1.

Introduction

A radar’s antenna size determines whether it is a real aperture radar (RAR) or a synthetic aperture radar (SAR). RAR includes the noncoherent radars that are regulated by the antenna’s length.¹ Active radar uses high-frequency radar waves from the antenna to deliver images to the desired area of the landscape. If the antenna’s length is large, then the image captured is also big and contains a lot of Earth’s surface information.² However, lengthening the antenna often is not a feasible solution to get high-resolution image, because on many satellites and aircraft, there is limited ability to attach large-size antennae. To get around this problem, engineers and scientists have created the synthetic aperture. When many small antennas are joined, the array of smaller antennae seems to be much bigger, allowing for better and bigger data resolution. High-resolution images of the entire Earth’s surface are captured by SAR, which is a coherent radar connected to satellites and aircraft.³ The antenna size on many satellites and aircraft is fixed in RAR, whereas in the case of SAR, the antenna mounted on satellites and aircraft is synthetic in nature, which means it moves forward and backward. As it moves forward, the SAR antenna also moves and keeps transmitting high-frequency radar waves toward the Earth’s surface. These high-frequency radar waves hit the target and reflect.⁴ The backscattered energy is received by the SAR and processed as well. This processing of received SAR data requires a lot of time, as it is a high computational process. The received SAR data in the form of high-resolution SAR images is resultant of consistent interaction of constructive and destructive interference of transmitted high-frequency radar waves with target on the Earth’s surface.⁵

This constructive and destructive interference causes a lot of information loss as well as quality degradation during the image acquisition phase. This information loss as well as quality degradation appears in the SAR image in the form of a scattering phenomenon.⁶ This scattering phenomenon is speckle noise. This speckle noise is inherently present in the SAR images. It badly degrades the image quality of SAR image.⁷ The speckle noise has granular pattern that looks like the image is severely affected by salt-and-pepper noise.⁸ The SAR images need to preprocess before performing any kind of classification operation on SAR images. This preprocessing is SAR image despeckling.⁹ The elimination of speckle noise from SAR image is called as SAR image despeckling. This preprocessing method is a mandatory step in any kind of SAR image processing as it improves the image quality and makes further processing easier.¹⁰ The effect of speckle noise is much bad in comparison to other kind of noise patterns.¹¹ Its main reason is its multiplicative nature. The pattern of speckle noise follows a gamma distribution.¹² Speckle noise is multiplicative in nature. So, it is a multiplicative noise. The effect of multiplicative noise is comparatively bad than additive noise. Here in multiplicative noise, the external noise components get multiplied by the reference SAR data resulting into noisy SAR image.¹³

There are multiple traditional and nontraditional SAR image despeckling methods. These methods are grouped into Bayesian and non-Bayesian methods. The traditional SAR image despeckling methods are mostly based on Bayesian techniques in spatial domain. While the nontraditional SAR image despeckling methods are based on Bayesian techniques in transform domain and non-Bayesian techniques. Bayesian and non-Bayesian approaches both have a few benefits and downsides. Bayesian employs probabilities of data and probabilities of both hypotheses, but non-Bayesian does not use or compute the probability of the hypothesis. Non-Bayesian approaches rely on the probabilities of seen and unobserved data and do not need the development of a prior. Bayesian approaches, on the other hand, rely on a prior and the likelihood of the observed data. There are increasing claims that Bayesian statistics is much more convenient for clinical research, as well as attempts to use both non-Bayesian and Bayesian statistics for data processing in clinical research. However, the significance of Bayesian statistics also rises, as it is fundamental for machine learning algorithms, i.e., artificial intelligence-based systems. Consequently, we should see Bayesian approaches as another potent instrument for processing our data.

Speckle noise is a serious issue in the SAR images that needs to address in the preprocessing stage. This step is considered as a mandatory step when dealing with SAR image. Because of this, a continuous advancement has been seen in this field. A lot of research has been done in multiple domains with great results as well. The traditional methods are effective in this domain but the nontraditional methods that are hybrid are more effective in this domain. The despeckling results of nontraditional methods are seen as better than the traditional methods. This paper presents a comparative analysis of many nontraditional methods for SAR image despeckling.

