20 nm bandwidth, and multiangularity.
In this communication, we evaluate the performance of the relevance vector machine (RVM) (Tipping,2000) for the estimation of biophysical parameters from remote sensing images. For illustration purposes, we focus on the estimation of chlorophyll concentrations from multispectral imagery, whose measurements are subject to high levels of uncertainty, both regarding the difficulties in ground-truth data acquisition, and when comparing <i>in situ</i> measurements against satellite-derived data. Moreover, acquired data are commonly affected by noise in the acquisition phase, and time mismatch between the acquired image and the recorded measurements, which is critical for instance for coastal water monitoring. In this context, robust and stable regressors that provide inverse models are desirable. Lately, the use of the support vector regressor (SVR) has produced good results to this end. However, the SVR has many deficiencies, which could be theoretically alleviated by the RVM. In this paper, performance of the RVM is benchmarked with SVR in terms of accuracy and bias of the estimations, sparseness of the solutions, distribution of the residuals, robustness to low number of training samples, and computational burden. In addition, some theoretical issues are discussed, such as the sensitivity to hyperparameters setting, kernel selection, and confidence intervals on the predictions. Results suggest that RVM offer an excellent compromise between accuracy and sparsity of the solution, and reveal itself as less sensitive to selection of the free parameters. Some disadvantages are also pointed, such as the unintuitive confidence intervals provided and the computational cost.
In this paper, we analyze regularized non-linear methods in the context of hyperspectral image classification. For this purpose, we compare regularized radial basis function neural networks (Reg-RBFNN), standard support vector machines (SVM), and kernel Fisher discriminant (KFD) analysis both theoretically and experimentally. We focus on the accuracy of methods when working in noisy environments, high input dimension, and limited number of training samples. In addition, some other important issues are discussed, such as the sparsity of the solutions, the computational burden, and the capability of the methods to provide probabilistic outputs. Although in general all methods yielded satisfactory results, SVM revealed more effective than KFD and Reg-RBFNN in standard situations regarding accuracy, robustness, sparsity, and computational cost.
In some key operational domains, users are not specially interested in obtaining an exhaustive map with all the thematic classes present in an area of interest, but rather in identifying accurately a single class of interest. In this paper, we present a novel partially supervised classification technique that faces this interesting practical and methodological problem. We have adopted a two-stage classification scheme based on an unsupervised approach, which allows us to introduce supervised information about the class of interest without an additional sample labeling. The first stage of the process consists in an initial clustering of the image using the Self-Organizing Map algorithm. The second stage consists in a partially supervised hierarchical joint of clusters. We modify the employed criterion of similarity by introducing fuzzy membership functions that make use of the supervised information. The method is tested on urban monitoring, where the objective is to produce an automatic classification of 'Urban/Non-Urban' by using optical and radar data (Landsat TM and 35-days interferometric pairs of ERS2 SAR). We compare classification accuracy of the proposed method to its parametric version, which uses the Expectation-Maximization algorithm. The good performance confirms the validity of the proposed approach: 90% classification accuracy using supervised information only in the coherence map.
In this paper, we propose a kernel-based approach for hyperspectral knowledge discovery, which is defined as a process that involves three steps: pre-processing, modeling and analysis of the classifier. Firstly, we select the most representative bands analyzing the surrogate and main splits of a Classification And Regression Trees (CART) approach. This yields three datasets with different reduced input dimensionality (6, 3 and 2 bands, respectively) along with the original one (128 bands). Secondly, we develop several crop cover classifiers for each of them. We use Support Vector Machines (SVM) and analyze its performance in terms
of efficiency and robustness, as compared to multilayer perceptrons (MLP) and radial basis functions (RBF) neural networks. Suitability to real-time working conditions, whenever a preprocessing stage is not possible, is evaluated by considering models with and without the CART-based feature selection stage. Finally, we analyze the support vectors distribution in the input space and through Principal Component Analysis (PCA) in order to gain knowledge about the problem. Several conclusions are drawn: (1) SVM yield better
outcomes than neural networks; (2) training neural models is unfeasible when working with high dimensional spaces; (3) SVM perform similarly in the four classification scenarios, which indicates that noisy bands are successfully detected and (4) relevant bands for the classification are identified.
In this paper, we propose a new approach to the classification of hyperspectral images. The main problem with supervised methods is that the learning process heavily depends on the quality of the training data set. In remote sensing, the training set is useful only for simultaneous images or for images with the same classes taken under the same conditions; and, even worse, the training set is frequently not available. On the other hand, unsupervised methods are not sensitive to the number of labelled samples since they work on the whole image. Nevertheless, relationship between clusters and classes is not ensured. In this context, we propose a combined strategy of supervised and unsupervised learning methods that avoids these drawbacks and automates the classification process. The method is based on the general formulation of the expectation-maximization (EM) algorithm. This method is applied to crop cover recognition of six hyperspectral images from the same area acquired with HyMap spectrometer during the DAISEX99 campaign. For classification purposes, six different classes are considered in this area: corn, wheat, sugar beet, barley, alfalfa, and soil. Classification accuracy results are compared to common methods: ISODATA, Learning Vector Quantization, Gaussian Maximum Likelihood, Expectation-Maximization, and Neural Networks. The good performance confirms the validity of the proposed approach in terms of accuracy and robustness.