Unsupervised SAR Research Articles

Optical knowledge assisted unsupervised cross-domain SAR target detection

Optical knowledge assisted unsupervised cross-domain SAR target detection

Detection of Forest Fires through Deep Unsupervised Learning Modeling of Sentinel-1 Time Series

With an increase in the amount of natural disasters, the combined use of cloud-penetrating Synthetic Aperture Radar and deep learning becomes unavoidable for their monitoring. This article proposes a methodology for forest fire detection using unsupervised location-expert autoencoders and Sentinel-1 SAR time series. The models are trained on SAR multitemporal images over a specific area using a reference period and extract any deviating time series over that same area for the test period. We present three variations of the autoencoder, incorporating either temporal features or spatiotemporal features, and we compare it against a state-of-the-art supervised autoencoder. Despite their limitations, we show that unsupervised approaches are on par with supervised techniques, performance-wise. A specific architecture, the fully temporal autoencoder, stands out as the best-performing unsupervised approach by leveraging temporal information of Sentinel-1 time series using one-dimensional convolutional layers. The approach is generic and can be applied to many applications, though we focus here on forest fire detection in Canadian boreal forests as a successful use case.

An Unsupervised CNN-Based Multichannel Interferometric Phase Denoising Method Applied to TomoSAR Imaging

Tomographic synthetic aperture radar (TomoSAR) is an advanced SAR interferometric technique to retrieve 3D spatial information. However, the standard deviation in the reconstructed elevation could be high due to the noise in the interferometric phases, which makes the denoising filter crucial before tomographic reconstruction. In this paper, we propose an unsupervised multichannel SAR interferometric phase denoising method based on the Convolution Neural Network (CNN). It utilizes the Weighted Least-Squares (WLS) regularization combining with the covariance of multichannel interferometric phases to minimize the standard deviation of phase noise, which leads to the accurate and complete TomoSAR reconstruction. This network is trained by real SAR images and the results of both simulated and real observations verify the effectiveness of our proposed method.

Unsupervised SAR Despeckling by Combining Online Speckle Generation and Unpaired Training

Unsupervised SAR Despeckling by Combining Online Speckle Generation and Unpaired Training

Unsupervised polarimetric SAR urban area classification based on model-based decomposition with cross scattering

Since it has been validated that cross-polarized scattering (HV) is caused not only by vegetation but also by rotated dihedrals, in this study, we use rotated dihedral corner reflectors to form a cross scattering matrix and propose an extended four-component model-based decomposition method for PolSAR data over urban areas. Unlike other urban area decomposition techniques which need to discriminate the urban and natural areas before decomposition, this proposed method is applied on PolSAR image directly. The building orientation angle is considered in this scattering matrix, making it flexible and adaptive in the decomposition. Therefore, we can separate cross scattering of urban areas from the overall HV component. Further, the cross and helix scattering components are also compared. Then, using these decomposed scattering powers, the buildings and natural areas can be easily discriminated from each other using a simple unsupervised K-means classifier. Moreover, buildings aligned and not aligned along the radar flight direction can be also distinguished clearly. Spaceborne RADARSAT-2 and airborne AIRSAR full polarimetric SAR data are used to validate the performance of our proposed method. The cross scattering power of oriented buildings is generated, leading to a better decomposition result for urban areas with respect to other state-of-the-art urban decomposition techniques. The decomposed scattering powers significantly improve the classification accuracy for urban areas.

Hierarchical semantic model and scattering mechanism based PolSAR image classification

The hybrid terrain type, such as the urban area, is generated by the backscattering of the similar ground objects aggregated together. It is a challenge to classify the hybrid terrain types as semantic homogeneous regions for unsupervised polarimetric SAR (PolSAR) image classification, since there are sharp bright-dark variations in intensity. In this paper, a hierarchical semantic model (HSM) is proposed for PolSAR images, which overcomes the above shortcoming by constructing the primal-level and middle-level semantics. The primal-level semantic is a polarimetric sketch map consisting of sketch segments, which is the sparse representation of a PolSAR image. The middle-level semantic is a region map which can extract hybrid terrain types as semantic homogeneous regions from the sketch map. Mapped by the region map, a complex PolSAR scene is partitioned into hybrid, linear and homogeneous subspaces. Then, according to the characteristics of the three subspaces, different methods are adopted, and furthermore polarimetric information is involved to improve the classification result. Experimental results on PolSAR data sets with different bands and sensors demonstrate that the proposed method is superior to the state-of-the-art methods in region homogeneity especially for hybrid terrain types.

