1. Introduction
A synthetic aperture radar (SAR) is a kind of active microwave imaging radar capable of imaging all day without relying on the light source. In addition, it has the advantage of penetrating clouds and rain because of the microwave band, which is able to achieve all-weather imaging. These advantages make SAR images an important data source in remote sensing applications and they are widely used in forestry monitoring [
1], geological prospecting [
2], disaster assessment [
3], and other military or civilian fields [
4,
5,
6]. However, due to the principle that SAR imaging is based on the coherent superposition of radar echoes reflected by a large amount of randomly distributed scattering, it is inevitable that a multiplicative random noise called speckle will be generated in SAR images [
7,
8,
9]. The speckle noise seriously reduces the visual quality of images and makes both manual and automatic interpretation difficult to implement effectively. Therefore, the suppression of speckle is an important part in SAR images processing, which is of great significance for improving the image quality and finally promoting the wider application of SAR images.
To effectively reduce and remove the influence of speckle noise, many filtering methods based on a single SAR image have been proposed. In general, they can be categorized into three methods: (i) multi-look processing method; (ii) spatial filtering method; and (iii) transform domain filtering method. The first approach utilizes multiple sub-views generated by dividing the Doppler bandwidth of the synthetic aperture before imaging to form a multi-look image, or averages the adjacent individual pixels after imaging to obtain the filtered image. Since the multi-look method does not take into account the statistical characteristics of speckle noise, it is simple to implement. However, this technique loses a great deal of detail information and seriously reduces the resolution of the image while improving the quality of the image. With the aid of mature estimation theory, the spatial domain filtering method [
10,
11,
12,
13,
14] has been widely studied. The second method relies on the statistical model of the speckles and scenes, and many classical algorithms have been proposed, such as Lee’s filtering [
15], Kuan’s filtering [
16], Frost filtering [
17], and plenty of their enhanced methods [
18,
19]. Touzi made a comprehensive summary and analysis of these methods in [
20] and proposed effective approaches for the steady-state and unsteady multiplicative speckle noise models. Most of the aforementioned filtering methods are pixel-based, and the adaptive spatial filtering is performed using a strategy that is adaptive to the coefficient variation (CV) in the local window of the pixel. Although such methods can smooth off the speckle noise and maintain the details of the image in homogeneous regions, they will reduce the resolution of the image within the sliding window in the heterogeneous areas, for example, the boundaries and edges are somewhat blurred. The last kind of method is based on the transform domain, due to the favorable performance of wavelet and other multi-scale transformations in suppressing the noise of common optical images, and many transform domain filtering methods for SAR images based on different transforms have been proposed successively. For instance, the Contourlet transform-based method [
21], Shearlet transform-based method [
22], wavelet-based method [
23,
24], etc. By adjusting the transform coefficients, a better denoised result can be obtained, and there is less loss of edge information.
Apart from the abovementioned methods, there is a kind of method called non-local mean (NLM) filtering proposed by Buades et al. [
25] in 2005, which was a breakthrough compared to the traditional local denoising models. The NLM algorithm fully considers the similarity and redundancy of natural image content and applies it to the image filtering, which obtains the filtered value by considering all the points being in a similar context. Deledalle et al. [
10] extended the NLM to SAR images by using the generalized likelihood ratio (GLR) and Kullback-Leibler (KL) divergence to measure the similarity between SAR patches. Parrilli et al. [
26] proposed the SARBM3D method based on the idea of block matching and collaborative filtering, which refines the results by taking advantage of the local linear minimum mean square in the wavelet domain.
Although many methods have been put forward for single-date SAR image filtering, there is always a loss of spatial resolution or insufficient detail information due to the limitation of only using spatial information. With the rapid increase of massive SAR data and large amounts of historical archive data in recent years, the temporal resolution of SAR images is getting higher and higher, which provides a good environment for the proposal of multi-date filtering methods. Using multi-date images, we can take advantage of both temporal and spatial information to improve the filtering results without losing the details and edge information [
27]. Expanding the spatial filtering method to 3-D accounts for a large part of the multi-date filtering method. By estimating the local statistics of the pixel block sequence located at the same position in the time dimension, Bruniquel and Lopes [
28] presented an approach that exploited the Kuan’s filter to obtain the filtered time-series images. Coltuc et al. [
29] transformed the time dimension of the multi-date images into the frequency domain by using a discrete cosine transform and then adaptively filtered each frequency component to reduce the noise interference, and finally the filtered images were obtained through an inverse transform. In [
30,
31], a method is proposed using the mean of all the spatial windows in the same position of time-series to reduce the variance by weighted summation and obtain the filtered pixels.
