双极化SAR图像特征提取在海冰图像分割中的应用

需积分: 10 106 浏览量更新于2024-09-10 收藏 803KB PDF 举报

"yu_feature_extraction_of_dual_pol_SAR_imagery_for_sea_ice_image_segmentation" 在海洋冰层图像分割领域，双极化合成孔径雷达（SAR）影像特征提取技术发挥着至关重要的作用。双极化SAR数据，如RADARSAT-2提供的数据，相较于单极化数据，能提供更丰富的信息来区分不同的海冰类型。这篇由Peter Yu、A.K. Qin和David A. Clausi合作的研究论文深入探讨了已发表的特征提取和融合技术，以优化利用这些额外信息进行无监督的海冰图像分割。研究中，海冰图像的分割过程通过两种方式对双极化数据进行处理：(a) 将其转换为一个新的二维特征空间（多元），以及 (b) 融合成一个单一通道的特征空间（一元）。这两种方法都涉及对实际和合成的双极化SAR海冰图像进行多种变换，然后使用公认的SAR分割算法（例如迭代重加权聚类（IRGS））进行分割。结果表明，未经转换的原始数据在某些情况下也能提供良好的分割效果，但通过对数据进行适当的特征提取和融合，可以显著提高海冰类型的识别精度。例如，利用极化散射矩阵（PSM）、相干矩阵或克拉克散射系数等方法，可以从双极化信号中提取出反映海冰物理特性的特征。这些特征可能包括海冰的粗糙度、厚度和含盐量等信息，这些信息对于理解冰层状态和预测气候变化至关重要。在多元特征空间中，不同极化通道之间的相关性和对比度被强调，这有助于识别不同海冰类型的独特散射模式。而在一元特征空间中，通过将双极化信息融合成一个通道，可以减少噪声影响，同时保留关键的散射特性，简化了后续的分割任务。此外，论文还可能讨论了特征选择、降维技术（如主成分分析PCA）和特征融合策略，以减少冗余信息并增强分类性能。这些技术的应用可以提升图像分割的稳定性和鲁棒性，尤其在面对复杂环境变化和数据质量不均匀时。这项研究揭示了双极化SAR影像在海冰图像分割中的潜力，并提供了优化特征提取和融合的策略，对于监测和研究海冰变化、提升冰情预报的准确性具有重要意义。这些发现不仅适用于海冰分析，还可以推广到其他需要利用SAR数据进行地表特性识别的领域。

The computed gradient is normalized so that the largest gra-

dient in the scene has a value of 1.0 to ensure that MIRGS

scales properly for scenes with diﬀerent dynamic ranges. Once

the watershed (Vincent and Soille, 1991) is generated from the

normalized image gradient, the image is represented by a re-

gion adjacency graph (RAG) (Li, 2001) data structure where

each node represents a watershed region and where each graph

edge connects spatially adjacent regions.

Each watershed region is assigned an initial label via a K-

means algorithm (Duda et al., 2001) to initialize MIRGS. The

region-based K-means algorithm used in MIRGS is described

in (Qin and Clausi, 2010). MIRGS then enters an iterative

phase to ﬁnd a conﬁguration of labels for the regions that

globally minimizes a cost function. At each iteration, a label-

ing process is performed with Gibbs sampling (Geman and

Geman, 1984) to move the segmentation towards the optimal

conﬁguration. After each iteration, regions with the same la-

bels are merged to reduce the number of nodes in the RAG

by combining adjacent regions, which makes subsequent iter-

ations more eﬃcient as fewer nodes have to be considered.

The cost function that MIRGS minimizes to produce the

optimal segmentation consists of a feature space model and a

spatial context model (Qin and Clausi, 2010). The cost func-

tion considers a segmentation more likely to be “true” when

the regions assigned to each class are similar to each other in

feature space and when spatially adjacent regions belong to

the same class if the edge strength between them is weak. This

is similar to the Markov random ﬁeld (MRF) based multi-

level logistic (MLL) segmentation model (Derin and Elliott,

1987) but MLL does not consider the edge strength in its spa-

tial context model. The MIRGS model agrees more closely

with intuition: if there is a strong edge between two regions,

they are more likely to be from diﬀerent classes than when

there is no edge. MLL, in contrast, makes no such distinction

and favours results where adjacent regions are assigned to the

same class regardless of the edge strength between them.

