Search

Despite the application of state-of-the-art fully Convolutional Neural Networks (CNNs) for semantic segmentation of very high-resolution optical imagery, their capacity has not yet been thoroughly examined for the classification of Synthetic Aperture Radar (SAR) images. The presence of speckle noise, the absence of efficient feature expression, and the limited availability of labelled SAR samples have hindered the application of the state-of-the-art CNNs for the classification of SAR imagery. This is of great concern for mapping complex land cover ecosystems, such as wetlands, where backscattering/spectrally similar signatures of land cover units further complicate the matter. Accordingly, we propose a new Fully Convolutional Network (FCN) architecture that can be trained in an end-to-end scheme and is specifically designed for the classification of wetland complexes using polarimetric SAR (PolSAR) imagery. The proposed architecture follows an encoder-decoder paradigm, wherein the input data are fed into a stack of convolutional filters (encoder) to extract high-level abstract features and a stack of transposed convolutional filters (decoder) to gradually up-sample the low resolution output to the spatial resolution of the original input image. The proposed network also benefits from recent advances in CNN designs, namely the addition of inception modules and skip connections with residual units. The former component improves multi-scale inference and enriches contextual information, while the latter contributes to the recovery of more detailed information and simplifies optimization. Moreover, an in-depth investigation of the learned features via opening the black box demonstrates that convolutional filters extract discriminative polarimetric features, thus mitigating the limitation of the feature engineering design in PolSAR image processing. Experimental results from full polarimetric RADARSAT-2 imagery illustrate that the proposed network outperforms the conventional random forest classifier and the state-of-the-art FCNs, such as FCN-32s, FCN-16s, FCN-8s, and SegNet, both visually and numerically for wetland mapping. [ABSTRACT FROM AUTHOR]

There are currently various classification algorithms, each with its own advantages and limitations. It is expected that fusing different classifiers in a way that the advantages of each are selected can boost the accuracy in the classification of complex land covers, such as wetlands, compared to using a single classifier. Classification of wetlands using remote-sensing methods is a challenging task because of considerable similarities between wetland classes. This fact is more important when utilizing synthetic aperture radar (SAR) data, which contain speckle noise. Consequently, discriminating wetland classes using only SAR data is generally not as accurate as using some other satellite data, such as optical imagery. In this study, a new Multiple Classifier System (MCS), which combines five different algorithms, was proposed to improve the classification accuracy of similar land covers. This system was then applied to classify wetlands in a study area in Newfoundland, Canada, using multi-source and multi-temporal SAR data. The results demonstrated that the proposed MCS was more accurate for the classification of wetlands in terms of both overall and class accuracies compared to applying one specific algorithm. Therefore, it is expected that the proposed system improves the classification accuracy of other complex landscapes. [ABSTRACT FROM AUTHOR]

Speckle degrades the radiometric quality of a Synthetic Aperture Radar (SAR) image. Previous methods for speckle reduction have used a fixed-size window for filtering the entire image. This, however, may not be effective for the entire image, as land covers of different sizes require different filtering windows. In this paper, a novel method is proposed by which each pixel in the image is filtered with a window appropriate for the size of object within it. The real in-phase and the imaginary quadrature components of the SAR images determine the best window size and the pixels in the intensity image are filtered using their own optimal windows. The proposed method is presented for both single- and multi-polarized SAR images, and the results of several common filters that were modified are presented. This approach is applied to two RADARSAT-2 images: one over San Francisco, California, USA and the other over St. John’s, Newfoundland and Labrador, Canada, producing results that were similar to, or outperformed, comparable filters while retaining details and suppressing speckle effectively. While the method was successful for single-look intensity data, it offers great potential for multi-look and amplitude data as well. [ABSTRACT FROM AUTHOR]

Speckle noise significantly degrades the radiometric quality of PolSAR image and, consequently, decreases the classification accuracy. This article proposes a new speckle reduction method for PolSAR imagery based on an adaptive Gaussian Markov Random Field model. We also introduce a new span image, called pseudo-span, obtained by the diagonal elements of the coherency matrix based on the least square analysis. The proposed de-speckling method was applied to full polarimetric C-band RADARSAT-2 data from the Avalon area, Newfoundland, Canada. The efficiency of the proposed method was evaluated in 2 different levels: de-speckled images and classified maps obtained by the Random Forest classifier. In terms of de-speckling, the proposed method illustrated approximately 19%, 43%, 46%, and 50% improvements in equivalent number of looks values, in comparison with SARBM3D, Enhanced Lee, Frost, and Kuan filter, respectively. Also, improvements of approximately 19%, 9%, 55%, and 32% were obtained in the overall classification accuracy using de-speckled PolSAR image by the proposed method compared with SARBM3D, Enhanced Lee, Frost, and Kuan filter, respectively. This new adaptive de-speckling method illustrates to be an efficient approach in terms of both speckle noise suppression and details/edges preservation, while having a great influence on the overall wetland classification accuracy. [ABSTRACT FROM AUTHOR]

Synthetic Aperture Radar (SAR) data have gained interest for a variety of remote sensing applications, given the capability of SAR sensors to operate independent of solar radiation and day/night conditions. However, the radiometric quality of SAR images is hindered by speckle noise, which affects further image processing and interpretation. As such, speckle reduction is a crucial pre-processing step in many remote sensing studies based on SAR imagery. This study proposes a new adaptive de-speckling method based on a Gaussian Markov Random Field (GMRF) model. The proposed method integrates both pixel-wised and contextual information using a weighted summation technique. As a by-product of the proposed method, a de-speckled pseudo-span image, which is obtained from the least-squares analysis of the de-speckled multi-polarization channels, is also produced. Experimental results from the medium resolution, fully polarimetric L-band ALOS PALSAR data demonstrate the effectiveness of the proposed algorithm compared to other well-known de-speckling approaches. The de-speckled images produced by the proposed method maintainthe mean value of the original image in homogenous areas, while preserving the edges of features in heterogeneous regions. In particular, the equivalent number of look (ENL) achieved using the proposed method improves by about 15% and 47% compared to the NL-SAR and SARBM3D de-speckling approaches, respectively. Other evaluation indices, such as the mean and variance of the ratio image also reveal the superiority of the proposed method relative to other de-speckling approaches examined in this study. [ABSTRACT FROM AUTHOR]