The local winds over the Adriatic Sea are strongly variable in space and time due to the interaction with islands, mountains and coastlines. In particular, at the Croatian side the winds can be quite strong (above 20 m/s) and variable (especially bora). After bora, sirocco (jugo) is the second most important local wind that can reach hurricane strength and affect the coastal environment. The name sirocco is given to hot and dry southeast local winds originating over the desert of North Africa, penetrating into the Mediterranean basin. Sirocco occurs in advance of atmospheric depressions (cyclones) moving eastward across the Mediterranean Sea or the Northern Africa. Sirocco events tend to occur during all year however, strong events are most common during the spring and autumn time. The average duration of continuous gale force winds during a sirocco is 10 to 24 hrs and occasionally up to several days and more. Unlike bora having jet-like structure and blowing across the sea, the sirocco wind blows along the eastern Croatian coast and produces more uniform signatures on the sea surface. Here we have analyzed several sirocco wind events, which signatures are frequently visible on the Sentinel-1A and Sentinel-1B synthetic aperture radar (SAR) images of the sea surface acquired in 2016−2018. The results of NCEP modeling have been used to verify SAR data. As a result, it is shown that the sea surface roughness patterns associated to these wind events captured by SAR yield information on the spatial extent and about the fine-scale structure of the sirocco wind field, which cannot be obtained by other remote sensing instruments. In particular, SAR is capable of resolving: 1) general wind direction, 2) island wakes and micro-jets generated by the interaction of south-easterly winds with the island coastal topography, and 3) atmospheric gravity (lee) waves. However, all these phenomena can be seen due to the interaction of the airflow with numerous islands in the Central Adriatic, and they actually visualize the sirocco. Quantitative information on the near surface wind field is derived from the SAR images by using the wind scatterometer model for converting radar backscatter values into wind speeds. These high resolution SAR data and fine resolution wind fields extracted from SAR images can be used to validate and improve meso-scale atmospheric models. It is finally concluded that SAR imagery can give very useful additional information about such a phenomenon as sirocco.

Area D - Board 418

A new automated algorithm for detecting offshore oil seepage is presented, exploiting the synergy between satellite-based Synthetic Aperture Radar (SAR) and optical datasets. This method has the potential to detect oil seepage and estimate liquid hydrocarbon release in the ocean with high accuracy and celerity, while also creating a viable tool for continuous monitoring of the seepage phenomenon in the offshore environment.

Naturally-occurring hydrocarbon seeps are found at many sites around the world, both on-shore and offshore. To reach the land or ocean surface, hydrocarbons migrate along geological layers or through fractures and fissures, which may be tectonic (e.g. folding, faulting) or erosional in origin. Oil seepage is considered a reliable exploration tool for characterisation of the potential of uninvestigated areas. In addition, hydrocarbon seepage may also contribute to the overall carbon cycle as well as to the impact on local ecosystems.

Identification and characterisation of oil seeps on land has been carried out since the early days of hydrocarbon exploration. Offshore seepage slick mapping has been of interest to the oil and gas industry since SAR technology was declassified. However, the reliable interpretation of sea surface features as seepage slicks and the relation of the observed slicks to characteristics of the sub-surface hydrocarbon reserves is a highly specialized task. Human analysis of currently available long time series of satellite acquisitions may not be a viable proposition and therefore automated techniques represent a cost effective alternative if they can be demonstrated to be sufficiently reliable.

Although semi-automatic oil slick detection algorithms have been developed, they still require a substantial visual inspection component. In the present work an algorithm based on an existing automated method is refined to utilise primarily Copernicus Sentinel-1 SAR and Sentinel-2 multispectral data, along with other complementary high and very high resolution datasets. The intention is to develop capabilities to automatically distinguish between seepage slicks, slicks due illegal discharges at sea, biogenic slicks, ocean and atmospheric phenomena. A time-series of SAR imagery is processed and possible natural slicks are extracted based on a set of rules designed to account for the geometry of the slick, their physical behaviour, texture parameters and the proximity to other elements on the surface of the sea. Similarly, various time coincident and collocated datasets of multispectral scenes are being used, focusing on the exploitation of sun-glint for extracting oil slicks, estimating their thickness. We will present the results obtained to date for a test area in the Black Sea. The algorithm is enhanced with contextual information such as SAR wind data and chlorophyll content estimation for ruling out look-alikes. Ultimately, the processing architecture is designed to ensure a highly adaptable structure that can accommodate the use of new datasets and methods (e.g. machine learning) to reduce the false positive rates.

Area D - Board 419

In our days one of the main environmental concerns is the pollution of the sea by oil spills because it can provoke major ecological disasters as it was dramatically proved by numerous accidents like: the Exxon Valdez accident in Alaska in 1989, the Prestige oil spill at the coast of Galicia, Spain in 2002 or the Deepwater Horizon platform oil spill accident in 2010 in the Gulf of Mexico. Large spills of oil related petroleum products in the marine environment cause serious biological and economic impacts. Remote sensing can play an increasingly important role in oil monitoring. Synthetic Aperture Radar SAR satellites are usually used to monitor the sea oil spill as they are more effective compared to the optical sensors as their signal can penetrate clouds, and it is in general less affective by the weather. Mapping and identification of the exact geometry of the spill constitutes a major importance’s step in the process of quantifying its volume and creating a spread model, while calculating other parameters crucial for the planning of an efficient response might be quite challenging. However, the presence of areas known as look-alikes which are caused by natural marine processes (algae bloom, threshold winds, e.t.c.) as well as the inherent speckle noise of SAR images makes the SAR processing results ambiguous. The final result of the oil spill detection procedure is based on the creation of a mask on areas classified from low to high levels of confidence. The launch of the first Sentinel-1 mission in April 2014 gave a new impulse to the specific research domain. Sentinel-1 mission provides continuous SAR images with a repeat time of 6 days and high spatial resolution, images that are freely available through ESA. Using the complementary SNAP software, satellite imagery processing has never been more accessible than it is today. Recently an automatic oil spill detection tool was integrated in ESA’s SNAP, in which an adaptive threshold value identifies dark spots and then clusters and discriminates them based on the dimension of the cluster and minimum cluster size. Furthermore, oil spill mapping can also be achieved without the use of the aforementioned tool by comparing two images, before and after the spill and applying a band math expression. On September 10th 2017 the Agia Zoni II tanker wrecked southwest of the island of Salamina at 02:30 GMT and almost all the transported petroleum products spilled out onto the water. The key point of this study is the mapping of the Agia Zoni II oil spill with the greatest possible precision using Sentinel-1 imagery and to evaluate the efficacy of the oil spill tool compared to a semi-automatic approach in terms of the final products quality but also considering the time required for each approach. Additionally, the validation of the results of the oil spill detection procedure was based on the digitization of the affected areas utilizing Sentinel-2 data. In general, pollution caused by oil spills, whether those come from tanker ships or offshore platforms pose a severe threat to the environment as they can put the entire ecosystem at risk and provokes serious economic damages. Understanding the important implications of an oil spill accident, to the environment and general public alike, can help in comprehension of the relevance and usefulness of such studies.

