Multibaseline and tomographic synthetic aperture radar (SAR) data are often affected by phase distortions known as phase screens. These distortions stem either from atmospheric effects or residual errors in platform motion. Calibrating and compensating for the phase screen is crucial to prevent spreading and defocusing in multidimensional tomographic imaging. Given the growing interest in artificial intelligence and deep learning, we aim to utilize their potential to develop a phase calibration process for SAR tomographic data. Our proposed framework is based upon a convolutional neural network (CNN) and generates training patches directly from the tomographic images under consideration, without relying on external references or resources. Once trained, the network effectively estimates phase distortions across the entire image; these are then used to calibrate the tomographic data. Experimental results from AfriSAR and UAVSAR tomographic datasets are included to showcase the effectiveness of the proposed solution.

In the above article [1], the correct reference associated with equation (2) on page 3, left column, is the below paper: 1) P. Imperatore and G. Fornaro, “Joint Phase-Screen Estimation in Airborne Multibaseline SAR Tomography Data Processing,” IEEE Transactions on Geoscience and Remote Sensing, vol. 62, pp. 1–14, 2024, Art. no. 4412614, doi: 10.1109/TGRS.2024.3446186.

The proposal of deep learning (DL) solutions for SAR image despeckling is recently widespread. Such solutions have been mainly designed in a DL perspective by leveraging the training and validation stage on the use of typical norm-based cost functions. For going beyond the DL perspective, in this paper we propose a SAR based validation stage by using SAR assessing metrics in the design and hyper-parameter selection of neural networks. In a first phase, SAR assessing metrics may be used only as validation metrics with the aim of highlighting critical issues that can not be spotted with standard image-processing quality metrics. In a second phase, the same SAR assessing metrics may be used directly for enhancing the DL solution by the addressing specific issues arisen during the previous SAR based validation stage. To this aim, three different DL SAR despeckling solutions and four different SAR assessing metrics have been considered. The outcome of this analysis shows the importance of including SAR knowledge in the training and validation stages of the design of a DL solution for SAR image despeckling.

This paper addresses the phase miscalibration within the broader scope of tomographic synthetic aperture radar (SAR) image focusing. Phase errors, typically independent from one acquisition to another, result in a dispersion and defocusing effect in the multi-dimensional imaging space. Dealing with this issue, particularly in the presence of volumetric scattering, poses a significant challenge. In this paper, we propose a novel framework for phase calibration in tomographic data. Our proposed framework employs deep learning solutions to estimate calibration phases, eliminating the need for model assumptions, external data sources, or phase unwrapping. In this approach, training data patch are directly derived from the data under consideration. Post-training, the network can effectively mitigate atmospheric phase screens or calibrate individual interferograms. Experimental results are included to demonstrate the effectiveness of the proposed method.

Synthetic Aperture Radar (SAR) serves as a fundamental system for Earth observation. Specifically, Polarimetric SAR (PolSAR) sensors capture images of diverse polarization scenes, enhancing the retrievable information. Due to their coherent nature, SAR images represent complex data affected by multiplicative noise known as speckle. The presence of this noise impedes image interpretation, making speckle removal a crucial pre-processing step for further applications. Recently, various Deep Learning (DL)-based methods have been proposed for speckle removal in PolSAR data, relying on different strategies for constructing training datasets. This complicates practical comparisons due to the absence of ground truth. In this paper, a comparison of the impact of different PolSAR despeckling methods on classification applications is studied. Some different PolSAR despeckling filters shows the different ability on improving the classification. In general, DL-based methods have greater improvement ability and potential.

The status of coastal zones has a great impact on economy and population and, therefore, the monitoring of shoreline is a crucial task. In order to have a wide scale and rapid monitoring, remote sensing represents a perfect opportunity, In particular, the ability of working all day in any meteorological conditions makes Synthetic aperture radar (SAR) system attractive for such task. At the same time, the ability of rapidly processing huge data makes deep learning an appealing solution. The aim of this work is to examine the effectiveness and potential of utilizing a deep learning solution for identifying and extracting coastlines from satellite SAR images. Firstly, a specific training dataset has been created using SAR data and ancillary information for retrieving position of coastline. Finally, the shoreline extraction has been performed as deep learning based segmentation task.

