Quantitative Analysis Of Approaches And Development Of Optimal Wavelet Based Denoising Technique For SAR Data

Abstract

The importance of microwave and mm-wave sensors in the field of remote sensing has been growing in leaps and bounds over the last three decades, due to their inherent advantages of all weather, day and night operations. These frequencies are recently being used in other fields too, such as, in the fields of surveillance and medical imaging. Signal data from these sensors have different characteristics as compared to optical and IR(Infra Red) sensors. Due to the methods of data acquisition data from microwave and mm-wave sensors, need extensive and complicated signal processing, filtering, and calibration for reconstructing the image and make them useful for further applications. Wavelet based denoising and enhancement techniques attempted on SAR (Synthetic Aperture Radar) type of data, can yield encouraging and useful results. Most of the techniques developed using wavelet based concepts have mostly been attempted on optical images and simulated speckle noise or medical images. Many of the recent papers have also illustrated the utility of wavelet based techniques on data sets from missions such as ERS, JERS, and Radarsat. High resolution and multi mode, SAR data from state of the art mission like that of ISRO’s RISAT-1(Radar Imaging SATellite)provide ample scope for exploring wavelet based reconstruction techniques and denoising methods. Wavelet analysis actually provide a scope to ‘analyse globally’ but ‘operate locally’, which is found to be a very useful technique in the image processing domain of applications such as image compression, denoising, zooming, enhancements etc. Discrete Wavelet Transform is used to transform digital signals in the image to discrete coefficients in the wavelet domain.Wavelet based approaches have the potential to give good results which are comparable to conventional techniques using spatial domain filtering. Since microwave remote sensing sensor’s signals are heavily tarnished by noise and RF interferences, it is deemed to be a promising field to study and arrive at suitable methods for data denoising by using wavelet techniques. In the field of image denoising extensive research has been carried out and is still an active topic of interest using wavelet domain techniques. A number of interesting results have been shown specially on optical images, which are affected by additive noise. The techniques have also been successfully tried on SAR images which are afflicted with multiplicative speckle noise. Denoising using wavelet techniques, has mostly been done by transforming the data into additive domain by homomorphic techniques, and subsequently applying models which are well suited for dealing with additive white Gaussian noise. Major Contribution For carrying out this research work a variety of data from the RISAT-1 Stripmap mode images have been chosen. The types of images have been chosen to encompass different types of scattering surfaces and different types of areas with a variety of contrasts. One data is from the rural area with uniform fields and boundaries, with moderate backscattering. Another data is chosen having high backscattering targets, such as the city area having buildings, roads, malls, over-bridges etc. This scene also contains the corner reflectors which were deployed by us, for analyzing the point target response. Another scene is chosen from the Sunderban area having uniform low backscatter region with well marked water bodies and mangrove forests, which are typically found in the Sunderban estuarine areas. These images are used to find out the response of the attempted filtering methodology, which are analysed. The signals are analysed in the wavelet transform domain, and the sub-bands are investigated. In this study we have investigated wavelet based denoising on higher resolution SAR data sets of around 3m from RISAT-1. The motivation was to study a few of the mother wavelets, and focus on one particular mother wavelet which is well suited for denoising of SAR images; to explore the means of arriving at a suitable thresholding method, and subsequently, find out the optimal order of the selected mother wavelet, and arrive at an optimal level of decomposition which gives the best denoised results. The sub-bands are statistically analysed to find out the noise characteristics, for different types of regions. The high frequency components or the detailed wavelet coefficients areanalysed, in an attempt to explore the trends of noise behaviour, since these are the components most affected by noise. After observing the behaviour of the noise or standard deviations, for the various types of data sets, a technique of reduction of the noise was attempted by using a threshold having a simple linear relationship with the noise standard deviation. As the multiplicative factor was increased an increasing trend in the noise reduction and contrast improvement was observed. However, this trend shows saturation, after about a multiplicative factor of about four, after which there is no significant improvement. This is observed for the different data types selected, and thus provided us with a strong basis for choosing the thresholding limit. Of course soft thresholding gives more consistent results as compared to hard thresholding, and less artifacts. So, only soft thresholding results are being discussed in this study. Approach for thresholding selection is derived by studying the behaviour of wavelet components of the sub-bands, and a heuristic approach is thus arrived at based on the above. The quality metrics to quantify the performance and tuning of noise removal algorithms, though well established in optical, and passive imageries, are not suitable for SAR images. The reason is the multiplicative nature of noise, here. Any attempt to reduce the noise, invariably degrades the spatial resolution and blurs the data. Hence, the other aspect of this research work is to use a suitable metrics for performance evaluation of SAR denoising techniques, which addresses both the radiometric as well as the geometric aspect of the image. For analyzing the performance metrics, denoised images using well established, spatial domain adaptive filters have been used. The mother wavelet chosen for this study is Daubechies orthogonal wavelet function. Orthogonal wavelet functions will have no overlap with each other (zero correlation) when computing the wavelet transform, while non-orthogonal wavelets will have some overlap (nonzero correlation). Using an orthogonal wavelet, we can transform to wavelet space and back with no loss of information. The other aspects which have significant roles in image denoising and quality determination, are the order of the wavelet chosen, and the levels of decomposition required to attain a desired radiometric resolution. However as with any denoising of SAR image, the goal is to achieve the best possible radiometric resolution, but with minimal blurring or spatial information degradation. As more number of levels of decomposition is time consuming, an optimal choice of the levels is also crucial for determining the practical choice of the level. The order of geometric resolution achieved is about 5m, and the radiometric resolution is about 1.8dB for 1 level of decomposition followed by an averaging filter in spatial domain, after denoising using wavelet domain. Depending on the specific end application one may opt for higher levels of decomposition but at the cost of geometric resolution and edge smoothing. More than two levels of decomposition, gives good radiometry but at the cost of blurring. The preservation of the geometric fidelity of the corner reflector image after denoising is an added attribute of the approach proposed here, as has been shown in the analysis section. For 1 level of decomposition with D24 wavelet denoising, if a further smoothing of 5 window is done, then the quality of the image is much improved to almost about 1.05dB, with minimal loss of spatial resolution which comes out to be of the order of 7m. This may be useful for applications in agriculture, such as crop assessment, monitoring crop growth, forest species discrimination and delineation, flood mapping etc., as the requirement for finer spatial resolution is not that critical. It is to be noted that conventional multilook techniques used for SAR denoising in frequency domain of approach, can generate an equivalent radiometric performance, only at a spatial resolution of about 10m. It should be noted that many applications need better radiometry and not geometry. In such cases the optimal choice would be just two levels of decomposition using D24. It is observed that the choice of D24 with level one decomposition followed by a smoothing filter, is best suited for most civilian applications as it preserves the geometric fidelity, and spatial resolution and also gives good radiometric performance.