Area D - Board 420

Classicaly marine oil seepages outbreaking at the sea surface (Sea Surface Outbreak-SSO) are considered as good oil potential indicators within any marine geological basins on the Earth. Effectively for more than two decades oil seeps flower structures extracted from multitemporal SAR images had been taken into account as a major key for Oil and Gaz (O&G) exploration. But many detailed petroleum studies deny the presence of highly interesting productive O&G reservoirs in their immediate vicinities. Consequently, O&G applied studies focused on the potential hydrocarbon pathways and percolations at depth in order to locate the potential original O&G reservoirs (issued from the primary O&G migration). But they reveal only few O&G highly productive reservoirs in such marine oil seepages surroundings. So what SSO oil seeps detection are then revealing ?
Following previous work (e.g.: MacDonald et al., 2004, Ding et al., 2010) that identify the presence of oil seepages and associated asphalts deposits at depth in Campeche Knolls (S. GoM), we perform herein a new stochastic method in order to get both the SSO and Sea Floor Sources (SFS) (Najoui et al., 2018A and B). So, if we confirm partly the previous published studies, following our topographic and morphostructural study, we newly reveal that Campeche Knolls corresponds to a huge deep submarine landslide (length from toe to crown: 260km* max. lateral extension: 400km) affecting both Banco de Campeche and Bay of Campeche continental plateaus. Its toe overrides the "Veracruz Tongue". This thick landslide, made of continental plateau materials, slide above the thick inferred late Jurassic in age salt ductile decollement layer. From many seismic reflexion profiles sedimentological truncations arguments, this major landslide should have been initiated in late Miocene time (see Bertagne, 1984) and have been regularily re-activated since that time (work in progress).
So we newly interprete marine oil seepages as excellent witnesses of marine sliding re-activation processes. We believe they evidence slow sliding of the margins as well as local landslide processes above ductile layers (salty for instance) situated above the thinned continental crust (see the S. GoM). Moreover, we find contrasting tectonic environment for oil seepages sea floor emission: the toe of the landslide where the higher oil seepage are due to the downstream gravitationnal compressive sliding which contrast to the low oil seepages emissions upstream which prevail a normal gravitationnal tectonic regime. Those active sliding reactivations processes re-open fractures (SFS, pockmarks) that expulse oil seeps to the sea surface (SSO). Finaly, we propose to re-interprete the marine oil seepages as excellent markers of the marine sliding and landsliding re-activations of any enriched in organic matter marine passive/active margins.
1. Zhour Najoui, Serge Riazanoff, Benoît Deffontaines, Jean-Paul Xavier, 2018A. A statistical approach to preprocess and enhance C-Band SAR images in order to detect automatically marine oil slicks, IEEE Transactions on Geoscience and Remote Sensing, Vol: 56 Issue: 5 print ISSN: 0196-2892, Online ISSN: 1558-0644, DOI: 10.1109/TGRS.2017.2760516
2. Zhour Najoui, Serge Riazanoff, Benoît Deffontaines, Jean-Paul Xavier, 2018B. Estimated location of the SeaFloor Sources of marine natural oil seeps from Sea Surface Outbreaks : Inputs of a new vertical drift model applied to the northern Gulf of Mexico, Marine Petroleum Geology. Vol.: 91, pp.: 190-201. see https://authors.elsevier.com/a/1WNXAyDcGAlf8, https://doi.org/10.1016/j.marpetgeo.2017.12.035.

Area D - Board 421

Satellite remote sensing for environmental monitoring is developing so fast in the last years, both for land and water resources. Thanks to the availability of data from various sensors with different spatial, temporal, radiometric and spectral resolutions, remote sensing data has become an essential source for a wide range of applications.
In the field of the maritime environment surveillance, the detection and mapping of oil on the water surface is one of the common use of the satellite images.
To respond to an oil spill effectively, response operations require accurate and timely information on the location, the quantity and characteristics of the oil spilled. In this perspective, remote sensing data can support the initial detection and assessment of the oil spilled and can support the synoptic monitoring of the phenomenon.
Capable of operating at night and day, in all weather conditions, Radar data is ideal for seas monitoring.
SAR image is a measure of surface roughness depending on the backscatter. Within SAR scene, oil slicks appear as relatively dark patches due to the dampening effect of oil on water. Therefore, affected area can be detected in correspondence of low values of backscatter.
To date numerous SAR satellite are available, this study has been focused on the use of Sentinel 1 data provided by the Copernicus programme.
The case study was offered by the ships collision occurred early in the morning of the 7th October 2018 off the coast of Corsica, in the Mediterranean sea. The Tunisian freighter Ulysse rammed into a Cyprus-registered container ship CSL Virginia anchored about 20 miles off the northern tip of Corsica causing a heavy fuel oil spill. The collision occurred in French waters but the clean-up operation was part of a joint pact among France, Italy and Monaco, to reduce as much as possible the pollution in Mediterranean sea. The spill was estimated to be around 600 m3 of heavy fuel oil.
The aim of this study has been the detection and the extent monitoring of the fuel spill.
A near real-time procedure has been applied to process the Sentinel 1 images acquired on the affected area with a good overpass after the accident. For this purpose, a semi-automatic, object-based classification approach was developed. Each scene was processed with the same ruleset: firstly, the image has been segmented, than a threshold classification based on backscatter values was applied to identify the image objects representing the fuel slicks and the probably fuel slicks.
The presented study highlighted the effectiveness of using of Sentinel 1 images for sea monitoring applications, also for near real time application thanks to the development of semiautomatic classification procedures.