SAR despeckling is a fundamental task for improving the interpretation of SAR images and the performance of further tasks (classification and detection). The spreading of deep learning solutions has highlighted one main common issue: the availability of a training dataset. The construction of a training dataset is, indeed, limited by the absence of real ground truth. In this paper, an analysis of different approaches for constructing training dataset based on either fully simulation or real data has been carried out. The results show the importance of including real data properties within the training dataset.

Synthetic Aperture Radar (SAR) are fundamental tools for target classification and detection in the different applicative scenarios (military, agriculture, etc…). Extracting geometrical features of a target strongly help in its detection and classification. Indeed, the extraction of Attribute Scattering Center (ASC) characteristics is widely used from improving SAR target recognition. ASC extraction is a challenging task that requires the accurate estimation of tiny details (shape, orientation, etc…) from the SAR target backscattering. In this work, the aim is to exploit the potential of deep learning for ASC extraction. Considering a simulated environment, a deep-learning based classification solution is defined for extracting the target characteristics.We proposed a multi classification heads VGG solution, which can extract scattering parameters from complex images and also guarantee the estimation accuracy.

The building height estimation of urban environments is a challenging problem for Synthetic Aperture Radar (SAR). SAR Tomography (TomoSAR) conducts a series of acquisitions to realize a 3D reconstruction. Classical 3D focusing algorithms’ performance tends to be affected by the limited number of acquisitions, and the uneven baselines. Inspired by the advanced performance of TSNN on forest height estimation, in this study, we apply TSNN to reconstruct building height and we compare the obtained results with a classical Tomography approach. The experimental results are based on the data acquired by the DLR’s ESAR sensor at L-band over Dresden, Germany. The results illustrate the possibility of using the deep learning-based approach for building height estimation on urban environments.

Forest measurement is crucial for tracking climate change and quantifying the global carbon cycle. Synthetic Aperture Radar Tomography has become an effective technique to realize 3D forest structure monitoring. Recently a Deep Learning approach, named TomoSAR Neural Network (TSNN), has been proven a valid method for forest height and underlying topography estimation with polarimetric TomoSAR data. In this study, we evaluate the generalization ability of TSNN in working in different target areas, with different sensors and acquisition parameters. The experimental results demonstrate the robustness and generalization ability of TSNN.

Polarimetric Synthetic Aperture Radar (PolSAR) data plays an important role in Earth observation. In this field, deep learning (DL) method can achieve high classification performance on PolSAR image dataset and, in particular, vision transformer(ViT) has achieved significant breakthroughs. Compared with convolutional layers, ViT is able to extract global feature and find the global relationship, which can help to improve the performance of classification. The aim of this work is to exploit the potential of ViT for PolSAR classification. In this case, we propose a simple classification method based on transformer, called Pol-Trans. The PolSAR data is pre-processed to get the coherency matrix. Then the image patch of the pixel to be classified is flattened as the tokens. Finally, with the class embedding, our transformer can output the classification result of the PolSAR data. Our experiments on the ALOS2 PolSAR dataset of San Francisco shows the effectiveness of our method.

Synthetic aperture radar (SAR) are fundamental system for Earth Observation. In particular, polarimetric SAR (PolSAR) sensors provide images of a scene at different polarizations, enriching the information that can be retrieved. Due to their coherent nature, SAR images are complex data affected by a multiplicative noise, called speckle. The presence of this noise hinders the interpretation of images, making speckle removal a fundamental preprocessing step for further applications. Several deep learning (DL)-based approaches have been recently proposed for speckle removal in PolSAR data relying, due to the lack of a real ground truth, on different strategies for constructing training datasets making a real comparison complicated. In this work, a study on the construction of a training dataset for PolSAR despeckling is proposed. In particular, considering the analysis recently conducted on the construction of the dataset for training supervised neural networks for SAR amplitude despeckling, the aim is to extend such studies to the PolSAR case. In particular, the commonly used multitemporal approach, relying on the stack of real data, is compared with the so-called hybrid approach, in which a mixture of real and synthetic data is proposed. A specific DL solution has been chosen for such comparison, but the analysis could be extended to whatever supervised neural network. Moreover, for the sake of completeness, results are also compared with a well-assessed PolSAR despeckling filter in the literature.