1.2.

Availability of Dataset and Material

Figure 1 shows SAR image dataset. The real SAR image dataset is taken from the open public database. The description and availability of all these images are available in the cited references. The SAR image in Fig. 1(a) belongs to Ka-band and its wavelength ranges from 1.1 to 0.8 cm available in Ref. 16. The SAR image in Fig. 1(b) belongs to C band and its wavelength ranges from 3.8 to 2.4 cm available at Ref. 17. The SAR image in Fig. 1(c) belongs to S-band and its wavelength range from 15 to 7.5 cm available at open database platform.^18,19 The SAR images in Figs. 1(d) and 1(e) are taken from the open database platform available at Ref. 20. SAR image in Fig. 1(d) belongs to same S band and SAR image in Fig. 1(e) belongs to P band and its wavelength ranges from 100 to 30 cm.

Fig. 1

SAR image dataset.

The results shown in Sec. 5.2 are evaluated and demonstrated under MATLAB environment, i.e., version: MATLAB R2022a. All the MATLAB codes are run on windows 10 pro–operating system. Processor: 11th Gen Intel(R) Core(TM), i5-1145G7 @ 2.60 GHz. Ram: 16.0 GB, Operating system type is 64-bit.

2.

Scattering and Polarization Processes

By adjusting the analyzed polarization in both transmit and receive pathways, radar may gather signals in various polarizations. The direction of oscillation of the transmitted radar signals is referred to as polarization. SAR sensors normally provide data that is linearly polarized, even though the sensors’ orientation may be at any angle. The symbol H stands for horizontal polarization, whereas the letter V stands for vertical polarization. Radar sensors have the benefit of being able to accurately regulate the polarization of the sent and received signals. For signals that are both V and H polarization, they would be labeled as VH. Another way to express this is to use the abbreviation HH for H and V signals, respectively.²²

The composition of the imaging surfaces is revealed by analyzing the signal intensity from these distinct polarizations, depending on the following scattering types: rough surface, volume, and double bounce.

The different scattering types are visually shown in Fig. 2.

Fig. 2

Different scattering type: rough surface, volume, and double bounce.²²

We must remember that various scattering types may have differing amounts of signal due to wavelength, since wavelength affects how far a signal may penetrate. As an example, a C-band signal can penetrate only the upper levels of a forest’s canopy, and so will encounter predominantly roughness scattering combined with a minor degree of volume scattering. However, an L-band or P-band signal will penetrate considerably deeper and so suffer much more volume scattering and more double-bounce scattering due to the tree trunk.²⁴ Different SAR bands with their defined wavelength are shown in Fig. 3.

Fig. 3

SAR band: (a) X-band (3 cm); (b) C-band (6 cm); (c) L-band (24 cm); (d) P-band (65 cm).²²

3.

Problem in SAR Image

Several factors may cause the SAR system’s performance to degrade and the picture quality to degrade. The nonlinearities of the SAR subsystem have a detrimental effect on the system’s capacity to resolve. Image acquisition is a substantial contributor to the noise in SAR imaging system. The sensor’s location and velocity inaccuracies, which cause geometric distortion in the SAR image, are further sources of noise introduction. There are mainly four types of problems found in SAR image.²⁵ The major source of their introduction is briefly discussed below:

Speckle noise is multiplicative in nature found mainly in SAR images and medical ultrasound images. The speckle noise in SAR images follows a gamma distribution as a noise pattern. The speckle noise in medical ultrasound images follows Rayleigh distribution as a noise pattern. The shape of gamma noise is very similar to Rayleigh distribution. The main difference between these two distributions is gamma distribution starts from zero while the Rayleigh distribution does not start from zero. Figure 4 shows classical speckle pattern. The speckle noise model is represented as

Fig. 4

Traditional speckle pattern.⁹

4.

