Oil Spill Detection Algorithm of a Fully Polarimetric SAR Based on Dual-EndNet

Dongmei Song; Mingyue Wang; Chengcong Hu; Jie Zhang; Bin Wang; Shanwei Liu; Dawei Wang; Bin Liu

doi:10.3788/LOP230660

Journals >Laser & Optoelectronics Progress >Volume 60 >Issue 24 >Page 2412002 > Article

Laser & Optoelectronics Progress
Vol. 60, Issue 24, 2412002 (2023)

Oil Spill Detection Algorithm of a Fully Polarimetric SAR Based on Dual-EndNet

Dongmei Song^1、2, Mingyue Wang^1、*, Chengcong Hu³, Jie Zhang^1、4, Bin Wang¹, Shanwei Liu¹, Dawei Wang¹, and Bin Liu⁵

Author Affiliations

¹College of Oceanography and Space Informatics, China University of Petroleum (East China), Qingdao 266580, Shandong, China

²Laboratory for Marine Mineral Resources, Qingdao National Laboratory for Marine Science and Technology, Qingdao 266071, Shandong, China

³China National Logging Corporation, Beijing 100101, China

⁴First Institute of Oceanology, Ministry of Natural Resources, Qingdao 266061, Shandong, China

⁵Qingdao Marine Science and Technology Center, Qingdao 266237, Shandong, China

show less

DOI: 10.3788/LOP230660 Cite this Article Set citation alerts

Dongmei Song, Mingyue Wang, Chengcong Hu, Jie Zhang, Bin Wang, Shanwei Liu, Dawei Wang, Bin Liu. Oil Spill Detection Algorithm of a Fully Polarimetric SAR Based on Dual-EndNet[J]. Laser & Optoelectronics Progress, 2023, 60(24): 2412002 Copy Citation Text

show less

Fig. 1. Overall flow of oil spill detection in fully polarimetric SAR images based on Dual-EndNet

Download full size

Fig. 2. Schematic of random forest algorithm

Download full size

Fig. 3. Experimental flowchart of polarimetric SAR oil spill detection based on Dual-EndNet

Download full size

Fig. 4. Structure of Dual-EndNet

Download full size

Fig. 5. Two PauliRGB images of polarimetric SAR oil spill

Download full size

Fig. 6. Oil spill detection results on dataset 1. (a) PauliGRB; (b) GroundTruth; (c) P-SVM; (d) F10-SVM; (e) F30-SVM; (f) FP-SVM; (g) P-CNN; (h) F10-CNN; (i) F30-CNN; (j) FP-CNN; (k)P-UNet; (l) F10-UNet; (m) F30-UNet; (n) FP-UNet; (o) P-FCN; (p) F10-FCN; (q) F30-FCN; (r) FP-FCN; (s) P-PSPNet; (t) F10-PSPNet; (u) F30-PSPNet; (v) FP-PSPNet; (w) P-Deeplabv3; (x) F10-Deeplabv3; (y) F30-Deeplabv3; (z) FP-Deeplabv3; (ab) Dual-EndNet

Download full size

Fig. 7. Oil spill detection results on dataset 2. (a) PauliGRB; (b) GroundTruth; (c) P-SVM; (d) F10-SVM; (e) F30-SVM; (f) FP-SVM; (g) P-CNN; (h) F10-CNN; (i) F30-CNN; (j) FP-CNN; (k) P-UNet; (l) F10-UNet; (m) F30-UNet; (n) FP-UNet; (o) P-FCN; (p) F10-FCN; (q) F30-FCN; (r) FP-FCN; (s) P-PSPNet; (t) F10-PSPNet; (u) F30-PSPNet; (v) FP-PSPNet; (w) P-Deeplabv3; (x) F10-Deeplabv3; (y) F30-Deeplabv3; (z) FP-Deeplabv3; (ab) Dual-EndNet