In the context of Synthetic Aperture Radar (SAR) data, polarization is an important source of information for Earth's surface monitoring. SAR Systems are often considered to transmit only one polarization. This constraint leads to either single or dual polarimetric SAR imaging modalities. Single polarimetric systems operate with a fixed single polarization of both transmitted and received electromagnetic (EM) waves, resulting in a single acquisition channel. Dual polarimetric systems, on the other hand, transmit in one fixed polarization and receive in two orthogonal polarizations, resulting in two acquisition channels. Dual polarimetric systems are obviously more informative than single polarimetric systems and are increasingly being used for a variety of remote sensing applications. In dual polarimetric systems, the choice of polarizations for the transmitter and the receiver is open. The choice of circular transmit polarization and coherent dual linear receive polarizations forms a special dual polarimetric system called hybrid polarimetry, which brings the properties of rotational invariance to geometrical orientations of features in the scene and optimizes the design of the radar in terms of reliability, mass, and power constraints. The complete characterization of target scattering, however, requires fully polarimetric data, which can be acquired with systems that transmit two orthogonal polarizations. This adds further complexity to data acquisition and shortens the coverage area or swath of fully polarimetric images, compared to the swath of dual or hybrid polarimetric images. The search for solutions to augment dual polarimetric data to full polarimetric data will therefore take advantage of full characterization and exploitation of the backscattered field over a wider coverage with less system complexity. Several methods for reconstructing fully polarimetric images using hybrid polarimetric data can be found in the literature. Although the improvements achieved by the newly investigated and experimented reconstruction techniques are undeniable, the existing methods are, however, mostly based upon model assumptions (especially the assumption of reflectance symmetry), which may limit their reliability and applicability to vegetation and forest scenarios. To overcome the problems of these techniques, this paper proposes a new framework for reconstructing fully polarimetric information from hybrid polarimetric data. The framework uses Deep Learning solutions to augment hybrid polarimetric data without relying on model assumptions. A convolutional neural network (CNN) with a specific architecture and loss function is defined for this augmentation problem by focusing on different scattering properties of the polarimetric data. In particular, the method controls the CNN training process with respect to several characteristic features of polarimetric images defined by the combination of different terms in the cost or loss function. The proposed method is experimentally validated with real data sets and compared with a well-known and standard approach from the literature. From the experiments, the reconstruction performance of the proposed framework is superior to conventional reconstruction methods. The pseudo fully polarimetric data reconstructed by the proposed method also agree well with the actual fully polarimetric images acquired by radar systems, confirming the reliability and efficiency of the proposed method.

SAR (Synthetic Aperture Radar) sensors are fundamental tools for the Earth Observation. Actually, SAR images are affected by a multiplicative noise speckle that require a filtering step crucial for further application: classification, detection, etc. As in all image processing task, deep learning has been widely used for SAR image despeckling in the last years. Many methods have been proposed with different architectures, cost functions, training approaches, showing impressive performance. Actually, differently from natural domain denoising, an extensive comparison of such methods is still missing. As matter fact, such methods focus their comparison on few testing images. The aim of this paper is to propose and carry out the comparison among DL (Deep Learning) based methods not only in the testing phase but explointg the validation dataset used during the training for evaluating performance on SAR based metrics. The aim is to give an assessment on a more wide scenarios.

Synthetic Aperture Radar (SAR) are fundamental systems for the Earth Observation, providing images in any meteorological condition, during day and night. Due to their coherent nature, SAR images are complex data affected by a multiplicative noise called speckle impairing their interpretation. Therefore, speckle removal is a fundamental task for further applications. Following the interesting results obtained for single-channel despeckling, a deep learning approach is proposed for Polarimetric SAR (PolSAR) despeckling. In particular, the aim is to extend the outcome obtained on the construction of the dataset for SAR amplitude despeckling to the PolSAR case. In order to take fully advantage of such approach a multi-objective neural network has been considered. In particular, the hybrid approach has been used for creating a dataset for training a network following supervised approach. Comparison with state of art methods on real PolSAR images have shown the good versatility of such approach: the hybrid approach together with the multi-objective cost function lead the network to a good trade-off between noise suppression and texture preservation.

