# Multi-Feature Segmentation for High-Resolution Polarimetric SAR Data Based on Fractal Net Evolution Approach

^{*}

## Abstract

**:**

^{0}distribution statistical heterogeneity in order to combine the shape and statistical features of PolSAR data. The statistical heterogeneity between two adjacent image objects is measured using a log likelihood function. Second, a modified SLIC algorithm is utilized to generate compact superpixels as the initial samples for the G

^{0}statistical model, which substitutes the polarimetric distance of the Pauli RGB composition for the CIELAB color distance. The segmentation results were obtained by weighting the G

^{0}statistical feature and the shape features, based on the FNEA framework. The validity and applicability of the proposed method was verified with extensive experiments on simulated data and three real-world high-resolution PolSAR images from airborne multi-look ESAR, spaceborne single-look RADARSAT-2, and multi-look TerraSAR-X data sets. The experimental results indicate that the proposed method obtains more accurate and precise segmentation results than the other methods for high-resolution PolSAR images.

## 1. Introduction

#### 1.1. Background

- (1)
- Spatial characteristics: The decrease in the resolution cell provides richer spatial details of ground objects [11], such as significant geometric shape features and texture information.
- (2)
- Statistical characteristics: The scattering vectors from the homogeneous regions of medium- or low-resolution PolSAR data can be modeled using Gaussian distributions. The corresponding coherency matrices have a complex Wishart distribution [12]. However, in high-resolution PolSAR data, a significantly reduced number of sub-scatterers within a resolution cell leads to a greater heterogeneity [13], particularly in urban areas, where clusters can no longer be modeled using a Gaussian process.

#### 1.2. Related Work

^{0}distribution [30], or KummerU distribution [31], respectively. Beaulieu and Touzi [16] presented a hierarchical segmentation method using the K-distribution model, and verified its effectiveness for textured forested areas. Bombrun et al. [32] proposed a hierarchical maximum likelihood segmentation for high-resolution PolSAR images using KummerU distribution heterogeneous clutter models, which provided a better performance compared to the classical Gaussian criterion. However, it is essential to robustly estimate the parameters of these multiplicative models with enough samples. Generally, the segmentation results have obvious dentate boundaries, as the initial samples are usually collected from image blocks within square windows [32]. Moreover, the accuracy of parameter estimation decreases due to the differences between the square blocks and the actual boundaries of targets in high-resolution PolSAR data.

#### 1.3. The Proposed Approach

^{0}distribution has been shown to be flexible, computationally inexpensive, and capable of modeling varying degrees of texture [30,36], we substitute the G

^{0}distribution of statistical heterogeneity for the spectral heterogeneity in the traditional FNEA. Furthermore, we also utilize a modified SLIC algorithm to generate compact, approximately homogeneous superpixels as initial samples for the statistical model, which utilizes the polarimetric distance of Pauli RGB composition instead of the CIELAB color distance.

## 2. Methodology

^{0}distribution model for high-resolution PolSAR data; (2) initial sample generation for the statistical model using the SLIC algorithm with polarimetric features; and (3) segmentation with the G

^{0}statistical and shape features, based on the FNEA framework. The details of these are explained in subsequent subsections.

#### 2.1. FNEA

#### 2.2. Statistical Heterogeneity Measure by the G^{0} Model

**S**[2]:

**S**can be transformed into a three-dimensional single-look scattering vector using the complex Pauli spin matrix basis set [2,27]:

**T**is computed to suppress speckle using the average of $\mathit{k}$ of the surrounding pixels, as follows [2,27]:

#### 2.2.1. G^{0} Model for High-Resolution PolSAR Data

**k**follows the G

^{0}distribution, which is characterized by the following PDF [37]:

**k**, $E[\xb7]$ denotes the mathematical expectation, and $|\xb7|$ represents the determinant, while ${(\xb7)}^{-1}$ denotes the inverse.

^{0}distribution [30], which is characterized by the PDF, as follows:

#### 2.2.2. Statistical Heterogeneity Measure

#### 2.2.3. Parameter Estimation

^{0}distribution model is parameterized by scale matrix Σ and shape parameter α. The correct and reasonable merging objects are based on the proper estimation of the involved parameters.