Download full size

No.	Polarimetric feature	Equation	Explanation
01	SPAN	$V_{S P A N} = {\|S_{H H}\|}^{2} + 2 {\|S_{H V}\|}^{2} + {\|S_{V V}\|}^{2}$
02	Geometric intensity	$V = {[d e t (T)]}^{1 / 3}$	$T$ is polarization coherence matrix
03	VV intensity	$V_{V V} = {\|S_{V V}\|}^{2}$
04	Copolarization phase difference	$σ_{ϕ C O} = \sqrt[]{〈{(φ_{H H} - φ_{V V})}^{2}〉 - {(〈φ_{H H} - φ_{V V}〉)}^{2}}$	$φ$ is the phase information
05	Copolarization power ratio	$γ_{C O} = 〈{\|S_{H H}\|}^{2}〉 / 〈{\|S_{V V}\|}^{2}〉$
06	Copolarization correlation coefficient	$ρ_{C O} = \|\frac{〈S_{H H} S_{V V}^{*}〉}{\sqrt[]{〈{\|S_{H H}\|}^{2}〉〈{\|S_{V V}\|}^{2}〉}}\|$
07	Real part of the copolarization cross product	$r_{C O} = \|R (〈S_{H H} S_{V V}^{*}〉)\|$
08	Muller polarization feature $M_{33}$	—
09	Consistency coefficient	$μ = \frac{2 [R e (S_{H H} S_{V V}^{*}) - {\|S_{H V}\|}^{2}]}{{\|S_{H H}\|}^{2} + 2 {\|S_{H V}\|}^{2} + {\|S_{V V}\|}^{2}}$
10	Polarization scattering entropy	$H = - \sum_{i = 1}^{3} p_{i} l o g_{3} p_{i}$	$p_{i}$ = $\frac{λ_{i}}{λ_{1} + λ_{2} + λ_{3}}$
11	Anisotropy $A$	$A = \frac{λ_{2} - λ_{3}}{λ_{2} + λ_{3}}$	λ_i is eigenvalue calculated from the polarimetric coherence matrix
12	Average scattering angle	$\bar{α} = \sum_{i = 1}^{3} p_{i} α_{i}$
13	Anisotropy $A_{12}$	$A_{12} = \frac{λ_{1} - λ_{2}}{λ_{1} + λ_{2}}$
14	Maximum eigenvalue	$λ_{m a x} = m a x (λ_{1}, λ_{2}, λ_{3})$
15	Pedestal height	$V_{P H} = \frac{m i n (λ_{1}, λ_{2}, λ_{3})}{m a x (λ_{1}, λ_{2}, λ_{3})}$
16	Averaged intensity	$I = λ_{1} \times p_{1} + λ_{2} \times p_{2} + λ_{3} \times p_{3}$
17	SERD	$V_{S E R D} = \frac{λ_{s} - λ_{3 n o s}}{λ_{s} + λ_{3 n o s}}$	$λ_{3 n o s} = 2 〈{\|S_{H V}\|}^{2}〉$
18	Polarization feature P	$P = \frac{〈{\|S_{H H} + S_{V V}\|}^{2}〉}{〈{\|S_{H H} - S_{V V}\|}^{2}〉}$
19	Bragg scattering energy proportion	$η = \frac{P_{B r a g g}}{V_{S P A N}} = \frac{T_{11} + {\|T_{12}\|}^{2} / T_{11}}{V_{S P A N}}$
20	Self-similarity parameter	$V_{r r} = \sum_{i = 1}^{3} λ_{i}^{2} / {(\sum_{i = 1}^{3} λ_{i})}^{2} = t r (T T^{H}) / {[t r (T)]}^{2}$
21	Scattering diversity	$V_{S D} = \frac{3}{2} (1 - {‖N‖}_{F}^{2}) = \frac{3}{2} [1 - {‖T / t r a c e (T)‖}_{F}^{2}]$
22	Surface scattering fraction	$N_{11} = \frac{{(S_{H H} + S_{V V})}^{2}}{V_{S P A N}}$
23	Combined feature parameter F	$F = [ρ_{C O} + \bar{H} + \bar{α} + A_{12}] / 4$
24	Combined polarimetric feature H_A₁₂	$H_A_{12} = H (1 - A_{12})$
25	Combined polarimetric feature H_A	$H_A = (1 - A) (1 - H)$
26	Polarimetric feature $V_{C T}$	$V_{C T} = C_{13} / \sqrt[]{T_{11} \times T_{33}}$
27	Coherence coefficient	$c_{h o} = \|〈T_{12}〉\| / \sqrt[]{〈T_{11}〉〈T_{22}〉}$
28	Cross-polarization ratio	$C = S_{H H} / S_{H V}$
29	Degree of polarization	$V_{D o P} = \sqrt[]{\frac{1}{3} [T r (M^{T} M) / M_{11}^{2} - 1]}$
30	Gini coefficient	$P_{g i n i} = 1 - \sum_{i = 1}^{3} p_{i}^{2}$

Table 1. Summary of the commonly used polarimetric features

Stage	Layer	Kernel size	Stride	Number
Encoder	Conv1	3×3	1	64
	Conv2	3×3	1	64
	Pooling1	2×2	2
	Conv3	3×3	1	128
	Conv4	3×3	1	128
	Pooling2	2×2	2
	Conv5	3×3	1	256
	Conv6	3×3	1	128
Decoder	Upsampling1	2×2
	Conv7	3×3	1	128
	Conv8	3×3	1	64
	Upsampling2	2×2
	Conv9	3×3	1	64
	Conv10	3×3	1	32
Feature fusion stage	Conv11	3×3	1	64
	Conv12	3×3	1	64
	Conv13	1×1	1	Class number
	Softmax