Literature Survey

Noise2noise despeckling, which can train a deep neural network using only speckled SAR images, has shown highly promising results in recent research studies. If the dataset is well-registered, then the approach employed in Ref. 29 presents a similarity estimation to adjust for the little time volatility in the data. A transformer-based network for despeckling SAR images is introduced in Ref. 30. An encoder based on a transformer enables the proposed despeckling network to learn global relationships between picture areas, which improves despeckling performance. Synthetically produced speckled images are used to train the system end-to-end. A dictionary learning and multiweighted sparse coding-based integrated SAR image despeckling model is proposed in Ref. 31. The dictionary is first trained using sets of comparable image patches, each of which has the same structural properties. For an orthogonal dictionary with good sparse representation capability, a correctly tight frame is necessary. The proposed model³¹ is built using quick and efficient solution stages that conduct orthogonal dictionary learning, weight parameter updating, sparse coding and picture reconstruction concurrently. The authors in Ref. 32 provide an SAR image despeckling model based on a nonlinear hyperbolic-parabolic coupled partial differential equation (PDE). In this case, the edge variable is calculated using a different equation, resulting in improved edge information in the despeckled SAR images. The authors of Ref. 33 looked at some of the most used indicators for evaluating edge preservation. Afterward, a new reference-free index is suggested. Speckled and despeckled images use the ratio gradient to distinguish between two nonoverlapping areas on each side of each pixel. Despeckling of SAR images using a bidimensional empirical mode decomposition (BEMD)-based adaptive filtering approach is proposed by the authors of Ref. 34. SAR images with high-frequency noise are first divided into bidimensional intrinsic mode function (BIMF) levels using BEMD, and then the first BIMF level is filtered to remove the high-frequency noise. CNNs based PolSAR despeckling has been presented in Ref. 35, and it makes use of a matrix logarithm invertible transformation to make the polarimetric SAR (PolSAR) data easier to get analyzed by CNN. Speckle noise is eliminated from a damaged images by using a residual learning technique, which involves training a CNN to detect these speckles. Filtering a single-look speckled data set using the Shannon and Rényi entropies under G0 model was presented in Ref. 36.

In Ref. 37, the authors suggest a unique truncated nonconvex nonsmooth model to reduce the speckle noise in SAR pictures. Regularization and I-divergence integrity terms are also included in the formula. Despeckling performance is considerably improved by replacing a single-stream structure of convolutional layers with a multiple-stream structure to extract feature representations with multidirectional and multiscale characteristics, as shown in Ref. 38. To collect abstract aspects of an image in a certain frequency and direction band, the Contourlet CNN (CCNN) is developed with numerous separate subnetworks using the Contourlet technology. Using a combination of a deep learning network with noise reference and a similarity-based block matching, the authors of Ref. 29 have presented a new single-image speckling approach. CNNs are used in this method’s denoising because they can handle tiny picture patches. Deeper CNNs have never been used previously to reduce speckle in noisy SAR pictures, according to the authors of Ref. 39. The ResNet model’s many skip connections are also used in the suggested architecture of the authors. A method for applying skip connections has also been devised to ensure consistency. To train the network more consistently, a hybrid loss function was devised. A new SAR image despeckling method was developed by the authors in Ref. 40 to increase the performance of an encoder–decoder CNN architecture. The smallest scale introduces a context block to collect multiscale data. Using a blend of a revised despeckling gain and a total variation loss function, the gradient descent technique is used to train the model. In Ref. 41, a weighted sparse representation-based approach for despeckling SAR images is described. First, multiplicative noise is transformed into additive noise via the homomorphic transformation. To acquire the flexible dictionaries and sparse coefficients based on nonlocal self-similarity constraint, related patches are clustered together in a second step. Coefficients are also subjected to weighted regularizations to boost efficiency. Finally, despeckling images may be generated using an exponential transform. For SAR image despeckling, a transformer-based network was presented in Ref. 30. An encoder-based on a transformer enables the proposed method to learn global relationships between image areas, which improves despeckling performance. Synthetically produced speckled images are used to train the system end-to-end. SSD-SAR-BS, a self-supervised deep learning approach based on Bernoulli-sampling-based self-supervised deep learning, was suggested by the authors of Ref. 42 in response to recent work on self-supervised denoising. As a result of training on actual SAR images, Bernoulli-sampled image pairs (input–target) were generated. It was after this that a network was trained on these pairings of images. To improve network performance, a dropout-based ensemble was implemented.