Area D - Board 422

The detection of non-ocean scatterers over the sea surface by using pulse-limited satellite altimeters involves a series of challenging targets, such as icebergs, lighthouses and ships, which have been investigated in the literature. In particular, past works focused on the hyperbolic features observed in the thermal noise area of the received waveforms, in order to detect the presence of such targets. The current availability of SAR-altimetry data offers a new opportunity to explore target-detection capabilities as a byproduct of radar altimetry missions. Following previous analysis made on CryoSat-2 waveforms (in SAR mode), this work makes a first analysis of the capabilities of Sentinel-3A SRAL delay-Doppler instrument for the detection of ships. In particular, we analyse the shape of the waveforms at the finest along-track sampling rate: 80 Hz (corresponding to an along-track distance between two consecutive measurements of about 85 m). In this work, we study the compatibility between the detected target(s) and the known ship traffic, by using the Automatic Identification System (AIS) data. Global ship traffic statistics may take benefit from the method described here, with its contribution to the identification of global traffic patterns and their temporal evolution by using free datasets, which are also particularly limited in size. We focused on two Sentinel-3A tracks (one ascending and one descending) from June 2016 to July 2018. These tracks cross the western aperture (Atlantic side) of the Strait of Gibraltar at 21:59 UTC (relative orbit #057) and at 10:49 UTC (relative orbit #322) every 27 days. In this particular zone high density ship traffic from AIS data is observed. The Sentinel-3 mission provides a constellation with global coverage (S3A and the new S3B) and free accessibility to the data, with a potential enhancement to the estimation of the number and characteristics of the ships with respect to past literature approaches still based on altimetry missions.

Area D - Board 423

Maritime surveillance is surely a topic of growing interest and can be regarded as key for its strategic importance and potential benefits for the future of the EU. Indeed, it is estimated that almost 90% of the external freight trade is seaborne, short sea shipping represent 40% of the intra-EU exchanges, and more than 400 million passengers embark and disembark in European ports each year [1].
In this scenario, Synthetic Aperture Radar (SAR) sensors are one of the most effective means thanks to their capability to get images independently from daylight and weather conditions. SAR sensors are considered a valid alternative to the traditional coastal-based surveillance systems (such as the Automatic Identification System, AIS) and are particularly suitable for the detection and the tracking of ships and vessels in open sea. In the SAR ship-detection field, many algorithms have been presented in literature for single and multi-polarization channels [2]. Traditionally, SAR ship-detection algorithms rely on Constant False Alarm Rate (CFAR) methods in which the sea clutter is modelled according to a distribution function and a threshold is then computed to achieve a desired probability of false alarm [2]. All the clusters of pixels with intensity greater than the set threshold are considered as potential ships. As a consequence, detectors are strongly influenced by thresholding techniques applied to the SAR amplitude or intensity image with two main negative results in the performance: 1) a high false alarm rate and 2) the inability to detect ships with Radar Cross Section (RCS) values in the order of the RCS of the surrounding sea clutter. A novel ship detector for SAR imagery based on the Generalized Likelihood Ratio Test (GLRT) has been already presented by the authors [3, 4]. It has been demonstrated that this detector overperforms the CFAR in terms of detection and false alarm rates in areas of heavy clutter (strong RCS relative to the sea background) where it is able to detect small vessels and boats.
The main objective of this paper is to present a new automated version of the GLRT ship detector for operational purposes. This new version does not rely anymore on any users’s intervention. At this aim, the clutter and the target parameters are evaluated and computed in moving windows over the SAR image and, as a consequence, the threshold test changes over the SAR image with a higher computational load than the supervised version of the GLRT detector. In particular, clutter and target parameters are computed over three moving windows: target, guard and background window as explained in [1].
In order to reduce the computational load, a two-step algorithm is implemented for the detector. In the first step, the clutter is assumed to be Gaussian distributed and a Cell Averaging (CA) CFAR is performed by choosing a high probability of false alarm to reject only obvious non target pixels. In the second step, the GLRT algorithm is applied to the pixels detected in the fisrt step. In particular, the distribution function of the sea background is here Gamma distributed in the intensity domain for homogenous clutter; on the other side, the distribution function of the target is derived from electromagnetic models within Geometric Optic (GO) and Kirchhoff Approximation (KA). Within these hypotheses, it has been proved that the target is Gamma distributed in the co-polarized (HH and VV) channels and Weibull distributed in the cross-polarized (HV and VH) channels at S-, C- and X- bands [5]. As for the CFAR algorithms, the GLRT is applied by using a moving window and computing the clutter and target parameters along with the threshold at each iteration (Adaptive Threshold algorithm).
The new GLRT algorithm is tested on SAR images acquired by the German TerraSAR-X and the European Sentinel-1 constellation operating at X- and C-band, respectively. Meaningful datasets are acquired over the coast of Japan, Australia and Panama for both constallations of sensors. In particular, Strimap, ScanSAR and Wide ScanSAR images are acquired by TerraSAR-X and Interferomteric Wide Swath by Sentinel-1. The new GLRT is tested and compared against the CA-CFAR algorithm and the previous supervised version of the GLRT itself. Outcomes will be shown at conference time in terms of detection rate, probability of false alarm, computational load and Receiver Operating Characteristc (ROC) curves.