Table 2. Detailed parameters of Dual-EndNet

Parameter	Dataset 1	Dataset 2
Satellite	Radarsat-2	Radarsat-2
Product type	SLC	SLC
Band	C	C
Polarimetric	Quad-pol	Quad-pol
Spatial resolution	4.7 m×4.8 m	4.7 m×4.8 m

Table 3. Detailed imaging parameters of two-view Radarsat-2 images

Polarimetric feature	Score of importance
Maximum eigenvalue	8.0094
Averaged intensity	7.7246
Real part of the copolarization cross product	5.7811
SPAN	5.4079
VV intensity	4.6914
Geometric intensity	3.7980
Surface scattering fraction	3.0953
Polarization scattering entropy	2.4766
Gini coefficient	2.0543
Polarimetric feature H_A₁₂	1.5344
Degree of polarization	1.4882
Anisotropy A₁₂	1.0643
Combined feature parameter F	0.8130
Consistency coefficient	0.4918
Cross-polarization ratio	0.4261
Self-similarity parameter	0.3293
Scattering diversity	0.3127
Bragg scattering energy proportion	0.2295
Copolarization correlation coefficient	0.1179
Polarization feature P	0.0575
Pedestal Height	0.0382
SERD	0.0297
Polarimetric feature CT	0.0146
Muller polarization feature M₃₃	0.0067
Copolarization power ratio	0.0022
Average scattering angle	0.0018
Coherence coefficient	0.0013
Combined polarimetric feature H_A	0.0009
Anisotropy	0.0009
Copolarization phase difference	0.0005

Table 4. Ranking of feature importance of random forest algorithm on dataset 1

Method	Accuracy /%		OA /%	AA /%	Kappa	F1-score	MIoU
Method	Oil spill	Sea water	OA /%	AA /%	Kappa	F1-score	MIoU
P-SVM	92.08	91.11	91.60	91.59	0.8319	0.9159	0.8449
F10-SVM	92.71	93.67	93.18	93.19	0.8635	0.9318	0.8722
F30-SVM	93.34	92.45	92.90	92.89	0.8579	0.9289	0.8673
FP-SVM	92.70	94.51	93.57	93.61	0.8714	0.9357	0.8792
P-CNN	92.50	96.06	94.18	94.28	0.8836	0.9418	0.8899
F10-CNN	93.41	96.66	94.95	95.03	0.8990	0.9495	0.9038
F30-CNN	95.42	92.53	93.95	93.97	0.8790	0.9395	0.8859
FP-CNN	93.76	93.67	93.72	93.72	0.8743	0.9372	0.8817
P-UNet	93.70	94.08	93.89	93.89	0.8777	0.9389	0.8848
F10-UNet	93.51	96.62	94.99	95.07	0.8344	0.9499	0.9045
F30-UNet	93.29	94.96	94.10	94.13	0.8819	0.9410	0.8885
FP-UNet	93.66	96.46	95.00	95.06	0.8999	0.9500	0.9047
P-FCN	92.35	95.21	93.71	93.78	0.8742	0.9371	0.8816
F10-FCN	95.28	96.32	95.79	95.80	0.9157	0.9579	0.9191
F30-FCN	93.96	90.53	92.21	92.25	0.8442	0.9220	0.8554
FP-FCN	94.31	96.36	95.30	95.34	0.9059	0.9530	0.9101
P-PSPNet	93.61	94.98	94.29	94.29	0.8857	0.9429	0.8919
F10-PSPNet	94.26	96.52	95.38	95.39	0.9075	0.9538	0.9116
F30-PSPNet	93.69	95.43	94.55	94.56	0.8910	0.9455	0.8966
FP-PSPNet	94.50	96.60	95.54	95.55	0.9107	0.9554	0.9145
P-Deeplabv3	93.70	94.99	94.34	94.35	0.8867	0.9434	0.8928
F10-Deeplabv3	94.31	96.63	95.54	95.47	0.9091	0.9545	0.9131
F30-Deeplabv3	93.72	95.52	94.61	94.62	0.8922	0.9461	0.8977
FP-Deeplabv3	94.60	96.68	95.63	95.64	0.9125	0.9563	0.9162
Dual-EndNet	95.78	96.98	96.36	96.38	0.9272	0.9636	0.9297