In Ref. 43, an improved Frost filtering technique for SAR images is presented. By employing Lee’s filter coefficient, the Frost filter model incorporates an adaptively controlled decay factor for SAR images that better characterizes homogenous and edge portions of the image. With the purpose of emphasizing sound suggestions for reasonable training, the authors in Ref. 44 evaluated experimentally the influence of training set design on the effectiveness of SAR image despeckling. For SAR image despeckling,⁴⁵ presents a dilated residual shrinkage network. A main network and a shrinking subnetwork are part of the suggested technique. Convolution and residual learning are paired with soft thresholding to make up the core of the network. Within our network, we have soft thresholding and shrinking. The goal of Ref. 46 is to suggest a multi-step despeckling process: first, a CNN instructed under the fully developed speckle theory with a numerical loss function is used for despeckling; second, the speckle noise is predicted by the system for detecting the not fully developed regions where the system will generate artefacts via a statistical technique and a ratio edge detector. The authors of Ref. 47 compare the advantages and disadvantages of several training methodologies (synthetic, multitemporal, and hybrid). Three datasets were used to train four CNN-based algorithms for evaluation. Each training method has been tested on genuine SAR pictures to highlight the differences in performance between them.

To improve speckle filtering and texture preservation at the same time in SAR images, a deep encoder–decoder CNN architecture is presented in Ref. 48. The U-Net CNN has been adapted to meet this goal and has been changed and optimized in accordance with it. An algorithm for estimating speckle noise distribution and despeckled images has been reported in Ref. 49. For SAR collections in a wide range of landcover situations, this method performs well since it does not rely on any noise model. The authors of Ref. 50 developed a novel approach for despeckling SAR images that uses a deep learning engine to do nonlocal filtering. Using nonlocal filtering for SAR despeckling has shown to be quite successful. The underlying principle is to use picture self-similarity to estimate the hidden signal. It is possible to estimate the target pixel using pixel-wise nonlocal methods, which use weights based on patch-wise similarity measurements. The paper⁵¹ provides a noisy reference-based SAR deep learning filter that uses complementary photos of the same region taken at different times as training references to solve the problem of speckle noise. Parameter-sharing CNN is used in Ref. 51 to better use the picture information. The similarity of each pixel pair across the multiple photos is also exploited to enhance the training process, which helps to decrease the training mistakes caused by changes in land-cover between different eras. The coherent imaging approach of SAR provides SAR images with intense and randomly dispersed speckle, which causes substantial interference to later applications. To cope with the impacted images,⁵² suggest a network combining wavelet characteristics to despeckled the images and then assess the outcomes.