References
[1] European Commission, “Mobility and Trasport”, Technical Report, https://ec.europa.eu/transport/modes/maritime_en (Accesed on 06/11/2018).
[2] D.J.Crisp, “The State-of-Art in Ship Detection in Synthetic Aperture Radar Imagery”, DSTO Information Science Laboratory, 2004.
[3] P. Iervolino and R. Guida, "A Novel Ship Detector Based on the Generalized-Likelihood Ratio Test for SAR Imagery," IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 10, no. 8, pp. 3616-3630, Aug. 2017.
[4] P. Iervolino et al., "Ship detection in SAR imagery: A comparison study," Proc. IGARSS, Fort Worth, TX, Jul. 2017, pp. 2050-2053.
[5] P. Iervolino, R. Guida, and P. Whittaker, “A model for the backscattering from a canonical ship in SAR imagery”, IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., vol. 9, no. 3, pp. 1163–1175, March 2016.

Area D - Board 424

The control of open seas areas or marine protected areas is usually performed based on the Automated Identification System (AIS) embarked within most of the vessels. The objective of the present document is to present a vessel detection tool for marine area surveillance based on Synthetic Aperture Radar (SAR) data. By comparing the detected targets with AIS data, this tool can determine the ratio of non-cooperative vessels (or not equipped with AIS) within any area.
The Sentinel-1 sensors offer wide coverage at medium to high resolution. The interferometric wide swath (IW) mode implemented with the terrain observation with progressive scans SAR (TOPSAR) technique is the standard acquisition mode over European waters and land mass.
The range resolution of SAR images is a function of the emitted radar signal bandwidth. Most recent SAR sensors use wide band signals in order to achieve metric range resolution. In IW mode, the Chirp bandwidth of the Sentinel-1 sensors is about 50 MHz. By comparison, the ENVISAR and ERS sensors used Chirp bandwidth of 15 MHz.
Instead of providing images of higher range resolution, we exploit wide band to split a single acquisition into sub-bands and generate several lower-resolution images, centered on slightly different frequencies, performing so a SAR spectral analysis [1]. One of the split-band applications is the spectral coherence consisting in computing the interferometric coherence between sub-images issued from a single acquisition.
It was shown that for surface scatterers having a random distribution, the spectral coherence is proportional to sub-band overlap [2]. As open sea surfaces can be considered as randomly distributed surface scatterers, for low sub-band overlap, spectral coherence will be almost totally lost on open sea areas, while it will be preserved on man-made structures, like vessels [2, 3].
The spectral coherence processing consists to divide the acquisition into sub-bands, compute the coherence between each pair of sub-bands, and, the modulus of each sub-band image. In order to increase the signal-to-noise ratio, it also computes their arithmetic average, resulting in the intensity and the spectral coherence images.
Several ship detectors have their principle based on thresholding SAR intensity images to identify bright spots [3, 4]. When the backscattering from the sea dominates the backscattering from small vessels, standard detectors based on intensity only produce false alarms and miss detection of small ships, especially the ones using non-adaptive thresholds. To circumvent these problems, statistical tests on the local clutter can be developed to adjust the threshold accordingly. The Constant False Alarm Rate (CFAR) algorithm is an adaptive thresholding algorithm commonly used to keep constant the probability of false alarms [4].
A first analysis, using the CFAR algorithm applied on the intensity and the spectral coherence image, was made for TerraSAR-X Spotlight data over the city of Venice (Italy) [1]. It shows that all vessels observable in the intensity image are easily detected in the spectral coherence image. Therefore, the spectral coherence and the intensity images can be considered as very good complementary information channels for vessel detection [1].
Our processing technique is divided in five or six stages, depending on availability of AIS data. It starts by (1) performing band split of a given SAR acquisition and estimating spectral coherence between split-band sub-images. The two next stages are (2) geo-projection, and (3) application of CFAR to the resulting images. The fourth stage consists in filtering the detected targets and selecting some of them by combining results from both channels (FLTCMB). The fifth stage extracts the targets that are outside the land areas, defined from shape files. The last stage, when AIS data are available, provides a comparison between the detected targets and the AIS data.
We applied our processing technique on a Sentinel-1 image (48.3MHz bandwidth) acquired on the Libyan Sea on April 12 2018. Spectral coherence processing was performed with 51 sub-bands of 5 MHz. CFAR processing was applied with several sets of window sizes. The best compromise between percentage of positive detections and false alarms was obtained for window sizes (target, guard, background) of 90, 210, and 270 m. We made 39 detections, 12 of which matching some of the 15 AIS targets. The 3 AIS targets not detected are either close to the coast or not visible in the original data. A comparison of both images provided by spectral coherence in the neighborhood of some detected targets demonstrate the usefulness of combining detection from both channels.
To evaluate our processing technique, we performed a comparison with the SNAP [5] “Ocean feature” tool, CFAR-based and commonly used by the Remote Sensing community. To show the benefit of exploiting spectral coherence, we applied this tool on the intensity image obtained after geo-projection. The two techniques provide 20 common detections. Our technique detects 19 additional targets, while the SNAP “Ocean feature” tool detects one additional target: the harbor. This comparison highlights the benefit of using the spectral coherence processing before applying CFAR.
To conclude, we present a processing technique for detecting potential vessels within open sea areas using SAR data. This technique is based on a dual modalities (intensity and spectral coherence) CFAR processor and includes selector to limit detection to open sea surfaces, excluding lands delimited from shapefiles. A comparison with AIS data is also presented. It allows determining the ratio of non-cooperative vessels (or vessels not equipped with AIS) within the area. This technique and the one of the SNAP “Ocean feature” tool, commonly used by the Remote Sensing community, were compared. Experimental results on SAR data acquired on the Libyan Sea show that this processing succeeds in detecting targets and results outperformed the ones of the SNAP “Ocean feature” tool. Some processing parts would still need some improvements, like the CFAR detection of targets close to the coasts.