Forests play an important role in the global carbon cycle, and subsequently global climate change. Synthetic Aperture Radar Tomography (TomoSAR) can achieve three-dimensional forest structures relying on the multibaseline image acquisition. At present, plenty of TomoSAR approaches are based on fully polarimetric TomoSAR datasets which require costly data acquisition. The aim of this paper is to exploit the potential of deep learning for retrieving forest height by using single polarimetric data, going beyond the limitation of the requirement for full polarization. We design a fully connected network handling the forest height reconstruction problem from a classification task perspective. The network is trained using the covariance matrix elements of single polarimetric images acquired by ONERA over Paracou region as input, while LiDAR data acts as reference. Experimental results generally show good performance for forest height and underlying topography reconstruction and, a good robustness if compared with the results driven by fully polarimetric images.

Forest characterization and monitoring are highly important for tracking climate change, utilizing ecology resources, and biodiversity applications. Synthetic Aperture Radar Tomography (TomoSAR) provides the opportunity to reconstruct three-dimensional structures of the penetrable media relying on multi-baseline image acquisition. In forest applications, TomoSAR serves as a powerful technical tool for reconstructing forest height and underlying topography. Presently, a number of reconstruction methods are based on fully polarimetric TomoSAR datasets which require costly data acquisition. The aim of this paper is to go beyond the limitation of the requirement for full polarization by extending Tomographic SAR Neural Network (TSNN), a neural network for TomoSAR, to the case of single-polarimetric (SP) and dual-polarimetric (DP) TomoSAR data for retrieving forest height and underlying topography. Experimental results indicate that TSNN trained by SP or DP TomoSAR data is a powerful candidate to estimate forest height and underlying topography with high accuracy.

Synthetic aperture radar (SAR) systems can be designed with different polarimetric modalities. Most spaceborne SAR systems acquire dual-polarimetric data to meet various operational requirements. They are designed to capture more information about the Earth’s surface than single-pol systems and to cover a wider area than full-pol modalities. Dual-polarimetric data may not be as informative as fully polarimetric images. Several methods exist to augment dual-polarimetric images to take the capabilities of fully polarimetric data. Such methods, nevertheless, are either specific to special dual-pol modalities, i.e., compact modes, or rely on model assumptions that may not be valid in various scattering scenarios. In this article, a new framework for reconstructing fully polarimetric information from typical modalities of dual-pol data is proposed. The framework uses deep learning solutions to augment dual-polarimetric data without relying on model assumptions. Besides the specific architecture of the network used, which makes it efficient to extract distinctive features, a specific loss function is defined to account for the different scattering properties of the polarimetric data. Experiments on different real data show that the reconstruction performance of the proposed framework is superior to the conventional reconstruction method that widely experimented in the literature. Moreover, the pseudo-fully polarimetric data reconstructed by the proposed method closely match the actual fully polarimetric images acquired by radar systems, confirming the reliability and effectiveness of the proposed method.

Synthetic Aperture Radar (SAR) images are impaired by the presence of speckle. Despite the deep interest of scholars in the last decades, SAR image despeckling is still an open issue. Among different approaches, recently, many Deep Learning (DL) methods have been proposed following both supervised and unsupervised training approaches. There are two main challenges within the supervised framework: training data, and cost functions. Our approach builds training datasets which are varied and realistic using a multi-category Generalized Gaussian Coherent SAR simulator. It allows modeling a variety of SAR scenarios beyond the fully developed speckle hypothesis, which is only valid in homogeneous areas. Such multi-category simulated speckle is then applied to a noise-free reference obtained by multi-looking a temporal stack of actual SAR images in order to obtain the noisy input. We design an effective multi-objective cost function that accounts for texture, edge, and statistical properties preservation. We show the superiority of our approach assessing numerically and quantitatively its performance with three different SAR datasets.