The authors of Ref. 53 came up with the image of using a recursive deep CNN model to remove speckles in SAR images. In the first place, the data-fitting and regularisation parts of the SAR variational model are split up into two separate problems: a data-fitting block and a deep CNN prior block, which are both separate problems. Finally, the gradient descent algorithm is used to solve the data-fitting block, and a predenoising residual channel attention network that uses dilation for the deep CNN prior block is used to train the network. The authors of Ref. 54 elaborate on the multilayer perceptron neural-network model for SAR image despeckling by employing a temporal series of SAR images. Unlike previous filtering approaches, this method may be taught using archival photos across an area of interest to understand the intensity features of test images and then flexibly decide the values and criteria for image despeckling by employing a neural network for image despeckling. The authors of Ref. 55 recommended using pretrained CNN models trained on additive white Gaussian noise to overcome the issue of speckle noise. The authors⁵⁵ further employed a multichannel logarithm method with Gaussian denoising to incorporate such CNNs. It was hypothesized in Ref. 56 that SAR images may be despeckled without the need of ground truth (despeckled images). Despeckled pictures are traditionally required for training in learning-based systems. Real-world situations are frequently difficult to get the ground truth, though. To that aim, we provide a method for restoring speckled pictures only using SAR data. Several various approaches to dealing with speckle noise are examined by Ref. 57, depending on the speckle removal job and the availability of multitemporal stacks of SAR data. As an alternative to the newly announced SAR speckle removal architecture, the first approach uses a CNN model trained to eliminate additive white Gaussian noise from real images. The authors in Ref. 58 proposed a wavelet thresholding-based SAR image despeckling techniques using 2D-DWT. Here, speckled SAR image is first preprocessed using iterative inverse variance-based nonhomomorphic filter. The low-frequency components are directed to bilateral filter, and high-frequency components are directed to modified Bayesian thresholding followed by interlevel method noise thresholding. The intralevel method noise thresholding is applied as a postprocessing operation to get the final despeckled image.⁵⁸ The main motive of research⁵⁹ is the identification of image counterfeiting using illumination discrepancies. The suggested method measures the illumination qualities of various surfaces or objects inside an image. The digital forensics algorithm finds illumination disparities in an image’s objects and delivers findings distinguishing between actual and false images. The objective of Ref. 60 is to construct an automated recognition model for the categorization of therapeutic plants using IoT and machine learning (ML) methods to improve the conventional medicinal system. Using Raspberry Pi 3 Model B+ (RPi) and the RPi camera, an intelligent system is presented to detect images of Indian medicinal plants in real-time and disclose their unique medicinal characteristics. The authors of Ref. 61 developed an enhanced energy-efficient fuzzy-based cognitive radio scheme for internet of things (IoT) networks based on the limits discovered in wireless communication networks. Cognitive radio-based heterogeneous wireless sensor area network, the standard technique, is compared to the suggested protocol.

5.

Comparative Analysis of SAR Image Despeckling Methods

This section mainly comparatively analyzes some of most cited and prevalent nontraditional methods. It also discusses some of the standard despeckling methods as well. All the nontraditional methods discussed are latest papers from 2018 to 2022. The comparison is done based on two perspectives. First comparison is theoretical analysis that is done based on objectives, methodologies, merits, and demerits. This comparison is shown in Table 1 of Sec. 5.1. Second comparison is done based on the analysis of qualitative and quantitative results. This is done in Figs. 5 and 6, Tables 2 and 3 of Sec. 5.2. These visual results are comparatively assessed based on factors like texture and edge preservation, speckle noise suppression, artifacts generation, etc. This detailed comparison is performed on some prevalent methods.^{1,3,4,29,31,34,38,39,46,54,62–67}

Table 1

Comparison of traditional and nontraditional despeckling methods.

Fig. 5

Despeckling results (a) 2018,¹ (b) 2020,³ (c) 2021,⁴ (d) 2022,²⁹ (e) 2022,³¹ (f) 2022,³⁴ (g) 2021,³⁸ (h) 2021,³⁹ (i) 2021,⁴⁶ and (j) 2019.⁵⁴

Fig. 6

Despeckling results (a) 2018,¹ (b) 2020,³ (c) 2021,⁴ (d) 2022,²⁹ (e) 2022,³¹ (f) 2022,³⁴ (g) 2021,³⁸ (h) 2021,³⁹ (i) 2021,⁴⁶ and (j) 2019.⁵⁴

Table 2

Comparative despeckling analysis based on NV, MSE, ENL, and CV on Fig. 1(c).

6.

Conclusion

This review surveys various nonconventional SAR image despeckling methods and a few traditional methods as well. It further discusses the basics of SAR imaging, including scattering and polarization effects, and problems present in SAR imaging. The compared methods range from the year 2018 to 2022. A more specific comparative analysis has been done of some of the major despeckling works in recent years. This comparison is done using theoretical analysis, qualitative analysis, and quantitative analysis. The theoretical analysis is done based on the objectives set, methods used, and merits and demerits of these papers. The qualitative analysis is done based on a certain set of factors that are discussed in the related section. The quantitative analysis is performed using parameters like NV, ENL, MSE, and CV. The analysis is done in detail by explaining all its perspectives. The survey performed on these methods mostly uses hybrid methodology. This will help the researchers to understand the standard methods used in the discussed papers and they will get an idea of how to design approaches for SAR image despeckling methods.