Area D - Board 425

Marine surveillance has recently gained a lot of interest interest because coastal areas are exploited by illegal fishing, traffickers and illegal immigration. Ship classification is becoming increasingly very important to monitor these situations since authorities want to guarantee the security of marine activities and ensure a safe navigation. Satellite synthetic aperture radar (SAR) missions are important instruments for wide area maritime monitoring thanks to their ability to provide all weather day and night imagery. Vessel detection by means of CFAR techniques on SAR images is relatively easy. On the contrary, it is difficult to classify and recognize the vessel type because the aspect of vessels in SAR images is very peculiar due the imaging mechanism of the SAR, characterized by double reflections, speckle, effects due to movement, etc.
This work aims at classifying vessels based on a deep learning (DL) approach using convolutional neural networks (CNN). Differently from machine learning classification problems based on feature spaces used in previous works [Xing et al. 2013, Lang et al. 2017], DL doesn’t need a preliminary analysis to select and extract features. In fact, DL extracts features from images directly (primitive features as points, edges, etc. in the first network stages and more complex ones in the successive stages). However, the disadvantage of DL is the need for large amount of ground-truth data to avoid overfitting (i.e. the problem of fitting noise and disturbances contained in the training data set, thus reducing generalization and prediction capabilities), possible if the network has a high number of parameters.
The SAR data used in this work are the Sentinel-1 (S-1) Interferometric Wide Ground Range Detected (IW GRD) imagery with a 10 m pixel-spacing. We used the information provided by the Automatic Identification System (AIS) in order to label samples (patches) extracted from S-1 SAR data to build the ground-truth data set. The AIS data are Very High Frequency (VHF) radio signals used among nearby vessels to broadcast information about position, speed, course and identification. The full dataset we built consisted in 5189 vessel clips. In order to perform the training and measure the achievable classification accuracies, we experimented different three‑splits of the dataset.
We tested two different network designs for DL, in terms of training time, classification accuracy and generalization capability, to classify three types of ships; cargo, tanker and passenger. The first CNN is based on the pre-trained network VGG-16 while the second followed an experimental design. Both the CNNs included:
• Data Augmentation technique to create more data through rotation, flipping and zooming of the tiles.
• Dropout technique that deactivate randomly a specific percentage of neurons to avoid overfitting.
• Convolutional layers that extract features through filters.
• Max-pooling layers that control overfitting reducing spatial dimensions and the number of parameters.
• ReLU layers that increase the non-linear properties of the network.
The performed tests shows how the two networks behave with different training dataset sizes and what classes of vessels are most problematic. Besides the differences between the two CNNs, the obtained results are very promising and confirm the validity of the proposed deep learning approach for vessel classification.

Area D - Board 426

Satellite based Synthetic-aperture radars (SAR) like Sentinel-1 are being used for maritime surveillance in Norwegian interest areas. Ships not transmitting their position on Automatic Identification System (AIS) or similar systems, can be detected in these satellite images, and is an important asset to more common AIS based surveillance.

Due to the large amount of daily SAR images available, there is a need for fast and automated vessel detection algorithms to be able to process all these images within accepted latency. The statistical algorithms commonly used today, struggle with areas containing ice, and need human supervision to remove false detections. There is a need for autonomous algorithms that can deal with these icy conditions with far less false detections than today. Such an algorithm would make it possible to cover larger areas, with less latency and lower cost.

A machine learning based vessel detection algorithm, called Nevron, has been developed to improve ship detection in icy waters. The algorithm is trained on a large amount of satellite images with corresponding vessel positions from different scenes, including images containing ice and land. This makes the algorithm able to distinguish ships from sea, ice and land, making it less dependent on masking.

The performance of the algorithm is compared to vessel detections delivered by a commercial company. These vessel positions are reviewed by a human operator and reported ship positions from e.g. AIS are used actively to add ships. The comparison show that the machine learning based algorithm can compete with the performance of the trained human operators. The vessel detector is also compared to the output of their statistical detection algorithm, where Nevron performs better, mainly due to far less false detections in regions with drift ice.

The accuracy of Nevron, makes it possible to run the algorithm autonomously without human intervention. The algorithm is optimized for speed by utilizing a fully convolutional neural network and a graphical processing unit (GPU), and is able to cover complete Sentinel-1 GRDH images in 1-3 minutes each. This makes it possible to survey much larger areas than commonly seen, with both good accuracy and little latency.

Area D - Board 427

Summary
The overall aim of this project was to analyse the conditions that allow the visibility of slicks in North Atlantic and Pacific oceans gyres. Our hypothesis is that increasing amounts of microplastics within the marine environment, produces a higher microbial activity in areas otherwise oligotrophic which on return produces slicks and damping of capillary waves. These microplastics have a direct impact on marine species from plastic entanglement to ingestion [1]. Microplastic debris is considered to be widely under-researched due to the difficulties associated with assessing the abundance of such small fragments (<0.5mm) and quantifying the density and distribution of the debris [2].

Detection of exogenous floating marine debris (EFMD) has been analysed due to the worldwide concern and possible long-term threats to the environment and human health, caused by microplastic [3]. We have recently seen that using SAR we can observe slick in the gyres [4]. Aim of this work is to investigate the conditions that are associated to this slicks, which special interest to Chlorophyll-a content.

