2. Information Extraction from Polarimetric SAR Images: A Review
observed that, presently, there is no algorithm that can suppress the sidelobes completely and simultaneously reduce the mainlobe width to further enhance the resolution when implemented on images sampled at non integer multiple of Nyquist rate. The present thrust is to formulate a simple algorithm which can completely wipe out the sidelobes and simultaneously reduce the mainlobe width to further enhance the resolution when implemented on SAR images sampled at non integer or unknown multiples of Nyquist rate.
In this thesis, we address these issues.
3
Landcover Mapping and Crop Classification Using PolSAR Images
Contents
3.1 An entropy based landcover classification scheme . . . 42 3.2 A Gini-index based landcover classification scheme . . . 47 3.3 Analysis of different entropies based landcover classification
schemes . . . 54 3.4 A Fully automated landcover classification scheme . . . 60 3.5 Summary . . . 63
3. Landcover Mapping and Crop Classification Using PolSAR Images
Anthropogenic changes in landcover is one of the alarming factors affecting global ecology.
Therefore the importance of acquiring complete, accurate and timely landcover information has grown over the years with the increasing concern about the ecological changes. Land- cover classification is one of the major remote sensing applications of radar polarimetry and has been a subject of recurring interest for the last few decades. Several techniques have been developed, recently, to correctly classify different landcover types. Various methods for supervised and unsupervised landcover classification have been reported in literature [24].
An unsupervised classification scheme based on the use of polarimetric entropy and alpha angle is widely used for land cover classification. In this chapter, we propose three dif- ferent landcover classification algorithms. These three algorithms are similar in principle to entropy/alpha based landcover classification scheme. We have shown that these algo- rithms perform better than the classic entropy/alpha based classification scheme in terms of classification accuracy and computational cost. The performance analysis of the proposed techniques are carried out using NASA/JPL AIRSAR L band datasets for San Francisco bay, USA and Flevoland, The Netherlands. The San Francisco data covers a mixture of ocean, urban, and vegetation land covers [7]. This dataset is used for terrain type classifica- tion. The Flevoland data covers a large agricultural area of horizontally flat topography and homogeneous soils. It contains several classes of crops and three other classes of bare soil, water, and forest [19]. This dataset is used for crop type classification and identification.
The rest of the chapter is organized as follows. In the next section, a brief description about entropy based landcover classification technique is provided. The three proposed algorithms are discussed in Sections 3.2, 3.3 and 3.4, respectively. Section 3.5 gives summary of the work presented in this chapter.
3.1 An entropy based landcover classification scheme
One of the successful unsupervised classification techniques based on entropy (H) and alpha angle (α) was proposed in [36]. The entropy and alpha parameters are derived from an eigenvector based decomposition of the target coherency matrix T [19]. According to
3.1 An entropy based landcover classification scheme
this method, the coherency matrix can be decomposed as follows.
T=
3
X
i=1
λiui·u∗iT, (3.1)
where λi and ui are the eigenvalues and eigenvectors of the coherency matrix, and ui’s are given by
ui =ejφi
cosαi sinαicosβiejδi sinαisinβiejγi T
, (3.2)
where α corresponds to different scattering mechanisms, β angle is twice the polarization orientation angle, φi’s are equivalent to target absolute phases and δ and γ are the phase differences of second and third terms relative to the first term.
Depending upon the scatterer types, the eigenvalues are different. Here, we consider three types of scatterers and find out the corresponding eigenvalues.
Case 1: For pure target, the coherency matrix has only one non zero eigenvalue.
Case 2: For distributed or random target, the coherency matrix has non zero and equal eigen- values.
Case 3: In between these two extremes lies the partial target. The coherency matrix of these targets has non equal and non zero eigenvalues.
From the eigenvalues and eigenvectors, two parameters, viz. entropy and alpha angle, are derived. The polarimetric entropy H, which is used to measure the target disorder is given by
H =−P1logP1−P2logP2 −P3logP3, (3.3) where P1, P2, and P3 are the pseudo-probabilities. These pseudo-probabilities are defined as
Pi = λi 3
P
j=1
λj
, for i= 1,2,3. (3.4)
Entropy values for various scatterer types are:
3. Landcover Mapping and Crop Classification Using PolSAR Images
Case 1: For pure target, entropy is equal to 0.
Case 2: For distributed target, entropy is equal to 1.
Case 3: For partial targets, entropy is in between 0 to 1.