Table 5. Comparison of oil spill detection accuracy of different algorithms on dataset 1

Polarimetric feature	Score of importance	Polarimetric feature	Score of importance
Maximum eigenvalue	6.1732	Consistency coefficient	1.0919
Averaged intensity	5.4363	Muller polarization feature M₃₃	1.0002
VV intensity	3.7743	Average scattering angle	0.9672
SERD	2.8461	Combined feature parameter F	0.9487
Surface scattering fraction	2.4692	Copolarization phase difference	0.8663
Real part of the copolarization cross product	2.3569	Anisotropy A	0.8624
Polarization scattering entropy	2.2217	Polarimetric feature CT	0.8191
Geometric intensity	2.1696	Cross-polarization ratio	0.7499
Gini coefficient	2.1405	Coherence coefficient	0.7151
Pedestal height	1.9866	Combined polarimetric feature H_A	0.6912
Combined polarimetric feature H_A₁₂	1.6533	Bragg scattering energy proportion	0.5881
Degree of polarization	1.5632	Scattering diversity	0.5822
SPAN	1.4700	Self-similarity parameter	0.5810
Copolarization power patio	1.1308	Polarization feature P	0.5154
Anisotropy A₁₂	1.1185	Copolarization correlation coefficient	0.5111

Table 6. Ranking of feature importance by random forest algorithm on dataset 2

Method	Accuracy /%				OA /%	AA /%	Kappa	F1-score	MIoU
Method	Sea water	Mineral	Emulsion	Bio-oil	OA /%	AA /%	Kappa	F1-score	MIoU
P-SVM	91.09	60.38	0.00	68.83	89.00	55.08	0.3989	0.4321	0.3617
F10-SVM	93.63	44.66	36.68	86.99	91.39	65.49	0.4657	0.5149	0.4120
F30-SVM	92.91	42.90	34.68	87.13	90.64	64.41	0.4405	0.5004	0.3998
FP-SVM	92.79	43.42	36.34	87.47	90.56	65.00	0.4389	0.5066	0.4044
P-CNN	97.25	93.40	79.43	64.21	96.56	83.57	0.7352	0.7152	0.6016
F10-CNN	97.64	86.03	79.37	83.88	96.92	86.73	0.7562	0.7294	0.6189
F30-CNN	96.93	93.15	64.62	89.77	96.45	86.12	0.7328	0.7080	0.6060
FP-CNN	96.94	92.03	83.77	88.05	96.55	90.20	0.7406	0.7288	0.6212
P-UNet	98.43	86.47	58.43	67.78	97.30	77.78	0.7727	0.7002	0.5954
F10-UNet	98.47	83.22	73.75	77.77	97.48	83.30	0.7867	0.7867	0.6274
F30-UNet	98.69	87.70	72.83	81.13	97.88	85.09	0.8166	0.7693	0.6645
FP-UNet	98.55	89.45	84.89	82.56	97.92	88.86	0.8226	0.7914	0.6869
P-FCN	98.91	86.30	72.10	70.15	97.89	81.87	0.8148	0.7543	0.6443
F10-FCN	98.66	90.98	88.78	85.54	98.15	90.99	0.8411	0.8164	0.7149
F30-FCN	98.06	92.15	83.14	90.27	97.63	90.90	0.8060	0.7872	0.6818
FP-FCN	98.26	88.26	84.43	90.19	97.70	90.28	0.8090	0.7834	0.6766
P-PSPNet	98.50	89.38	89.69	85.50	97.95	90.77	0.8262	0.8012	0.6976
F10-PSPNet	98.42	88.50	86.37	79.68	97.74	88.24	0.8091	0.7798	0.6727
F30-PSPNet	98.70	91.06	88.90	85.56	98.19	91.06	0.8443	0.8118	0.7113
FP-PSPNet	98.72	91.17	89.20	86.12	98.22	91.30	0.8471	0.8181	0.7198
P-Deeplabv3	98.62	89.78	82.32	80.72	97.96	87.86	0.8247	0.7888	0.6848
F10-Deeplabv3	98.65	87.72	76.07	83.21	97.89	86.41	0.8186	0.7793	0.6733
F30-Deeplabv3	98.79	91.25	89.70	85.62	98.29	91.34	0.8521	0.8218	0.7245
FP-Deeplabv3	98.87	91.26	89.71	86.68	98.38	91.63	0.8590	0.8285	0.7329
Dual-EndNet	99.04	94.32	95.71	92.28	98.76	95.34	0.8913	0.8788	0.7952

Table 7. Comparison of oil spill detection accuracy of different methods on dataset 2

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information