Methodology:
The aim of this research was to investigate the use of remote sensing satellites, such as the Copernicus and Sentinel satellite systems, to determine the conditions that are associated to the slick observed in [4]. With special interest for chlorophyll-a concentrations. The detection of slicks can be determined from wave activity, however other factors have been considered, such as meteorological conditions. Slicks were detected using Sentinle-1 data. Chlorophyll-a concentrations were extracted from the Terra MODIS satellite via NASA’s Ocean Biology Processing Group. Meteorological conditions were gained using Copernicus and European Centre for Medium-Range Weather Forecasts (ECMWF). Finally all the data were processed using Sentinels Application Program (SNAP).

Sentinel-1 data can be used to determine surface roughness and dampening of the capillary waves [5]. We can therefore use them to identify the slicks. Sea oil spills, shipping oils and surfactants on the sea surface all produce a reduction of surface roughness as chlorophyll-a biogenic slicks. We want therefore monitor to Chlorophyll-a of these slicks in the gyres to understand if their origin is due to algal blooms, or if they were produce by other microbial activity related to other causes. Meteorological conditions can also factor into the visibility of surfactants on the sea surface. We used ECMWF datasets. These were subsequently analysed using the ESA Copernicus Climate Data Store.

Results:
In all the gyres slicks analyses so far, we always found that the wind conditions were in the regimes that allows visibility (2 to 9 m/s) while the value of Chlorophyll-a were very low and generally too low to be associated to algal blooms. There are however microbes that can produce slicks but not have a large Chlorophyll-a content. Such microbes are also the ones that would colonise microplastic debris (i.e. using them as support). This results therefore seems to provide more evidences that micro-plastic in some conditions may produce slicks.

References
[1] Cole M, Lindeque P, Halsband C, Galloway Tamara S., 2011, Microplastics as contaminants in the marine environment: a review, Marine Pollution Bulletin 62. 2588-2597
[2] Doyle, M.J., Watson, W., Bowlin, N.M., Sheavly S.B. (2011). Plastic particles in coastal pelagic ecosystems of the Northeast Pacific ocean. Marine Environmental Research, 71(2011), pp. 41-52
[3] Paige, M.L.B., Painho, P. (2015). Detection of exogenous floating marine debris: an overview of techniques associated with remote sensing. WIT Transactions on Ecology and the Environment, 196, pp. 537-548.
[4] Davaasuren, N., Marino, A., Ackermann, N., Alparone, M., Nunziata, F. and Boardman, C. (2018). Detecting microplastics pollution in world oceans using SAR remote sensing. In: IGARSS 2018: IEEE International Geoscience and Remote Sensing Symposium, 22-27 Jul 2018, Valencia, Spain.
[5] Alpers, W. & Espedal, H. A. (2004) Chapter 11. Oils and Surfactants. In: Jackson C. R., & Apel J.R. (eds.) Synthetic Aperture Radar Marine User’s Manual. Washington, DC, U.S Department of Commerce National Oceanic and Atmospheric Administration (NOAA), pp. 263-276.

Area D - Board 428

Summary
In this work, we tackle the problem of monitoring changes in coastal areas using SAR polarimetric data with two frameworks for change detection. The two methodologies are based on the use of the Lagrangian optimisations of two distinct operators: a power ratio and a power difference. The two detectors are based on different signal models, a multiplicative and an additive model. In this work the difference between these two signal models is tested over the task of detecting erosion and phenology changes in salt marshes. The results clearly reveal that the eigenvalues of the multiplicative model produce a clearer detection of changes while the eigenvectors of the additive model allow a straightforward physical interpretation of the dominant scattering mechanisms that have been changed in the marsh (i.e. classification of changes). The two detectors are therefore complementary in providing information. The area under investigation is the Solway Firth in Scotland. The most of the data used are Sentinel-1 dual-pol however we also exploited 2 ALOS-1 quad-pol scenes and 3 RADARSAT-2 quad-pol scenes. The latter allow to appreciate the different results between dual-pol and quad-pol data. The changes were validated using drone aerial photography that we collected on the area between August 2018 and March 2019.

Monitoring Salt marshes:
Salt marshes and coastal grassland areas are ecologically and economically valuable ecosystems. They support a wide range of biodiversity including a rich invertebrate community and important wading birds and wildfowl. They also play an important role in carbon sequestration, act as nursery grounds for economically important fish and reduce the impact of storms on coastal defences. They are, however, fragile ecosystems [Gedan et al. 2009].
Unfortunately, salt marshes are under treat in many places all around the World. The main causes of changes and degradation of marshes are anthropogenic (pollution, agriculture, buildings) and related to climate changes (e.g. the increase in storm frequency and strength). The only way to make sure our marshes are properly protected is by carry out periodic surveys. Although these are spectacular ecosystems, surveying them is very time demanding due to the uneven terrain (presence of creeks, ponds and mud). Additionally, many feature are more easily visible from above (e.g. draining system). The community is therefore in need of some methodologies able to provide periodic and inexpensive surveys of marshes.
We want to demonstrate that satellite remote sensing could be the key. In this project we use satellite Synthetic Aperture Radar (SAR). SAR is able to obtain images of the environment from space using microwave. It allows to acquire images independent of weather condition and solar illumination which is very valuable in areas with frequent cloud cover. We will also polarimetry (PolSAR). The advantage of PolSAR is that we can use the polarisation of the radar echo to obtain more images and therefore more information about objects in the scene [Cloude 2009].

Signal Models
The starting point of this work are the change detectors based on the optimisation of the power ratio [MarinoOpt, AlonsoChange] and difference. The optimisations are accomplished diagonalising specific matrices. The eigenvalues can be used to detect the changes and the eigenvectors can be used to classify them.