#### 2.3. Initial Samples Generation for Statistical Model

_{p}) and $g$ The equivalent weight between ${d}_{p}$ and ${d}_{s}$ is utilized to calculate final distance D [35,43]. Once each pixel has been associated to the nearest cluster center, the cluster centers adjust to be the mean ${\left[R\text{}G\text{}B\text{}x\text{}y\right]}^{T}$ vector of all the pixels belonging to the cluster. This step can be repeated for 10 iterations, which is enough for most images [35].

#### 2.4. Segmentation with Statistical and Shape Features

^{0}distribution statistical heterogeneity for spectral heterogeneity in the original FNEA, and then combine the G

^{0}statistical and shape features to segment PolSAR data.

^{0}statistical and shape features can be obtained by weighting as follows:

^{0}statistics information and shape features. The value of g

^{2}is related to the start time using statistics information. A greater value of g

^{2}delays the use of statistical information.

Algorithm 1. FNEA-based Multi-Feature PolSAR Segmentation. |

1: INPUT: PolSAR data, samples number ${g}^{2}$, shape weight ${w}_{shp}$, scale parameter t. |

2: OUTPUT: segmentation result. |

3: Generate superpixels of PolSAR data using new SLIC algorithm by Equation (21). |

4: Produce initial image objects with the superpixels. |

5: do |

6: Get the number of image objects ${N}_{O}$, $a={N}_{O}$. |

7: for each image object i do |

8: for each adjacent object j of image object i do |

9: Estimate scale matrix Σ, shape parameter α using pixels in object by Equations (19) and (20). |

10: Compute the change of heterogeneity $\mathsf{\Delta}h\text{'}$ by Equations (3), (4), (14), (18), and (22). |

11: end for |

12: Compare all the $\mathsf{\Delta}h\text{'}$ to obtain the minimum and set it as $\mathsf{\Delta}{h}_{min}$ |

13: if $\mathsf{\Delta}{h}_{min}\le t$ then |

14: Merge image objects i and j, create a new image object i$\bigcup}j$, and delete objects i, j. |

15: $a=a-1$. |

16: end if |

17: end for |

18: while ($a<{N}_{O}$) |

19: Produce segmentation image and the vector of objects boundaries. |

## 3. Experiment and Results

#### 3.1. Description of the Experimental Data Sets

^{0}distribution data were adopted to generate the other classes. The initial scattering vectors and distribution parameters were estimated from a real data set. The Pauli RGB image is shown in Figure 1a, and the corresponding reference map is shown in Figure 1b. Figure 2 depicts the theoretical PDF of each class.

#### 3.2. Evaluation and Comparison

^{0}statistical features into the FNEA framework and pre-segmenting using SLIC, segmentation experiments were performed using different methods, namely: (a) FNEA segmentation based on Freeman decomposition without SAR statistical features (FFD); (b) FNEA segmentation with Pauli RGB image without SAR statistical features (FPD); (c) improved FNEA with G

^{0}statistical features start from square blocks (IFGB); (d) improved FNEA using Wishart statistical features with pre-segmenting by SLIC (IFWS); (e) improved FNEA using K statistical features with pre-segmenting by SLIC (IFKS); (f) improved FNEA using the G

^{0}statistical features with pre-segmenting by SLIC (IFGS); and (g) segmentation using the G

^{0}statistical features without shape features based on SLIC pre-segmenting (IFGS-S).

^{0}-based IFGS method obtained the best segmentation results for the areas with different degrees of heterogeneity, with a 98.77% detection rate and a 97.57% quality rate. The contrast between Figure 6c,f demonstrates that dentate boundaries appear when square blocks are taken as the initial samples for the statistical model. The boundaries of straight roads or other regular areas deviated. Compared to the IFGS method, the IFGS-S method obtained the wrong segmentation boundaries in partial regular building areas, due to the absence of utilizing the shape features as shown in Figure 6g. In conclusion, the proposed superpixel-based IFGS method, utilizing the G

^{0}distribution and shape features, obtained accurate and precise segmentation boundaries for the different areas.