Forest height and underlying terrain reconstruction is one of the main aims in dealing with forested areas. Theoretically, Synthetic Aperture Radar Tomography (TomoSAR) offers the possibility to solve the layover problem, making it possible to estimate the elevation of scatters located in the same resolution cell. This paper describes a deep learning approach, named Tomographic SAR Neural Network (TSNN), that aims at reconstructing forest and ground height using multipolarimetric multibaseline (MPMB) SAR data and Light Detection and Ranging (LiDAR) based data. The reconstruction of the forest and ground height is formulated as a classification problem, in which TSNN, a feed-forward network, is trained using covariance matrix elements as input vectors and quantized LiDAR-based data as the reference. In our work, TSNN is trained and tested with P-band MPMB data acquired by ONERA over Paracou region of French Guiana in the frame of the European Space Agency’s campaign TROPISAR and LiDAR-based data provided by the French Agricultural Research Center. The novelty of the proposed TSNN is related to its ability to estimate height with a high agreement with LiDAR-based measurement and actual height with no requirement for phase calibration. Experimental results of different covariance window sizes are included to demonstrate TSNN conducts height measurement with high spatial resolution and vertical accuracy outperforming the other two TomoSAR methods. Moreover, the conducted experiments on the effects of phase errors in different ranges show that TSNN has a good tolerance for small errors and is still able to precisely reconstruct forest heights.

PolSAR (Fully Polarimetric Synthetic Aperture Radar) imagery is used in Earth observation and remote sensing for various applications such as land use and land cover classification, change detection, etc. The occurrence of speckle in PolSAR images, however, degrades the performance of all image processing techniques and therefore prevents their full use for various applications. Several speckle reduction methods have been developed in the literature over the last forty years, highlighting the importance of this issue. Despite this extensive knowledge, speckle removal is yet an open problem that is far from being fully solved. Recently, Deep Learning (DL) has achieved great success in speckle reduction of SAR images. The data-driven nature of this technique provides improved flexibility and the ability to capture a variety of features from PolSAR images, thereby enhancing the performance of the speckle reduction process. In this paper, a new despeckling technique is proposed in the context of deep convolutional neural networks for de-noising the polarimetric covariance or coherence matrix. The method controls the training process with respect to several characteristic features of PolSAR images defined by the combination of three different cost functions. In particular, the goal is to balance the different features, including spatial details and speckle statistical properties, in the denoising process. The proposed method is experimentally validated with real airborne datasets and compared with existing despeckling approaches.

A fast and precise detection of floating plastics debris is necessary for monitoring and saving the sea ecosystem. Recent studies have demonstrated how remote sensing (and in particular satellites) can be helpful in such detection. In particular, data provided by satellite sensors allow to continuously monitoring wide areas of our planet interested by plastic litters. In this work, the possibility of exploiting different optical remote sensing satellite methods is investigated: the analysis is conducted starting from multi-spectral data and moving tom hyperspectral one. Data acquired from Sentinel 2 and from PRISMA sensors are considered showing the added value of these systems in the detection of marine litter within marine areas.

InSAR phase are affected by noise that impairs the performance of applications such as topography, 3D reconstruction, DEM profile, etc. Therefore a denoising step is fundamental. In the last decades many methods for InSAR phase denoising have been proposed such as Local, Non Local and other kind of filters. Inspired by the great success of deep learning in image denoising, methods relying on convolutional neural networks have been proposed recently. In this work, inspired by the outcomes of an amplitude despeckling filter, a multi-objective neural network for interferometric phase denoising is proposed: a cost function composed of three terms takes into account fringe, edges and statistical preservation. The encouraging results are validated quantitatively and qualitatively on a simulated dataset.

Elevation estimation of canopy and ground is one of the main aims in dealing with forest scenario using Synthetic Aperture Radar (SAR) Tomography. Theoretically, SAR Tomography (TomoSAR) provides layover solution, allowing to reconstruct the elevation of the different contributions collapsing in the same resolution cell. TomoSAR is commonly applied on both urban and vegetated areas. Within the latter scenario, one of the most interesting outcomes of TomoSAR is the possibility of separating the canopy and ground, allowing the reconstruction of their height maps. Within this paper, we propose a Deep Learning (DL) based method for TomoSAR. In particular, a neural network was trained for predicting the elevation value of canopy and ground of an area under investigation, based on a stack of SAR fully polarimetric multi-baseline acquisitions. The method uses the Light Detection And Ranging (LiDAR) data as reference and exploit a classification approach. The process was operated on a tropical forest over the TropiSAR2009 test site in Paracou, French Guiana. Testing results on real data are presented showing interesting results.