Results:
The output of the detector can be represented using RGB images, where the intensity is related to the eigenvalue and the colour related to the eigenvectors. It can be observed that the power detector is able to monitor more easily changes in coastal areas. This is because the ratio has a large dynamic compared to the difference. However, when it comes to the interpretation of the results, the eigenvectors of the difference allow a better estimation of the scattering mechanisms that are producing the change. They therefore allow a better classification of the change.
Combining both optimisation, we can use the ratio to identify/detect the area and the difference to interpret the change.
A validation with drone data is performed, where we evaluate situations in which we are able to see the changes and the one where we cannot monitor them.

References:
[Alonso et al 2015]: A. Alonso-González, K. Papathanassiou, I. Hajnsek (2015): Physical interpretation of the Polarimetric SAR changes observed over agricultural time series., BioGeo SAR, Harwell, UK, 2015.
[Cloude 2009]: S. Cloude (2009), Polarisation: Applications in Remote Sensing} Oxford University Press, Oxford, UK, 2009.
[Gedan et al. 2009]: K.B. Gedan, B.R. Silliman, M.D. Bertness (2009), Centuries of Human-Driven Change in Salt Marsh Ecosystems. Annual Review of Marine Science, 1, 117–141.
[Marino et al 2014]: A. Marino, I. Hajnsek (2014), A Change Detector Based on an Optimization With Polarimetric SAR Imagery, IEEE Transactions on Geoscience and Remote Sensing, 2014, 52, 4781-4798.

Area D - Board 429

Summary:
The presence of icebergs in cold waters represents a danger to navigation. Tracking and detection of icebergs has seen a lot of work carried out in the past decades using synthetic aperture radar (SAR) altimeters and scatterometers. However, previous work was limited in that the detection was dependant on iceberg size. Smaller icebergs continue to pose a challenge for detection, particularly those embedded in sea ice.

The scattering behaviour of icebergs is the main focus of our work. Using Sentinel-1 data acquired from Antarctica, and the coasts of East and West Greenland, we conducted a polarimetric and radiometric analysis using the application of several detectors and polarimetric observables. The datasets used are mostly Sentinle-1 dual-polarimetric, but we also exploited quad-pol ALOS-2 and RADARSAT-2 to evaluate the added benefit of the third polarisation channels. Using quad-pol data we could also explore some decompositions, including Cloud Pottier, model based and coherent. Finally we used few Sentinel-2 data to validate presence of small icebergs in some of the Sentinel-1 data.

Methodology:
Following the extensive analysis, a solution based on combination of previously developed detectors is proposed. In particular we make large use of the intensity Dual Pol Ratio Anomaly Detector or iDPolRAD (1), the Polarimetric Notch Filter (2) and the Cell Averaging CFAR. The parameters of these three detectors are modified producing different versions that can be focused on specific scattering mechanisms, derived by the polarimetric analysis. In the ordinary version of the iDPolRAD two boxcar filters are applied over the HV and HH intensity images, exploiting two different window sizes: a smaller window or test window Wtest and a larger window or training window Wtrain. The detector can be writes as the different of the “HV filtered with the test window” minus “HV filtered with the train window”, all divided by the “HH filtered with the train window”. In this work we propose a new iDPolRAD were the intensity images are obtained considering the projections of the scattering vectors over the eigenvectors for the dual-pol case (or quad-pol if available). The intensity image at the denominator is the one that form the maximum eigenvalue for an area with an iceberg. While the intensities at the numerator are obtained using the eigenvector that is minimum over an iceberg area. Since different icebergs have different signatures, we apply different iDPolRAD versions and combine the results at the end, together with the PNF. The combination is simply done by summing the results of each detector. In the case of the PNF, the parameter Pt (power of target) is used and not its normalised version gamma. In this preliminary version, at the end we apply a CA-CFAR detector on this detector sum.

Results:
The findings establish that icebergs have different polarimetric behaviours depending on conditions. However, the different polarimetric signatures seem to be well characterised by either volume or multiple reflections. The results of the detection exercise show that on quad-pol data the detection performance is enhanced if we use the Cloude Pottier decomposition on historical data to extract the two most common iceberg signatures. On dual-pol data, the improvement is not very evident and a single iDPolRAD may still be the solution if we want to have a quick detector and not reply on a priori information.

References:
[1] A. Marino, W Dierking, and C. Wesche, “A depolarization ratio anomaly detector to identify icebergs in sea ice using dual-polarization SAR images,” vol. 54, no. 9, pp. 5602–5615, 2016.
[2] A. Marino (2013) A notch filter for ship detection with polarimetric SAR data, IEEE Journal of Selected Topics in Applied Earth Observation and Remote Sensing, 6(3) pp. 1219–1232.

Area D - Board 430

Satellite observations have been demonstrated to represent a valuable tool for the detection of objects over sea, like ships, oil slicks, icebergs, wind turbines and fish farms. While optical sensors are not able to detect the Earth surface under cloud coverage conditions, Synthetic Aperture Radar (SAR) is capable of observing the Earth surface under almost all weather conditions.
Some climatic events only occur under certain weather conditions. In example, Sirocco wind blowing in the Mediterranean basin is a wind regime usually accompanied by prolonged high cloud cover conditions, and represents a major hydro-climatic stressor that intensively determine sediment resuspension and consequently coastal erosion. Due to high cloud cover and heavy rainfall conditions during the Sirocco wind regime, optical satellite data have been demonstrated to be uncapable of observing the fingerprint of such climate physical forcings, like the river plume dispersal patterns. On the other side, the capability of SAR images to detect water regions exhibiting higher turbidity levels has not been demonstrated yet. During intense rainfall, river flood peak brings floating material, both organic material and waste material, to the sea. The suitability of satellite observations for the identification of such floating material, drifting to the open sea according to the local circulation and frequently accumulating on the coastal beaches, has not been fully investigated.
This research study show the evidence of the capability of high resolution Sentinel-1 SAR and Sentinel-2 MSI optical sensor in the identification of floating drifting materials over the sea surface, like floating materials at the boundary of river plumes and other floating organic materials. It makes use of both Sentinel-1 GRD and Sentinel-2 MSI data to identify different kind of floating drifting materials from radar backscatter anomalies, optical spectral signatures and spatial patterns. Different case studies are used to show the capability of SAR and optical high resolution data to locate drifting floating materials and the dispersal pattern of river plumes.
Results demonstrate that the detection of new fingerprints of climate physical forcings on SAR signal over sea surface may represent one of the potential new advances in satellite SAR. Moreover, the use of satellite data for the detection of new types of drifting floating materials over sea could be used to improve the security of the maritime space.