^{0}-based IFGS method obtained better segmentation results (such as for roads in the forest) than the Wishart-based IFWS or the K-based IFKS method. For the IFGS-S method, the blurred segmentation boundaries occurred in areas with little change in statistics and polarimetric information, especially for the inner areas of forest, urban, and farmland in Figure 8g. According to Table 3, the proposed IFGS method obtained the highest accuracy with the least number of segmentation objects, similar to that of the single-look RADARSAT-2 image. Specifically, the detection and quality rates of the IFGS method were 88.46% and 72.07%, respectively. This demonstrates the effectiveness of the proposed method for multi-look high-resolution PolSAR images.

## 4. Discussion

#### 4.1. Main Features of the Proposed Method

^{0}distribution statistical features, is proposed. The proposed method was successfully applied to simulated and real-world PolSAR data sets.

^{0}distribution and shape information, based on FNEA. In related studies, traditional FNEA used polarimetric and geometric shape features for PolSAR image segmentation [8,9,25,26], which is easily influenced by speckle noise. Given the absence of statistical characteristics for PolSAR data, the statistical feature is introduced into the FNEA framework in the proposed approach. Many other methods use the classical complex Wishart distribution in order to represent scattering matrix statistics for PolSAR image segmentation [15,17,19,20,21,27]. Considering the ability to modeling varying degrees of texture [30,36], G

^{0}distribution is more suitable for heterogeneous or homogeneous areas in high-resolution PolSAR data compared to the Wishart or K distribution. Thus, the proposed method adopts the G

^{0}distribution to suppress speckle noise and obtains consistent segmentation results compared with the traditional FNEA method.

^{0}statistical model. It is essential to robustly estimate model parameters with enough samples. Most of the previous work collected initial samples from image blocks within square windows [32], which led to segmentation results with obvious dentate boundaries. To handle this problem, superpixels were introduced for pre-segmentation in the proposed approach. Given the excellent boundary adherence, SLIC was used in our approach, which combined the polarimetric distance of Pauli RGB compositions and spatial distance. This approach is capable of achieving accurate segmentation results with precise boundaries between the different areas.

#### 4.2. Sensitivity Analysis of the Parameters

#### 4.2.1. Number of Initial Samples

^{0}distribution in our method. When the size of the superpixels was too small, parameter estimation of the G

^{0}distribution became unstable, which made the calculation of statistical heterogeneity inaccurate. On the other hand, statistical features were not adapted for superpixel generation. The time utilizing this statistic’s feature for the superpixel-based FNEA can be delayed when the size of superpixels ${g}^{2}$ become too big, which could affect the subsequent segmentation accuracy. Therefore, the proper size of a superpixel is one of key issues for the segmentation experiment.

^{0}distribution, and there was little in terms of statistical heterogeneity differences between adjacent objects in one class. Three different simulated PolSAR images were used in the experiment, which only contained forest, crops, and roads, and is mentioned here in Section 3.1. The three simulated PolSAR images were divided into different sizes of blocks to calculate the standard deviation of the normalized G

^{0}heterogeneity ($\mathsf{\Delta}{h}_{stt}/{g}^{2}$) between adjacent objects [32]. Figure 11 shows the changes of $\mathsf{\Delta}{h}_{stt}/{g}^{2}$ with different sizes of blocks. As observed in Figure 11, the standard deviation became stable when the size of the blocks was large enough. In contrast, the standard deviation increased sharply when the size of the blocks was less than 16 pixels, which means that there were large statistical heterogeneity differences, due to the unstable parameter estimation using a small number of samples. Consequently, the superpixels should contain at least 16 pixels in order to ensure the stable calculation for G

^{0}heterogeneity.

^{0}distribution for the proposed method.

#### 4.2.2. Weight of Features

#### 4.2.3. Scale of Segmentation

#### 4.3. Time Performance Analysis

#### 4.4. Accuracies, Errors, and Uncertainties

^{0}distribution plays an important role in the statistical feature-based segmentation method. For the proposed approach, measurement of statistical heterogeneity and correct object merging depend on the proper estimation of the involved parameters. The determination of the scale parameters is another factor that causes segmentation errors for the proposed method. The scales of different types of targets exhibit differences due to their inequitable heterogeneities. An inappropriate scale parameter leads to under-segmentation in homogeneous regions or over-segmentation in heterogeneous regions, reducing the segmentation accuracy of the target of interest. Future development of this approach should include an accurate parameter estimation method of the G

^{0}distribution and the determination of the segmentation scale.