In the framework of deep learning for synthetic aperture radar (SAR) speckle reduction, the methods presented in the literature mainly focus on the definition of new architectures and cost functions for better catching and preserving the properties of a real SAR image. The achieved results are interesting and promising but with many left open issues. The main critical problem, shared by all the methods, is the construction of a training dataset. This is due to the lack of a noise-free reference. In this work, a comparison among different training approaches (synthetic, multitemporal, and hybrid) is carried out in order to analyze their benefits and drawbacks. Four convolutional neural network (CNN)-based methods have been trained with the three different datasets for their assessment. Results on real SAR images have been carried out showing the peculiarities of each training approach.

Among the large number of synthetic aperture radar (SAR) image despeckling approaches existing in literature, nonlocal (NL) filters have received a desirable boost. However, often NL approaches define the similarity criterion based on model assumptions, such as a fully developed speckle model. This assumption may not be verified in high-resolution images of urban environments. To address this issue, a standalone model-free despeckling framework is proposed in this article. The presented approach provides a generic framework for denoising a variety of SAR products, from a single-intensity/amplitude image to polarimetric and interferometric SAR data. In particular, the method is based on the empirical distributional similarity between the patch containing the pixel to be recovered and the patch containing a similar candidate pixel. To decide whether the patches follow a similar distribution, the Kolmogorov–Smirnov test is adapted. Finally, the restoration process aggregates the selected similar pixels based on their relative importance derived from their distribution similarities. To mitigate the blurring effect and preserve the resolution, the inhomogeneity of the ratio image is used to perform the bias reduction step. The designed generic despeckling filter was tested on different products of SAR data. The results show that the method proves to be an unbiased restoration approach and is able to preserve structures and textures. It works fully automatically and efficiently with single and multilook (and multichannel) images.

In recent years, there has been a growing interest in deep learning-based pansharpening. Thus far, research has mainly focused on architectures. Nonetheless, model training is an equally important issue. A first problem is the absence of ground truths, unavoidable in pansharpening. This is often addressed by training networks in a reduced-resolution domain and using the original data as ground truth, relying on an implicit scale invariance assumption. However, on full-resolution images, results are often disappointing, suggesting such invariance not to hold. A further problem is the scarcity of training data, which causes a limited generalization ability and a poor performance on off-training-test images. In this article, we propose a full-resolution training framework for deep learning-based pansharpening. The framework is fully general and can be used for any deep learning-based pansharpening model. Training takes place in the high-resolution domain, relying only on the original data, thus avoiding any loss of information. To ensure spectral and spatial fidelity, a suitable two-component loss is defined. The spectral component enforces consistency between the pansharpened output and the low-resolution multispectral input. The spatial component, computed at high resolution, maximizes the local correlation between each pansharpened band and the panchromatic input. At testing time, the target-adaptive operating modality is adopted, achieving good generalization with a limited computational overhead. Experiments carried out on WorldView-3, WorldView-2, and GeoEye-1 images show that methods trained with the proposed framework guarantee a pretty good performance in terms of both full-resolution numerical indexes and visual quality.

Interferometric Synthetic Aperture Radar is an effective and widely adopted tool for earth observation. Based on interferograms it is possible to infer several information about the observed area. Two main problems affecting the interferogram can limit its application: phase noise and phase wrapping. In this paper the attention is focused on the first issue. Several algorithms have been developed for interferogram restoration. Given the wide spread of Deep Learning (DL) in the framework of image processing, DL based algorithms have been proposed for interferogram denoising. Most of the efforts have been devoted in designing specific network architectures or training dataset, rather than on the definition of a specific cost function, well suited for the problem under investigation. The aim of this manuscript is to define a new multi-objective cost function, specifically thought for the interferograms restoration problem: the idea is to provide a cost function able to take into account multiple aspects of the data under investigation (i.e. multi-objective). The cost function is implemented within a Convolutional Neural Network and a specific realistic training dataset is built, to account the main characteristics of real interferograms. The final outcome of the paper is the proposal of a new robust and accurate interferometeric phase denoising algorithm (namely InSAR-MONet ), able to remove undesired noise and, at the same time, able to preserve important phase details. The assessment of the method is conducted on simulated and real datasets, comparing quantitatively and qualitatively InSAR-MONet with the state of the art interferometric denoising algorithms.

