Deep neural network for oil spill detection using Sentinel-1 data: application to Egyptian coastal regions

Figures & data

Figure 1. Study region in the Northern Shore of Egypt.

Table 1. A summary of dates, areas, and locations of the raw SAR images used for model training and testing.

Year	Month	Day	Area	Region	Year	Month	Day	Area	Region
2015	4	2 *	4.01	Port Said	2017	4	22	17.12	Port Said
		19	24.15				27	12.09
		26 *	11.33			5	2	12.0	Rahseed
	7	19 *	6.22				3	19.9	Port Said
2016	8	6	17.6				9	79.3
2017	1	3	7.83				21	39.8
		21	4.24				26	28.2
		16 *	8.22	Baltim		6	1	22.5	Damietta
		21	25.2				2	7.83	Port Said
	2	2	13.01	Port Said			8	94.0
		6 *	6.95	Alexandria		7	20	21.2
		7	1.11	Port Said			26	12.3	Baltim
		8	2.51			10	1	3.71	Port Said
		9	5.14			12	30	29.87
		19	58.69	Rahseed	2018	3	5	9.71
		20	2.67	Port Said			17	3.36
		26	2.29				23	3.43	Rahseed
	3	4	12.35			5	22	7.95	Port Said
		10	5.91			9	20	1.8
		29	10.89		2019	3	18	4.61
	4	3	26.19			4	5	4.68
		8	26.19				11	2.16
		15	10.05	Baltim		5	21	4.64	Damietta

*Dataset reserved for model testing and was not used in training

Figure 2. The U-Net architecture used for oil spills detection.

Figure 3. The Seg-Net architecture used for oil spills detection.

Figure 4. The discriminator architecture used for oil spills detection.

Table 2. Average evaluation parameters after 300 epochs.

Evaluation indices	SSIM	PSNR	RMSE	Accuracy	Precision	Recall	F1-Score	IoU
U-net	81.119	64.644	0.0245	92.386	34.391	36.737	34.286	88.12
Seg-Net	41.208	56.648	0.0378	98.463	32.207	38.674	34.069	61.16
CGAN	90.593	73.434	0.0272	83.591	33.114	34.348	32.950	96.38
Seg-Net-based GAN	97.747	125.898	0.0187	99.043	32.629	32.573	30.485	96.59

Figure 5. Detection results of the four detection models (U-Net, Seg-NET, CGAN, Seg-Net-based CGAN) on different sizes and shapes of oil spill patches from the testing set.

Figure 6. The Proposed Methodology for Oil Spills Detection from raw SAR images using a Seg-Net-based CGAN.

Figure 7. The average detection performance results during training, testing, and validating Seg-Net-based CGAN model using unseen images.

Figure 8. The performance of the proposed Seg-Net-based CGAN model in detecting oil spills during the testing phase.

Table 3. The average of evaluation for patches with oil spills only during train, test, and validation of Seg-Net-based CGAN.

Evaluation parameters %	precision	recall	f1-score
Train	87.81	79.12	76.85
Validation	85.24	84.97	79.54
Test	73.83	78.46	71.73

Figure 9. (a) Application region in the Arabian Gulf, (b) location of oil spill at 10 August 2017 overlayed by Sentinel-1GRD (VV) image, (c) Sentinel-1 GRD (VV) image showing the areas of oil spills and lookalikes, (d) results of oil spills segmentation using the proposed model.

Data availability statement

The dataset used in this analysis will be available to the public via https://drive.google.com/file/d/15WYzzFZvAHmqSIW0PXXRTp_YVd_868l8/view?usp=sharing