All the three eigenvalues of the coherency matrix are roll-invariant and hence, the pseudo-probabilities are also roll-invariant. It then follows that the polarimetric entropy is also a roll-invariant parameter. This is one of the major advantages of entropy, as this property ensures that this parameter is independent of the orientation of targets about the line of sight. The drawback with entropy parameter is that if any of the eigenvalues is zero, then the corresponding Pi becomes zero and hence entropy cannot be evaluated. For that reason a small absolute value is added toPi which results in a small amount of error in the analysis of polarimetric SAR data.
Alpha angle is obtained from theαi angle of each eigenvector as follows.
α =
3
X
i=1
Piαi. (3.5)
Alpha angle denotes the average or dominant scattering mechanism. α = 00 indicates surface scattering, α= 450 indicates volume scattering, while α= 900 represents a dihedral scattering from metallic surfaces [68].
Cloude and Pottier proposed an unsupervised classification scheme based on the use of the H/α plane [36]. All random scattering mechanisms can be represented in this H/α plane. However because of spatial averaging, not all regions are equally populated. There are bounds on the maximum and minimum values of alpha angle that can be obtained for a particular value of entropy. For example, when H=1 there is only one possible value for alpha and for zero entropy, alpha can take any value in between 00 to 900. The possible variations of alpha as a function of entropy is defined by curve I and II, as shown in Figure 3.1. The curve I and II are determined by the H/αvariation for a coherency matrix with degenerate minor eigenvalues varying in between 0 to 1. The feasible region in H/α plane is bounded by these two curves. These two curves tend to the same alpha point (600)
3.1 An entropy based landcover classification scheme
for maximum entropy H = 1. From this figure, it can be seen that the possible choices of alpha reduces as entropy increases. This means the discrimination capability of this classification scheme reduces as underlying entropy increases. The H/αclassification space was sub-divided into nine basic classes of different scattering behavior, out of which eight zones lie in the feasible region. These nine zones are numbered as Z1 to Z9. Pixels in a particular zone belong to a class associated with that zone. The location of the boundaries within the permittedH and α space is set based on the general properties of the scattering mechanisms. The physical scattering characteristics of each of the nine zones are outlined in [36]. The classification map alongside the pixels distribution in H/α plane for San Francisco area are shown in Figure 3.2.
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0 10 20 30 40 50 60 70 80 90
Entropy
Alpha
Z7
Z9
Z4 Z1
Z3 Curve II
Curve I Z8
Z6
Z5 Z2
Figure 3.1: Feasible region inH/αclassification plane.
3.1.1 Wishart-H/α classification technique
TheH/αclassification technique is used to segment the image in terms of physical scattering mechanisms. However, the classification results are not satisfactory in some cases, due to
3. Landcover Mapping and Crop Classification Using PolSAR Images
the fact that not all the polarimetric information contained in the target coherency matrix has been used. Also the fixed linear decision boundaries in the H/α plane greatly affect the classification results. A cluster may fall on a boundary or may not be confined in a single zone. Besides, it may be possible that a group of clusters may be enclosed in the same zone.
Therefore to improve the classification accuracy, a combined use of unsupervised classi- fication based onH/αand supervised algorithm based on statistics of the coherency matrix has been introduced in [34]. This supervised algorithm is a maximum likelihood (ML) clas- sifier based on the complex Wishart distribution for the coherency matrix. The complex Wishart distribution of coherency matrix is given by [34]
p(n)T (hTi) = nqn|hTi|n−qexp [−nTr (V−1hTi)]
K(n, q)|V|n , (3.6)
where
K(n, q) =π(12)q(q−1)Γ (n)...Γ (n−q+ 1). (3.7) Here, q = 3 is the case for monostatic backscattering, Tr(·) and | · | denote the trace and determinant of a matrix, n is the number of looks and K is a normalization factor and V =E[hTi].
The classified pixels in each zone of H/α plane are taken as an initial training set for classification based on complex Wishart distribution. The classification results are then used as training sets for the next iteration. Improvements in classification accuracy through each iteration has been observed as the cluster centers in the H/αplane are updated after each iteration. The iteration stops when the number of pixels, switching classes, becomes smaller than a predetermined value, or when other termination criteria are met [34]. The classification results of the combined technique (Wishart-H/α) is shown in Figure 3.3(a).
The Wishart-H/α technique gives better classification results over H/α technique. Grass fields and city blocks are much clearly defined. The minute details such as polo field and golf course which were indistinguishable inH/αclassification map (Figure 3.2(b)), are more clearly visible now.