Unsupervised SAR Image Segmentation Using Higher Order Neighborhood-Based Triplet Markov Fields Model

The triplet Markov fields (TMF) model has been successfully applied to statistical segmentation of nonstationary images by introducing the auxiliary field, which represents the different stationarities of images. Commonly, the TMF adopts a four-nearest neighborhood. This limits the modeling ability for complex priors. Therefore, this paper suggests using a higher order neighborhood-based TMF (HN-TMF). In the HN-TMF, the autocovariance analysis is applied to reveal the local fluctuation at each site. The auxiliary field is then redefined based on the local fluctuation information to denote homogeneity or heterogeneity. Based on the auxiliary field, the local energy function in HN-TMF is constructed either in a homogeneous or heterogeneous way, and hence, the local structure can be embedded in the energy function to improve the prior modeling ability. Along with the newly constructed energy function, new initializations of HN-TMF parameters are given to fulfill the physical interpretation of the energy function. The experiments performed on both synthetic and real synthetic aperture radar images demonstrate the effectiveness of the proposed HN-TMF in both speckle noise reduction and heterogeneous region segmentation accuracy.

Characterizing Scattering Behaviour and Assessing Potential for Classification of Arctic Shore and Near-Shore Land Covers with Fine Quad-Pol RADARSAT-2 Data

Freeman–Durden and Cloude–Pottier decompositions were applied to polarimetric RADARSAT-2 data to characterize the scattering behaviour of shore and near-shore land cover types for two study areas in the Beaufort Sea, Canada: Richards Island and Tuktoyaktuk Harbour. The impact of incidence angle was evaluated through comparison of Single Look Complex Fine Quad-Pol images acquired at three angles: shallow (45.3°–49.5°), medium (39.3°–41.6°), and steep (20.9°–24.2°), and the potential for accurate classification was assessed using three unsupervised polarimetric SAR classifiers: Wishart-entropy/alpha, Wishart-entropy/anisotropy/alpha, and Freeman-Wishart. Results indicate that overall, medium, and shallow incidence angle images provide more useful information than steep angle images and that these classifiers are capable of distinguishing some general land cover types, including substrates, wetlands, vegetated tundra, and water. Results from this analysis will inform future research related to classification and interpretation of RADARSAT-2 data for shoreline mapping.Résumé. Des décompositions Freeman–Durden et Cloude–Pottier ont été appliquées à des données polarimétriques RADARSAT-2 pour caractériser le comportement de diffusion des différents types de couverture terrestre des milieux littoraux et côtiers pour deux zones d'étude dans la mer de Beaufort, au Canada: l'île Richards et le port de Tuktoyaktuk. L'impact de l'angle d'incidence a été évalué par comparaison des images singulières complexes <<Single Look Complex (SLC)>> quad-pol fines acquises à trois angles; plus rasants (45.3°–49.5°), moyens (39.3°–41.6°) et abrupts (20.9°–24.2°), tandis que le potentiel pour une classification précise a été évalué à l'aide de trois classificateurs RSO polarimétriques non dirigés: Wishart-entropie/alpha, Wishart-entropie/anisotropie/alpha et Freeman–Wishart. Les résultats indiquent que, dans l'ensemble, les images avec des angles d'incidence moyens à rasants fournissent des informations plus utiles que les images aux angles abrupts, et que ces classificateurs sont capables de distinguer certains types généraux de couverture terrestre, y compris: les substrats, les zones humides, la toundra végétalisée et l'eau. Les résultats de cette analyse seront utiles dans les recherches futures liées à la classification et à l'interprétation des données RADARSAT-2 pour la cartographie des rivages.