## 5. Conclusions

^{0}distribution statistical heterogeneity to combine shape features and statistical features of PolSAR data. Second, a modified SLIC algorithm was utilized to generate compact, approximately homogeneous superpixels as the initial samples for the G

^{0}statistical model, which substituted the polarimetric distance of Pauli RGB composition for the CIELAB color distance.

## Acknowledgments

## Author Contributions

## Conflicts of Interest

## References

- Jiao, L.; Liu, F. Wishart deep stacking network for fast POLSAR image classification. IEEE Trans. Image Process.
**2016**, 25, 3273–3286. [Google Scholar] [CrossRef] [PubMed] - Yang, S.; Chen, Q.; Yuan, X.; Liu, X. Adaptive coherency matrix estimation for polarimetric SAR imagery based on local heterogeneity coefficients. IEEE Trans. Geosci. Remote Sens.
**2016**, 54, 6732–6745. [Google Scholar] [CrossRef] - Ressel, R.; Singha, S. Comparing near coincident space borne C and X band fully polarimetric SAR data for arctic sea ice classification. Remote Sens.
**2016**, 8, 198. [Google Scholar] [CrossRef] - Wei, J.; Zhang, J.; Huang, G.; Zhao, Z. On the use of cross-correlation between volume scattering and helix scattering from polarimetric SAR data for the improvement of ship detection. Remote Sens.
**2016**, 8, 74. [Google Scholar] [CrossRef] - Lang, F.; Yang, J.; Li, D.; Zhao, L.; Shi, L. Polarimetric SAR image segmentation using statistical region merging. IEEE Geosci. Remote Sens. Lett.
**2014**, 11, 509–513. [Google Scholar] [CrossRef] - Liu, H.; Wang, Y.; Yang, S.; Wang, S.; Feng, J.; Jiao, L. Large polarimetric SAR data semi-supervised classification with spatial-anchor graph. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.
**2016**, 9, 1439–1458. [Google Scholar] [CrossRef] - Cheng, J.; Ji, Y.; Liu, H. Segmentation-based PolSAR image classification using visual features: RHLBP and color features. Remote Sens.
**2015**, 7, 6079–6106. [Google Scholar] [CrossRef] - Qi, Z.; Yeh, A.G.; Li, X.; Lin, Z. A novel algorithm for land use and land cover classification using RADARSAT-2 polarimetric SAR data. Remote Sens Environ.
**2012**, 118, 21–39. [Google Scholar] [CrossRef] - Qi, Z.; Yeh, A.G.; Li, X.; Xian, S.; Zhang, X. Monthly short-term detection of land development using RADARSAT-2 polarimetric SAR imagery. Remote Sens Environ.
**2015**, 164, 179–196. [Google Scholar] [CrossRef] - Suresh, G.; Melsheimer, C.; Koerber, J.; Bohrmann, G. Automatic estimation of oil seep locations in synthetic aperture radar images. IEEE Trans. Geosci. Remote Sens.
**2015**, 53, 4218–4230. [Google Scholar] [CrossRef] - Vasile, G.; Ovarlez, J.; Pascal, F.; Tison, C. Coherency matrix estimation of heterogeneous clutter in high-resolution polarimetric SAR images. IEEE Trans. Geosci. Remote Sens.
**2010**, 48, 1809–1826. [Google Scholar] [CrossRef] - Goodman, N.R. Statistical analysis based on a certain multivariate complex Gaussian distribution (an introduction). Ann. Math. Stat.
**1963**, 34, 152–177. [Google Scholar] [CrossRef] - Tison, C.; Nicolas, J.M.; Tupin, F.; Maitre, H. A new statistical model for Markovian classification of urban areas in high-resolution SAR images. IEEE Trans. Geosci. Remote Sens.
**2004**, 42, 2046–2057. [Google Scholar] [CrossRef] - Rignot, E.; Chellappa, R. Segmentation of polarimetric synthetic aperture radar data. IEEE Trans. Image Process.