In this paper the effectiveness of a CNN based interferometric phase unwrapping algorithm combined with phase noise filtering is analysed. In particular, the considered processing chain relies on a pre-processing step with the nonlocal filter InSAR-BM3D followed by a deep CNN solution for restoring the absolute phase. The analyses is conducted on simulated data with different coherence values and aims at comparing the performance of the unwrapping with and without the pre-processing step. This paper is the first step towards a unique deep learning solution for jointly unwrapping and restoring the absolute phase.

SAR image interpretation is always impaired by speckle that is a multiplicative noise due to interference among the backscatterings from targets inside a resolution cell. Many algorithms for both single and multi-channel SAR despeckling have been proposed in the last forty years following different approaches. Recently, a multi-objective convolutional neural network, named MONet, has been proposed for single channel SAR despeckling. It relies on a mulit-objectvie cost function that takes into account three main aspects of the SAR images: noise removal, details and statistics preservation. Inspired by MONet, in this paper a deep learning method for InSAR phase filtering is proposed. The idea is to benefit from the multi-objective cost function defined in MONet that seems to perfectly fit with the interferogram denoising. This is the first step towards a solution able to provide a complete processed multi-channel product.

Among the different deep learning-based methods proposed for SAR image despeckling, the main issue seems to construct reliable training data sets. In the statistical-based solution MONet, which assumes square root Gamma distributed speckle in the simulation, the authors showed that despeckling results on actual SAR images are stringently related to the considered training dataset and its statistical distributions. This paper develops realistic simulated data sets for feeding the MONet architecture, including backscattering mechanisms arising in different existing SAR scenarios. We consider a generalized Gaussian coherent scatterer model for SAR correlated clutter simulation for this aim. The use of such simulation has a twofold effect within the considered framework: from one side, it allows generating several noisy patches, used as input data; on the other, it allows including different speckle distributions for different actual SAR scenarios. Results on SAR images show the effectiveness of such simulation.

Deep learning (DL) in remote sensing has nowadays become an effective operative tool: it is largely used in applications, such as change detection, image restoration, segmentation, detection, and classification. With reference to the synthetic aperture radar (SAR) domain, the application of DL techniques is not straightforward due to the nontrivial interpretation of SAR images, especially caused by the presence of speckle. Several DL solutions for SAR despeckling have been proposed in the last few years. Most of these solutions focus on the definition of different network architectures with similar cost functions, not involving SAR image properties. In this article, a convolutional neural network (CNN) with a multi-objective cost function taking care of spatial and statistical properties of the SAR image is proposed. This is achieved by the definition of a peculiar loss function obtained by the weighted combination of three different terms. Each of these terms is dedicated mainly to one of the following SAR image characteristics: spatial details, speckle statistical properties, and strong scatterers identification. Their combination allows balancing these effects. Moreover, a specifically designed architecture is proposed to effectively extract distinctive features within the considered framework. Experiments on simulated and real SAR images show the accuracy of the proposed method compared with the state-of-art despeckling algorithms, both from a quantitative and qualitative point of view. The importance of considering such SAR properties in the cost function is crucial for correct noise rejection and details preservation in different underlined scenarios, such as homogeneous, heterogeneous, and extremely heterogeneous.

SAR images are affected by multiplicative noise that impairs their interpretations. In the last decades several methods for SAR denoising have been proposed and in the last years great attention has moved towards deep learning based solutions. Based on our last proposed convolutional neural network for SAR despeckling, here we exploit the effect of the complexity of the network. More precisely, once a dataset has been fixed, we carry out an analysis of the network performance with respect to the number of layers and numbers of features the network is composed of. Evaluation on simulated and real data are carried out. The results show that deeper networks better generalize on both simulated and real images.

To cope with the lack of input-output training samples, deep learning (DL) methods for pansharpening usually resort to Wald's protocol or other similar downscaling processes. By doing so, the scaled versions of the multispectral (MS) and panchromatic (PAN) components serve as input while the original MS plays as output during the training phase. As a side effect, the informational gap between reduced and full scales causes a mismatch between the training and test phases. In fact, DL methods typically provide a pretty good performance at reduced scale, with a good margin over traditional solutions that tends to vanish in the full-resolution framework. In this work, we propose a training framework that involves both the reduced and the full scale versions of the multiresolution image samples. This is achieved thanks to a suitably defined loss which comprises costs for both scales. Our numerical and visual experimental results confirm that the proposed approach provides an improved performance in the full-resolution case.