Most multitemporal filtering methods can realize the suppression for speckle noise without the loss of spatial resolution, but they assume that the spatial pixels remain unchanged over time, which makes them simply suitable for short-term multi-date images with homogeneous areas. However, the ground objects may change frequently due to human or natural influences in most actual situations, so an algorithm was proposed to improve the application scope of traditional multi-date filtering methods by finding out the similar pixels in time-series [
32]. In addition, Su et al. [
33] and Chierchia et al. [
34] have proposed two kinds of multitemporal SAR filtering approaches combined with non-local mean idea and time-series images, respectively, which achieved outstanding results in SAR image denoising. One of the most important parts in the above methods is the selection and design of the similarity measurement method, which determines the accuracy of the last-found unchanged points. Nevertheless, the similar measurement called the CV test used in [
32] was followed by the spatial adaptive filtering method, and it is not well applied to high-dimensional data. To solve those problems, a multi-date filtering approach considering the data structure should be developed, indicating a better suitability for long-term time-series data with changeable areas.
In this paper, we propose a novel method called the sliding time-series likelihood ratio (STSLR) test that is appropriate for the similarity measurement of the 3-D case, which makes full use of the multi-dimensional structure and the characteristics of time-series data to enhance the accuracy of the similar points search and reduce the misdetection rate. Firstly, the similarity comparison between patches is conducted by bi-date analysis using a two-sample KS test in order to generate the 3-D patch stacks. Secondly, an STSLR test is used to further improve the detection rate of similar points. Finally, the statistical mean is calculated according to the similar pixels obtained along the time axis.
This paper is organized as follows:
Section 2 presents the overall framework of the proposed multitemporal filtering method in detail. In
Section 3, the experimental results and analysis based on simulated data and real SAR datasets are shown to verify the effectiveness of the proposed method.
Section 4 discusses the results. Finally, the conclusions are presented in
Section 5.
4. Discussion
Compared with other multitemporal filtering approaches, the method proposed in this paper can remarkably improve the filtering effect, mainly owing to two factors: (i) Adopting the strategy of a two-step similarity measure, which improves the accuracy of finding unchanged points; and (ii) according to the data pattern, two more appropriate methods and criterion of similarity measurement based on hypothesis testing are applied to further enhance the success rate of similar points detection and to simultaneously reduce the commission error. For the aforementioned reasons, our method has great potential for the suppression of speckle noise and the preservation of spatiotemporal information in complex and changeable regions.
To illustrate the process of searching unchanged points, the analysis results of a pixel within the building areas are shown in
Figure 13, where the buildings were constructed on the penultimate image in Dataset 1.
Figure 13a is the similarity measure matrix obtained by the two-sample KS test from the first step, and
Figure 13b shows the result of the second step using the STSLR test, which includes the similarity information of all pixels on the time-series in this position. It can be observed that the first matrix contains some erroneous results considered to be dissimilar, while in the second one, the wrong detection results are eliminated completely.
Figure 13c shows the filtered results of all images based on the first and second similarity measure matrix, respectively, which indicates the advantages of the two-step strategy combining the proposed hypothesis testing method.
By suppressing the speckle noise, the multitemporal filtered SAR images can restore the structural features, including the contours, edges, and textures of real geographical information, considerably improve the image quality, and enable better interpretation of the image, especially for the high resolution SAR images, providing more reliable data for subsequent multitemporal applications such as forest monitoring, land cover change detection, and target recognition.
5. Conclusions
The traditional speckle suppression methods for SAR images either have the problem of resolution and detail information loss, or changes of the spatiotemporal information of the filtered images. To solve these problems, a novel multitemporal filtering method based on hypothesis testing was proposed, which can not only suppress the noise without the loss of spatial resolution, but also ensure the consistency of the images before and after filtering as far as possible. To combine the hypothesis testing approaches and improve the filtering results, the strategy of a two-step similarity measurement was conducted. In step 1, the similar patches for each date were extracted according to the bi-date analysis using the two-sample KS test method. Then, in step 2, a new test method called the STSLR test was proposed to process the patch stacks generated by the first step, which exploited the peculiarity of the multi-dimensional structure to further enhance the filtering results. By averaging the unchanged pixels, the final filtered images were obtained. To demonstrate the performance of the proposed method, we carried out the experiments using synthetic multitemporal SAR data and two real time-series datasets in Tongzhou District of Beijing, China, which contained different land cover types and were acquired from 2012 to 2016 by TerraSAR-X. Our method showed the best performance in terms of both the qualitative and quantitative results. However, there are still some deficiencies in this work that need to be improved. For example, the computational efficiency of the algorithm can be accelerated by parallel processing. Apart from that, more experiments with different scenes and resolutions can be performed to validate the effectiveness more adequately. In future research, we will consider combining the non-local method and our approach to further improve the filtering effect.