**1992**, 1, 281–300. [Google Scholar] [CrossRef] [PubMed] - Liu, B.; Zhang, Z.; Liu, X.; Yu, W. Representation and spatially adaptive segmentation for PolSAR images based on wedgelet analysis. IEEE Trans. Geosci. Remote Sens.
**2015**, 53, 1–13. [Google Scholar] [CrossRef] - Beaulieu, J.M.; Touzi, R. Segmentation of textured polarimetric SAR scenes by likelihood approximation. IEEE Trans. Geosci. Remote Sens.
**2004**, 42, 2063–2072. [Google Scholar] [CrossRef] - Alonso-Gonzalez, A.; Lopez-Martinez, C.; Salembier, P. Filtering and segmentation of polarimetric SAR data based on binary partition Trees. IEEE Trans. Geosci. Remote Sens.
**2012**, 50, 593–605. [Google Scholar] [CrossRef] - Liu, B.; Hu, H.; Wang, H.; Wang, K.; Liu, X.; Yu, W. Superpixel-based classification with an adaptive number of classes for polarimetric SAR images. IEEE Trans. Geosci. Remote Sens.
**2013**, 51, 907–924. [Google Scholar] [CrossRef] - Qin, F.; Guo, J.; Lang, F. Superpixel segmentation for polarimetric SAR imagery using local iterative clustering. IEEE Geosci. Remote Sens. Lett.
**2015**, 12, 13–17. [Google Scholar] - Zhang, Y.; Zou, H.; Luo, T.; Qin, X.; Zhou, S.; Ji, K. A fast superpixel segmentation algorithm for PolSAR images based on edge refinement and revised Wishart distance. Sensors
**2016**, 16, 1687. [Google Scholar] [CrossRef] [PubMed] - Ersahin, K.; Cumming, I.G.; Ward, R.K. Segmentation and classification of polarimetric SAR data using spectral graph partitioning. IEEE Trans. Geosci. Remote Sens.
**2010**, 48, 164–174. [Google Scholar] [CrossRef] - Benz, U.C.; Hofmann, P.; Willhauck, G.; Lingenfelder, I.; Heynen, M. Multi-resolution, object-oriented fuzzy analysis of remote sensing data for GIS-ready information. ISPRS J. Photogramm. Remote Sens.
**2004**, 58, 239–258. [Google Scholar] [CrossRef] - Hay, G.J.; Blaschke, T.; Marceau, D.J.; Bouchard, A. A comparison of three image-object methods for the multiscale analysis of landscape structure. ISPRS J. Photogramm. Remote Sens.
**2003**, 57, 327–345. [Google Scholar] [CrossRef] - Burnett, C.; Blaschke, T. A multi-scale segmentation/object relationship modelling methodolgy for landscape analysis. Ecol. Model.
**2003**, 168, 233–249. [Google Scholar] [CrossRef] - Benz, U.; Pottier, E. Object based analysis of polarimetric SAR data in alpha-entropy-anisotropy decomposition using fuzzy classification by eCognition. In Proceedings of the IEEE International Geoscience and Remote Sensing Symposium, Sydney, Australia, 9–13 July 2001; pp. 1427–1429. [Google Scholar]
- Gao, H.; Yang, K.; Jia, Y.L. Segmentation of polarimetric SAR image using object-oriented strategy. In Proceedings of the 2nd International Conference on Remote Sensing, Environment and Transportation Engineering, Nanjing, China, 1–3 June 2012; pp. 1–5. [Google Scholar]
- Cao, F.; Hong, W. An unsupervised segmentation with an adaptive number of clusters using the SPAN/H/α/A space and the complex Wishart clustering for fully polarimetric SAR data analysis. IEEE Trans. Geosci. Remote Sens.
**2007**, 45, 3454–3467. [Google Scholar] [CrossRef] - Ulaby, F.T.; Kouyate, F.; Brisco, B.; Williams, T.H.L. Textural information in SAR images. IEEE Trans. Geosci. Remote Sens.
**1986**, GE-24, 235–245. [Google Scholar] [CrossRef] - Quegan, S.; Rhodes, I.; Caves, R. Statistical models for polarimetric SAR data. In Proceedings of the IEEE International Geoscience and Remote Sensing Symposium, Pasadena, CA, USA, 8–12 August 1994; pp. 1371–1373. [Google Scholar]
- Freitas, C.; Frery, A.; Correia, A. The polarimetric G distribution for SAR data analysis. Environmetrics