3. Research objectives

To generate an accurate and consistent segmentation, the

MIRGS algorithm requires data with suﬃcient feature space

separability for diﬀerent ice classes, i.e. it should be possible

to discern a diﬀerence in feature space values between diﬀer-

ent classes (for example, one class might appear darker than

another class in the image). Additionally, MIRGS requires

the proper generation of the initial watershed and an image

gradient that presents strong boundaries between regions of

diﬀerent ice classes. As will be seen in Section 4, information

from both dual-polarization channels (HH and HV) are neces-

sary. Many strategies exist to use dual-polarization RS-2 data

to satisfy these requirements. The objective of this study is

to determine which of these strategies is the most eﬀective.

The following three strategies will be tested:

1. Direct MIRGS implementation: The most basic strat-

egy is to use the backscatter values from the HH and

HV channels directly in the multivariate formulation of

MIRGS, using the VFG gradient method that is already

implemented to create the watershed and image gradient.

2. Gradient combination: While the feature space separabil-

ity provided by the dual-polarization data is fully utilized

by Strategy 1, the VFG image gradient was not designed

with domain knowledge of dual-polarization data. VFG

tends to assign the highest strength only to edges that are

strongest in both the HH and HV channels while strong

edges that appear in only one of the two channels are as-

signed a lower value. However, strong edges that appear

in at least one of the channels are equally meaningful as

they denote a boundary between ice classes. Thus, there

is motivation for testing and comparing various gradi-

ent generation strategies that combine information from

both channels.

3. Feature extraction and image fusion: Another strategy

for making use of dual-polarization data is to fuse the

information from both channels into a single image ﬁrst

with feature extraction or image fusion techniques be-

fore segmentation in MIRGS. If feature space separability

can be maintained between all ice classes after mapping

each two dimensional feature vector to a one dimensional

value, then both the separability and the image gradient

requirements can be satisﬁed: all ice classes will have a

diﬀerent brightness in the fused image, which will nat-

urally cause edges between them to appear in the fused

image.

The experiments will test whether the basic multivariate

strategy, a modiﬁed gradient combination approach or feature

extraction will give the best results for the RS-2 data.

4. Data

ScanSAR Wide A has a pixel resolution of 100 m × 100 m,

with a pixel spacing of 50 m × 50 m. The full 500km swath

width spans approximately 10000 × 10000 pixels. The CIS

expects to use data from the co-polarization (σ

◦

) and the

cross-polarization (σ

◦

) channels for their operations and has

provided real-valued RS-2 imagery for testing. Each pixel in

the image is represented by a two dimensional feature vector

whose elements are σ

◦

and σ

◦

. The HH channel con-

tains the same information as that available from the single-

polarization RADARSAT-1 (RS-1) satellite. Complex-valued

images are not considered in this paper because these are not

used operationally by CIS.

A Gulf of St. Lawrence scene recorded on February 25,

2008 was tested in this paper. CIS provided operational ice

charts for the area on this date, which were created from RS-

1 data since CIS had not yet integrated RS-2 imagery into

their operational pipeline at that time. A manually segmented

ground-truth image was produced based on the ice charts for

a small part of the RS-2 scene (depicting an area north of

Anticosti Island) to use for validation purposes (Fig. 2). This

image represents ice appearance for an incidence angle range

of less than 10

◦

There are still ambiguities in the manual segmentation be-

cause certain ice types cannot always be reliably identiﬁed

from the backscatter images alone and because each polygon

in the CIS ice chart contains a mix of ice types but not the

exact pixel location of each type. However, there are small

patches of the original, full RS-2 scene where the ice type is

known, such as within polygons that have only one ice type

or a mix of distinctive ice types. Although the ice type in

剩余13页未读，继续阅读

baidu_37489064

粉丝: 0
资源: 2

双极化SAR图像特征提取在海冰图像分割中的应用

双极点雷达算法(DualPol) 的Python接口_python_代码_下载

SAR_DATA1.rar_SAR MATLAB_SAR 处理_SAR_原始数据_SAR数据_raw data

Beihang_SAR.rar_SAR MATLAB_matlab SAR_sar_matlab_北航SAR_北航SAR代码

sar_sift_SAR配准_配准_SAR_SAR-SIFT_SAR配准_

OpenSARSim.rar_C SAR_SAR_SAR仿真_SAR目标_sar simulation

GMTI_algorithm.rar_GMTI SAR_GMTI_ SAR_SAR_ GMTI_SAR_ 动目标_动目标检测

SAR_RDA_Time.zip_SAR RDA_SAR_RDA_SAR，原始_原始 SAR_多普勒，SAR

SAR_Filter.rar_SAR 滤波_SAR图像滤波_SAR滤波_matlab SAR_sar filter

RD.rar_RDA for SAR_RDA_ SAR_SAR_RDA_sar rd_多普勒雷达

SAR_Image.rar_SAR_SAR 分割_SAR分割_sar图像分割

最新资源