Currently, some of the implemented atmospheric correction processors for remote sensing spectral sensors, use masking algorithms based on thresholding of spectral indices with sensor Top-Of-Atmosphere (TOA) reflectance. This concept allows the use of a limited amount of spectral bands, which is optimal for multi-spectral sensors (~10-20 bands), but for the case of the high spectral dimensionality of hyper-spectral sensors, spectral thresholding underutilizes the number of available bands. Given this limitation, we propose a masking algorithm which performs spatial and spectral feature extraction based on a 2D convolutional neural network, fitting the model with the available classification maps from the Python-based Atmospheric COrrection (PACO) processor. The training samples are selected based on their uncertainty to belong to a given class, The validation is performed using two independent human expert labelled datasets. The resulting classification maps show an improvement from the original ones of PACO.

Synthetic Aperture Radar (SAR) images exhibit the presence of the so-called speckle, often considered noise, resulting from the fluctuations of the elemental scatterers present in a given resolution cell. Speckle hinders the interpretation of images and several other applications. In this regard, training a neural network to despeckle requires the presence of matched ground truth (clean images), which doesn’t exist for SAR imagery. Recent training strategies attempt to solve the issue of lack of training data (via self-supervised training) by exploiting the usage of several acquisitions (temporal diversity), the spatial diversity of image patches, or by utilizing the real/imagery information of the Single-Look Complex (SLC) SAR data. However, the information present in the spectral domain has not yet been explored to train a self-supervised network. This study presents a new way of obtaining training data based on the sublook processing of SAR data to train a self-supervised despeckling network. Our proposed approach is tested using TerraSAR-X data, with experimental results reflecting the validity of our novel training method, paving the way for a new self-supervised strategy for training neural networks for despeckling.

Most common focusing techniques for Synthetic Aperture Radar (SAR) Tomography (TomoSAR), e.g. Matched Spatial Filtering and Capon, make use of the conventional sample covariance matrix, obtained from a finite number of observations. Yet, structured covariance matrix estimates can be employed in lieu of the sample covariance matrix. Accordingly, our simulation study shows that Capon’s performance improves with the use of structured covariance matrices. These are obtained with the Subspace Fitting approach, properly adapted to TomoSAR. Numerical comparisons between structure and unstructured covariance matrices are presented.

Nonlinear inverse problems like Synthetic Aperture Radar Tomography are often ill-posed, since their solutions are very sensitive to small perturbations in the input data and are, therefore, difficult to compute numerically. Ill-posed problems are commonly tackled with regularization approaches; however, there is a crucial problem in regularization, related to the selection of regularization parameters. In the search for optimal values of such regularization parameters, this article addresses an extension of the L-curve method, called Θ-curve. Furthermore, aimed at reducing converge time, the k-criterion is added, based on the first and second derivatives of the L-curve.

Polarimetric focusing techniques for synthetic aperture radar (SAR) tomography (TomoSAR) pursue finding optimal polarization combinations to extract the associated scattering mechanisms and height of reflectors. Regularization approaches like weighted covariance fitting (WCF), implemented in an iterative manner, attain finer resolution than conventional focusing techniques (e.g., Capon). Such approaches normally entail the selection of a regularization parameter and a first estimate of the power spectrum pattern. Regularization parameter selection via L-Curve method requires providing the scattering vector; nonetheless, it may not be always available, especially when not working at full resolution. Manipulations previously done to the data covariance matrix (e.g., pre-summing) must be equivalent in the scattering vector, which may not be at all times feasible. Accordingly, this article suggests modifying the L-Curve method to work exclusively with data covariance matrices. The proposed novel strategy is applied to WCF, considering single and dual channels. Iterations are stopped based on the Akaike information criterion.

Speckle is inherent to all coherent imaging systems and affects SAR imagery in the form of strong intensity variations in pixels with similar backscattering coefficient, difficulting the interpretation of SAR data. In the context of the Beltrami filter, a polarimetric distance is utilized as part of a region growing algorithm to find and then average similar covariance matrices within a central window using an iterative scheme. The Beltrami filter has shown good results using a computationally expensive geodesic distance that takes into account the Hermitian positive definite nature of the polarimetric covariance matrices. The flexible nature of the Beltrami distance allows for the use of any polarimetric distance, allowing the study on the utilization of less computationally complex distances and their impact on speckle reduction. In this paper, an analysis on the effect of some of the commonly utilized polarimetric synthetic aperture radar (PolSAR) distances measures within the Beltrami despeckling filter will be presented.

Digital approaches concerning the I/Q imbalance problem in zero-intermediate frequency (zero IF) mixing scheme have one of two main objectives: the first one is knowing the parameters of a widely—linear system (WLS) that models this problem with the purpose of devising a compensation scheme (calibration). The second one is estimating the system gain and phase imbalance values before and after calibrating the system to determine the quality of the manufacturing and calibrating processes (testing). State—of—the—art techniques for both objectives should involve three main aspects to be optimized: computational complexity, reliability and performance. In addition, a framework to use a single technique for both calibration and testing has not been proposed yet. This paper presents a low—complexity and statistically—efficient WLS parameters estimation technique based on the auto—correlation property of Golay complementary sequences (GCS). The proposed technique allows for disregarding the power amplifier (PA) nonlinearities affecting the model, making it reliable. In addition, a framework to extract the phase and gain imbalances in the I/Q branches from the WLS estimations is presented. The low complexity operations of the proposed estimator allow for a cheap hardware implementation. The performance of this technique is illustrated under extreme test cases.

In this work we propose a low complexity detector for hybrid MIMO space-time codes based on linear dispersion coding. This detector is designed for systems with fewer receiver than transmitter antennas. We also present a fast, scalable hardware architecture (suitable for FPGA implementation) that performs the matrix inversion required by the detector. This architecture is based on the QR matrix decomposition, implemented with a low-complexity, parallel and pipelined CORDIC with fixed-point arithmetic. This results in a 60% reduction in complexity, compared with other reported algorithms. We show that the proposed detector achieves better bit-error rates than comparable double-space-time transmit diversity (DSTTD) fixed-point detectors.

A new architecture for detection of layered quasi-orthogonal space-time block codes (LQOSTBC) with receiver antenna selection over Rayleigh fading channels is presented. The LQOSTBC transmitter consists of a full-rate quasi-orthogonal space-time block code unit, plus one antenna operating as V-BLAST. The LQOSTBC receiver selects nRout of a total of NRavailable antennas using a low-complexity, nearly optimal algorithm. The proposed receiver is based on successive interference cancellation (SIC) and the QR decomposition, suggesting a simple hardware implementation. It is designed for zero-forcing (ZF) criterion and a spatially uncorrelated channel. Reduced complexity is achieved by means of a convenient rearrangement of the channel matrix elements. The detection scheme proposed is compared to similar, recently reported techniques.