**2005**, 16, 13–31. [Google Scholar] [CrossRef] - Bombrun, L.; Beaulieu, J.M. Fisher distribution for texture modeling of polarimetric SAR data. IEEE Geosci. Remote Sens. Lett.
**2008**, 5, 512–516. [Google Scholar] [CrossRef] - Bombrun, L.; Vasile, G.; Gay, M.; Totir, F. Hierarchical segmentation of polarimetric SAR images using heterogeneous clutter models. IEEE Trans. Geosci. Remote Sens.
**2011**, 49, 726–737. [Google Scholar] [CrossRef] - Salembier, P.; Foucher, S. Optimum graph cuts for pruning binary partition trees of polarimetric SAR images. IEEE Trans. Geosci. Remote Sens.
**2016**, 54, 5493–5502. [Google Scholar] [CrossRef] - Feng, J.; Cao, Z.; Pi, Y. Polarimetric contextual classification of PolSAR images using sparse representation and superpixels. Remote Sens.
**2014**, 6, 7158–7181. [Google Scholar] [CrossRef] - Achanta, R.; Shaji, A.; Smith, K.; Lucchi, A.; Fua, P.; Süsstrunk, S. SLIC superpixels compared to state-of-the-art superpixel methods. IEEE Trans. Pattern Anal. Mach. Intell.
**2012**, 34, 2274–2282. [Google Scholar] [CrossRef] [PubMed] - Khan, S.; Guida, R. Application of Mellin-kind statistics to polarimetric G distribution for SAR data. IEEE Trans. Geosci. Remote Sens.
**2014**, 52, 3513–3528. [Google Scholar] [CrossRef] - Khan, S.; Guida, R. On single-look multivariate G distribution for PolSAR data. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.
**2012**, 5, 1149–1163. [Google Scholar] [CrossRef] - Gini, F.; Greco, M. Covariance matrix estimation for CFAR detection in correlated heavy tailed clutter. Signal Process.
**2002**, 82, 1847–1859. [Google Scholar] [CrossRef] - Anfinsen, S.N.; Eltoft, T. Application of the matrix-variate Mellin transform to analysis of polarimetric radar images. IEEE Trans. Geosci. Remote Sens.
**2011**, 49, 2281–2295. [Google Scholar] [CrossRef] - Doulgeris, A.P.; Anfinsen, S.N.; Eltoft, T. Classification with a non-Gaussian model for PolSAR data. IEEE Trans. Geosci. Remote Sens.
**2008**, 46, 2999–3009. [Google Scholar] [CrossRef] - Cloude, S.R.; Pottier, E. A review of target decomposition theorems in radar polarimetry. IEEE Trans. Geosci. Remote Sens.
**1996**, 34, 498–518. [Google Scholar] [CrossRef] - Jiao, H.; Luo, Y.; Wang, N.; Qi, L.; Dong, J.; Lei, H. Underwater multi-spectral photometric stereo reconstruction from a single RGBD image. In Proceedings of the Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA), Jeju, South Korea, 13–16 December 2016; pp. 1–4. [Google Scholar]
- Xu, Q.; Chen, Q.; Yang, S.; Liu, X. Superpixel-based classification using K distribution and spatial context for polarimetric SAR images. Remote Sens.
**2016**, 8, 619. [Google Scholar] [CrossRef] - Möller, M.; Lymburner, L.; Volk, M. The comparison index: A tool for assessing the accuracy of image segmentation. IEEE J. Sel. Top. Appl. Earth Obs.
**2007**, 9, 311–321. [Google Scholar] [CrossRef] - Clinton, N.; Holt, A.; Scarborough, J.; Yan, L.; Gong, P. Accuracy assessment measures for object-based image segmentation goodness. Photogramm. Eng. Remote Sens.
**2010**, 76, 289–299. [Google Scholar] [CrossRef]