Area D - Board 432

Icebergs are considered a threat to marine operations. Satellite monitoring of icebergs is one option to aid in the development of iceberg hazard maps. Satellite synthetic aperture radar (SAR) is an obvious choice because of its weather independence, day and night operation. The SAR response from icebergs is generally complicated due to the random geometric shapes of icebergs. To discriminate icebergs from other ocean targets such as sea ice and vessels, automated classification is important. Therefore, a study of polarimetric properties of icebergs is vital to optimize the surveillance of icebergs and enhance iceberg detection in background clutter.
Polarimetric decompositions have been extensively used to study the physical property of scattering behaviour of remotely sensed target for decades. Model based decomposition such as Freeman Durden decomposition aims to measure the three elemental scattering mechanisms from the polarimetric covariance matrices of a target: surface scattering from ground, double bounce from the interaction with corner reflector and volume scattering from the uniformly randomly oriented dipole. Nonetheless, several studies have found drawbacks in the model-based decomposition. The major flaw of the model-based decomposition is the over-estimation of volume scattering. This decomposition classifies all the cross-polarization backscatter component as volume scattering but some of the cross-polarized returns might be the contribution from the rough surface scattering or terrain effects. The pure volume scattering should be reflected from randomly oriented dipole scatterers which have not been taken into account. One study has proposed a non-negative eigenvalue-based decomposition that can overcome these limitations by eliminating the negative eigenvalues to remove the over-estimation of randomly oriented dipole scattering in model-based decomposition. For vegetation terrain, the decomposition results have reduced the over-estimation of volume scattering by 15-20%. This study used vegetation terrain at L-band for the analysis. The over-estimation here is reasonable since the image scenes are forest areas and L-band energy has better penetration (relative to other SAR frequencies, such as C or X) through the canopy and interacts with the ground.
Consideration has been given to the over-estimation of volume scattering in the case of icebergs and whether the model-based decomposition is valid in this case. This presentation is focused on investigating the over-estimation of the volume scattering from iceberg targets in Freeman Durden decomposition by implementing non-negative eigenvalue decomposition. Then, by comparing both decompositions, the results will confirm and quantify the over-estimation of volume scattering for iceberg targets. This, analysis have been performed on a validated RADARSAT-2 dataset at C-band. Iceberg targets have been extracted from five RADARSAT-2 images acquired in the Newfoundland and Labrador region. Given that the physical properties of iceberg backscatter is different from that of vegetation and the shorter C-band wavelength typically interact more with randomly oriented scatterers, the over-estimation of volume scattering in model-based decomposition is expected to be less than for vegetation terrain at L-band.

Area D - Board 433

Iceberg detection and characterization is of high interest not only for Near Real Time Maritime awareness but also for climate and physical/biological oceanography studies. Whereas tabular icebergs are well tracked by scatterometer (see BYU/NIC database), the distribution of smaller icebergs (< 1-2 km) is still largely unknown.
Thanks to the use of altimeter sensors (inter alia, ERS-1, ERS-2, Jason-1, Jason-2, CryoSat-2, Topex, Envisat, AltiKa…), (Tournadre et al. 2008) showed that smaller icebergs can be detected bringing an unprecedented description of the iceberg distribution at unprecedented time and space resolutions (see Altiberg database). Given the 1-dimensional sampling of altimeter data, SAR images can provide new insights on the actual distribution of these smaller icebergs. Especially, the systematic acquisition of Sentinel-1 Wave Mode (WM) imagettes in the open ocean (and from 2002 with ENVISAT/ASAR) could complement the Altiberg database with even smaller icebergs (> 10-100 meters) and better characterization (icebergs area).

In this study, we have thus massively processed the entire ENVISAT and Sentinel-1 missions looking for icebergs in Wave Mode SAR imagettes from 2002 to now. The general idea is to attempt a climatology study of iceberg presence/size in complement to Altiberg database.

We have implemented a new iceberg detection/characterization chain for ENVISAT and Sentinel-1 WM products with the following processing units:
- A pre-detection scheme based on Inception-v3 Convolutional Neural Networks (CNN) framework enables to efficiently discard all the imagettes without any potential icebergs.
- To locate the icebergs within the imagette, a multi-scale detection approach is used based on various configuration of the Constant False Alarm Rate algorithm (size of the target cell but also of the background, various False Alarm Probability…). For the WM data acquired at 23°, icebergs sometimes appear as "black iceberg" (it affects mostly S-1 but not ENVISAT acquisitions). An inverse CFAR detection looking for black echoes has been thus integrated.
- All these CFAR-detected points are then clustered based on distance criterion
- A classification process checks whether the echo should be considered as potential iceberg or false alarm. A coherency test based on sub-look SLC processing enables to remove short-lifetime target such as breaking wave. In addition, a radiometric evaluation removes bright echoes with low backscattering or due to CFAR detection in noise-polluted background.
- A multi-scale segmentation algorithm is built accounting for both large icebergs (> 1km) but also very small icebergs (< 100 meters), taking into consideration whether the target has been detected as either black, bright or both black/bright echo.

This chain has been applied to the entire archives of ENVISAT/ASAR and S-1 a/b SLC Wave Mode products thanks to the IFREMER Datarmor supercomputer. The results are presented and compared to the Altiberg database.



This study is being funded by the CNRS/LOPS via the SARICE project.

Area D - Board 434