In this paper, the potential of a deep learning approach for SAR tomography (TomoSAR) is investigated. TomoSAR is a powerful technique that allows the 3D reconstruction of objects lying on the Earth surface, by separating multiple scatterers with different elevations laying in the same range-azimuth resolution cell. In urban applications, the number of interfering scatterers is typically very small, so that the reconstruction of the elevation reflectivity profile can be faced as a statistical detection problem. Detection performance depends on how well the adopted statistical model fits to the observed scene. For complex urban scenarios this issue can greatly impair achievable accuracy of results. Then, we propose to exploit the neural networks' capabilities to learn the data generative model, in order to face the problem of signal model inaccuracies. In particular, in the assumption of a single scatterer, a neural network can be trained to solve a simple classification problem. Results on simulated and real data are presented.

Pansharpening is a classical data fusion task that is often necessary when dealing with data sensed through multiresolution acquisition systems. These systems, in fact, provide a single panchromatic band at full spatial resolution coupled with a multispectral lower resolution image of the same scene, which must be fused (pansharpened) to generate a full spatial-spectral resolution datacube. In the last few years, there has been a methodological shift in pansharpening towards the deep learning (DL) paradigm. Most DL solutions proposed thus far use self-supervised learning. Training is carried out on data at downgraded resolution, where ground truth data are also available. Then, the trained network is applied to perform pansharpening on native resolution data. As a consequence, such solutions show good results on low-resolution datasets, but less convincing results on full-resolution data, due to limited generalization ability. In this work, to address this problem, we enrich the training loss function with a perceptual term computed on full-resolution data, obtaining promising experimental results.

In SAR domain many application like classification, detection and segmentation are impaired by speckle. Hence, despeckling of SAR images is the key for scene understanding. Usually despeckling filters face the trade-off of speckle suppression and information preservation. In the last years deep learning solutions for speckle reduction have been proposed. One the biggest issue for these methods is how to train a network given the lack of a reference. In this work we proposed a convolutional neural network based solution trained on simulated data. We propose the use of a cost function taking into account both spatial and statistical properties. The aim is two fold: overcome the trade-off between speckle suppression and details suppression; find a suitable cost function for despeckling in unsupervised learning. The algorithm is validated on both real and simulated data, showing interesting performances.

We propose a new method for synthetic aperture radar (SAR) image despeckling, which leverages information drawn from coregistered optical imagery. Filtering is performed by patchwise nonlocal means, working exclusively on SAR data. However, the filtering weights are computed by taking into account also the optical guide, which is much cleaner than the SAR image, and hence more discriminative. To avoid injecting optical-domain information into the filtered image, an SAR-domain statistical test is preliminarily performed to reject right away any risky predictor. Experiments on two SAR-optical data sets prove the proposed method to suppress very effectively the speckle, preserving structural details, and without introducing significant filtering artifacts. Overall, the proposed method compares favorably with all the state-of-the-art despeckling filters, and also with our own previous optical-guided filter.

Fusing a multispectral image with a co-registered higher resolution single panchromatic band, provided by any multiresolution satellite systems, to rise the resolution of the former to that of the latter is known as pansharpening, and can be regarded as a guided super-resolution problem. Recently the use of convolutional neural networks (CNNs) has been extended to the pansharpening problem achieving state-of-the-art performance. Following this research line, the objective of this work was two-fold: provide a trained CNN model fitted to a specific sensor (WorldView-3) and explore a range of architectural configurations varied in both width and depth, seeking for the optimal one. Numerical and visual results show that the proposed solution compares favourably against reference methods.

We recently proposed a convolutional neural network (CNN) for remote sensing image pansharpening obtaining a significant performance gain over the state of the art. In this paper, we explore a number of architectural and training variations to this baseline, achieving further performance gains with a lightweight network that trains very fast. Leveraging on this latter property, we propose a target-adaptive usage modality that ensures a very good performance also in the presence of a mismatch with respect to the training set and even across different sensors. The proposed method, published online as an off-the-shelf software tool, allows users to perform fast and high-quality CNN-based pansharpening of their own target images on general-purpose hardware.