**Figure 1.**Simulated single-look polarimetric synthetic aperture radar (PolSAR) data as the first data set: (

**a**) Pauli RGB image and (

**b**) the object spatial distribution reference map of the simulated image.

**Figure 3.**C-band, single-look RADARSAT-2 PolSAR image in Flevoland as the second data set: (

**a**) Pauli RGB image and (

**b**) the ground truth map of (

**a**).

**Figure 4.**L-band, two-look ESAR PolSAR image in Oberpfaffenhofen as the third data set: (

**a**) Pauli RGB image and (

**b**) the ground truth map of (

**a**).

**Figure 5.**X-band, six-look TerraSAR-X PolSAR image of Deggendorf as the fourth data set: (

**a**) Pauli RGB image and (

**b**) the reference image from Google Earth.

**Figure 6.**Segmentation results of the simulated image using different methods: (

**a**) FNEA segmentation based on Freeman decomposition without SAR statistical features (FFD); (

**b**) FNEA segmentation with Pauli RGB image without SAR statistical features (FPD); (

**c**) improved FNEA with G

^{0}statistical features start from square blocks (IFGB); (

**d**) improved FNEA using Wishart statistical features with pre-segmenting by simple linear iterative clustering (SLIC) (IFWS); (

**e**) improved FNEA using K statistical features with pre-segmenting by SLIC (IFKS); (

**f**) improved FNEA using the G

^{0}statistical features with pre-segmenting by SLIC (IFGS); (

**g**) segmentation using the G

^{0}statistical features without shape features based on SLIC pre-segmenting (IFGS-S); and (

**h**) SLIC (local enlarged drawing of lower-left part of the superpixel map).

**Figure 7.**Segmentation results of the single-look RADARSAT-2 PolSAR image using different methods: (

**a**) FFD; (

**b**) FPD; (

**c**) IFGB; (

**d**) IFWS; (

**e**) IFKS; (

**f**) IFGS; (

**g**) IFGS-S; and (

**h**) SLIC (local enlarged drawing of central left part of the superpixel map). (

**a**–

**e**,

**g**) show screenshots of the results in the red dashed rectangle of (

**f**).

**Figure 8.**Segmentation results of the two-look ESAR PolSAR image using different methods: (

**a**) FFD; (

**b**) FPD; (

**c**) IFGB; (

**d**) IFWS; (

**e**) IFKS; (

**f**) IFGS; (

**g**) IFGS-S; and (

**h**) SLIC (local enlarged drawing of central left part of the superpixel map). (

**a**–

**e**,

**g**) show screenshots of the results in the red dashed rectangle of (

**f**).

**Figure 9.**Representation maps (Pauli RGB images) of segmentation results for the ESAR data using different methods: (

**a**) GSRM; (

**b**) MW-WMRF; (

**c**) IFGS, t = 17, ${w}_{shp}$ = 0.05, ${g}^{2}$ = 16; and (

**d**) IFGS, t = 13, ${w}_{shp}$ = 0.05, ${g}^{2}$ = 16.

**Figure 10.**Segmentation results of the six-look TerraSAR-X PolSAR image using the proposed method: (

**a**) IFGS, t = 16, ${w}_{shp}$ = 0.05, ${g}^{2}$ = 16; (

**b**) IFGS, t = 12, ${w}_{shp}$ = 0.05, ${g}^{2}$ = 16.

**Figure 12.**Segmentation accuracy obtained using the IFGS method for the simulated data, ESAR, and RADARSAT-2 data, with different desired sizes of superpixels: (

**a**) detection rate; (

**b**) quality rate.

**Figure 13.**Segmentation accuracy obtained using the IFGS method for the simulated data, ESAR, and RADARSAT-2 data with different shape weights: (

**a**) detection rate; (

**b**) quality rate.

**Figure 14.**Segmentation accuracy obtained using the IFGS method for the simulated data, ESAR, and RADARSAT-2 data with different scales: (

**a**) detection rate; (

**b**) the number of result objects.

**Figure 15.**Segmentation results of the single-look RADARSAT-2 PolSAR image with different scales: (

**a**) scale = 12; (

**b**) scale = 16; (

**c**) scale = 20; (

**d**) scale = 25.

Method | ${\mathit{\rho}}_{\mathit{d}}$ (%) | ${\mathit{\rho}}_{\mathit{q}}$ (%) | TN |
---|---|---|---|