Change Detection in SAR Images Based on Semi-Supervised Learning

Support Vector Machine (SVM) is a supervised approach, which needs large numbers of labeled samples. However, it is difficult to obtain such samples for change detection based on SAR images and the available labeled samples are very limited. this paper proposes a semi-supervised support vector machine (S3VM) unsupervised SAR image change detection. Using of K-means clustering method obtain threshold of image; introduce offsets which are automatically selected to achieve a pseudo-training set and unlabeled set; Finally, based on the statistical characteristics of semi-supervised support vector machines for image change and non-change class. The experimental results showed that: In the case without using noise reduction and the reduction in the number of samples, the proposed algorithm can maintain better classification, generalization and more stable detection accuracy.

Unsupervised SAR Image Segmentation Based on Conditional Triplet Markov Fields

Conditional random field (CRF) has been widely used in optical image and remote sensing image segmentation because of the advantage of directly modeling the posterior distribution and capturing arbitrary dependencies among observations. However, for nonstationary SAR images, applications of CRF often fail because of their nonstationary property. The triplet Markov field (TMF) model is well appropriate for nonstationary SAR image processing, owing to the introduction of an auxiliary field which reflects the nonstationarity. Therefore, we introduce an auxiliary field to describe the nonstationarity of the posterior distribution and propose an unsupervised SAR image segmentation algorithm based on a conditional TMF (CTMF) framework which combines the advantages of both CRF and TMF. The proposed CTMF framework explicitly takes into account the nonstationary property of SAR images, directly models the posterior distribution, and considers the interactions among the observed data. Experimental results on real SAR images validate the effectiveness of the algorithm proposed in this letter.

Unsupervised Full-Polarimetric SAR Data Segmentation as a Tool for Classification of Agricultural Areas

Versatile, robust and computational efficient methods for radar image segmentation, which preserve the full polarimetric information content, are of importance as research tools, as well as for practical applications in land surface monitoring. The method introduced here consists of several steps. The first step is a (reversible) transform of the full polarimetric radar information content into nine backscatter intensity values. The next steps relate to unsupervised clustering encompassing a simple region-growing segmentation (incomplete and over-segmented) followed by model-based agglomerative clustering and expectation-maximization on the pixels of these segments. Classification is achieved by Markov random field filtering on the original data. The result is a series of segmented maps, which differ in the number of (unsupervised) classes. For a (compatible) supervised approach, only the first and last step have to be applied. Results are discussed for the agricultural areas Flevoland in The Netherlands (AirSAR data) and DEMMIN in Germany, using the NASA/JPL AirSAR system and the DLR ESAR system, respectively. The applications include the use of groundtruth for legend development, the check for groundtruth completeness, and the construction of a bottom-up hierarchy of the characteristics that can be distinguished in the radar data. The latter gives important insights in physics of polarimetric radar backscattering mechanisms. Moreover, the relative importance of crop differences, (full-polarimetric) incidence angle effects and sub-classes (related to factors such as crop varieties, row direction or development stage) may be assessed. The overall classification results range between 84.3% and 98.0%, depending on number of observations dates and radar band(s) used, with higher values for the supervised approach, and substantially more thematic detail for the unsupervised approach.

An Analysis of Texture Measures in PCA-Based Unsupervised Classification of SAR Images

In single-band single-polarized SAR images, intensity and texture are the information source available for unsupervised land cover classification. Every textural feature measure identifies texture patterns by different approaches. For efficient land cover classification, textural measures have to be chosen suitably. Therefore, in this letter, the role of various intensity and textural measures is analyzed for their discriminative ability for unsupervised SAR image classification into various land cover types like water, urban, and vegetation areas. To make the algorithm adaptable, these textural features are fused using principal component analysis (PCA), and principal components are used for classification purposes. To highlight the effectiveness of PCA, the difference between PCA- and non-PCA-based classifications is also analyzed. Analysis of the role of texture measures for unsupervised classification of real-world SAR data with application of PCA is presented in this letter. The analysis of how every individual feature measure contributes for classification process is presented, and then, textural measures for a feature set are chosen according to their role in improving classification accuracy. By analysis, it is observed that the feature set comprising mean, variance, wavelet components, semivariogram, lacunarity, and weighted rank fill ratio provides good classification accuracy of up to 90.4% than by using individual textural measures, and this increased accuracy justifies the complexity involved in the process.