FFD | 94.33 | 89.27 | 26 |

FPD | 98.22 | 96.50 | 26 |

IFGB | 97.12 | 94.40 | 26 |

IFWS | 98.47 | 96.98 | 26 |

IFKS | 98.11 | 96.29 | 27 |

IFGS | 98.77 | 97.57 | 25 |

IFGS-S | 98.70 | 97.44 | 29 |

Method | ${\mathit{\rho}}_{\mathit{d}}$ (%) | ${\mathit{\rho}}_{\mathit{q}}$ (%) | TN |
---|---|---|---|

FFD | 89.30 | 78.26 | 558 |

FPD | 88.49 | 77.07 | 553 |

IFGB | 90.35 | 80.10 | 546 |

IFWS | 89.14 | 78.17 | 564 |

IFKS | 89.17 | 78.03 | 543 |

IFGS | 91.33 | 81.53 | 541 |

IFGS-S | 89.53 | 78.68 | 590 |

Method | ${\mathit{\rho}}_{\mathit{d}}$ (%) | ${\mathit{\rho}}_{\mathit{q}}$ (%) | TN |
---|---|---|---|

FFD | 83.31 | 65.13 | 348 |

FPD | 86.65 | 69.57 | 335 |

IFGB | 87.90 | 71.29 | 335 |

IFWS | 87.94 | 71.34 | 342 |

IFKS | 87.57 | 70.82 | 334 |

IFGS | 88.46 | 72.07 | 334 |

IFGS-S | 85.38 | 67.85 | 343 |

**Table 4.**Ratios ($\overline{\mathsf{\Delta}{h}_{shp}}/\overline{\mathsf{\Delta}{h}_{stt}}$) calculated for the simulated data, ESAR, and RADARSAT-2 image in segmentation with the shape weight of 0 and 1.

PolSAR Data | $\overline{\mathsf{\Delta}{\mathit{h}}_{\mathit{s}\mathit{h}\mathit{p}}}/\overline{\mathsf{\Delta}{\mathit{h}}_{\mathit{s}\mathit{t}\mathit{t}}}$ | |
---|---|---|

${\mathit{w}}_{\mathit{s}\mathit{h}\mathit{p}}=0$ | ${\mathit{w}}_{\mathit{s}\mathit{h}\mathit{p}}=1$ | |

Simulated image | 1.48 | 2.07 |

ESAR image | 2.09 | 2.25 |

RADARSAT-2 image | 6.55 | 8.25 |

PolSAR Data | Size | Scale t | Time (s) | ||
---|---|---|---|---|---|

SLIC | SP-FNEA | Total | |||

Simulated Data | 400 × 400 | 22 | 22.52 | 12.32 | 34.84 |

RADARSAT-2 | 1400 × 1400 | 16 | 267.85 | 238.60 | 506.45 |

ESAR | 800 × 800 | 17 | 88.04 | 64.66 | 152.70 |

TerraSAR-X | 541 × 541 | 16 | 40.07 | 27.20 | 67.27 |

© 2017 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (http://creativecommons.org/licenses/by/4.0/).

## Share and Cite

**MDPI and ACS Style**

Chen, Q.; Li, L.; Xu, Q.; Yang, S.; Shi, X.; Liu, X.
Multi-Feature Segmentation for High-Resolution Polarimetric SAR Data Based on Fractal Net Evolution Approach. *Remote Sens.* **2017**, *9*, 570.
https://doi.org/10.3390/rs9060570

**AMA Style**

Chen Q, Li L, Xu Q, Yang S, Shi X, Liu X.
Multi-Feature Segmentation for High-Resolution Polarimetric SAR Data Based on Fractal Net Evolution Approach. *Remote Sensing*. 2017; 9(6):570.
https://doi.org/10.3390/rs9060570

**Chicago/Turabian Style**

Chen, Qihao, Linlin Li, Qiao Xu, Shuai Yang, Xuguo Shi, and Xiuguo Liu.
2017. "Multi-Feature Segmentation for High-Resolution Polarimetric SAR Data Based on Fractal Net Evolution Approach" *Remote Sensing* 9, no. 6: 570.
https://doi.org/10.3390/rs9060570