Conference Agenda

Arctic region is experiencing faster warming than that in the middle and low latitude region, which triggers extensive change of physical and biological Arctic systems. Sea ice, as one of the most important indicator of climate change in Arctic, is showing an accelerating declination trend. The loss of sea ice in the Arctic is not only a simple outcome of climate change, but it also has strong feedback to regional and global climate. In addition to the climate change issue, Arctic is being paid more attentions than ever before, as increasingly accessible for extensive shipping (e.g., the recently opened Northern passages in summer) and oil and gas exploration. Satellite observation have been exploited for Arctic sea ice monitoring for more than a few decades, as they have advantages of large coverage and flexible acquisitions. In contrast to optical imaging, microwave sensors are not impeded by cloud coverage or lack of daylight. Among the microwave sensors, Synthetic Aperture Radar (SAR) has unique advantages of high resolution, as well as large coverage.

For Arctic sea ice monitoring, in the first step, we focus on deriving static and dynamic information, e.g., classification of ice types, ice-water discrimination, ice concentration and drift of sea ice. We will present our recently collaboration on deriving such information based on C- and X-band SAR, including TerraSAR-X, Radarsat-2 and the Chinese recent launched GaoFen-3 (GF-3). Different algorithms and approaches have been developed for these spaceborne SAR data derive sea ice information. In particular, the SAR data acquired in multiple polarizaiton are exploited. In addition to derive sea ice information from SAR data, we will also present some preliminary studies on interaction of swell and ice in the marginal ice zone (MIZ).

A fully automatic SAR oil spill monitoring system briefly consists of three parts, 1) dark target segmentation on SAR image 2) dark target feature extraction 3) dark target classification. In order to build a fully automatic system with high detection and high classification performance, the first step requires all oil spill targets to be extracted in the first step, the second step requires the category distinction of dark targets in the feature space to be as high as possible to reduce the classification complexity and the third step requires the classifier to deal with imbalance problems.

The classification complexity can be simply expressed as the size of the overlapping regions of the distribution of the two types in the feature space. The larger the overlap, the more difficult it is to train a high performance classifier. The negative effect of imbalance is equivalent to a multiplicative amplification factor attached to the complexity. The imbalance problem would disappear if complexity is reduced zero. However, the complexity is impossible to be zero, therefore the technical effort to build a high performance automatic oil spill monitoring system is to try to reduce classification complexity and to deal with imbalance problem.

In order to ensure that all oil spills are extracted, the existing segmentation algorithm for extracting dark targets from SAR images will lead to a large number of look-alikes appear, which contain not only dark targets caused by common atmospheric or oceanic phenomena, but also lots of unexplained ones. That is the source of the imbalance problem. In the segmentation stage, an adaptive threshold segmentation algorithm based on multi-scale background normalization is adopted at first to ensure that all oil spills are extracted, and then a post-processing filter chain is used to reduce the imbalance of the targets from 1:100 to n1:5 below.

77 features collected from different researchers are used to improve the category distinction of the types in the feature space, that is, to reduce the complexity of the classification problem. The more the features, the more weights the network needs to adjust, therefore the more training samples are required. However, all 77 features are still used in our study since there are sufficient samples available and computation amount is not so high. Some wrong category labels of samples due to human error may increase classification complexity, especially in the case of a large number of samples. By using an iterative sample category designation method combined with manual classification and machine classification, the number of error labels is significantly reduced and thus the classification complexity is reduced either.

Double-hidden-layer neural network has strong capability to fit complex decision surface, but it also easy to fall in over-fitting and suffer high false discovery rate under the circumstance of imbalance. The Adaboost integrating multiple double-hidden-layer neural networks plays a role in smoothing the decision surface, thus weakening the over-fitting effect and enhancing the generalization performance of the classifier. Moreover, Adaboost trends to reduce the false alarms in imbalance condition since its component classifiers pay more attention on the samples misclassified by the preceding ones. Therefore the Adaboost based on neuronal network has capability to deal with the imbalance problem.

By applying the techniques mentioned above, a fully automatic SAR oil spill monitoring system with detection rate of 81%, false discovery rate of 17%, false alarm rate of 5% and correct recognition rate of 92% is obtained for the sample set consisting of 23768 dark targets extracted from 337 scences of Envisat ASAR and Cosmo-SkyMed images.

In 2008, an extensive Ulva Prolifera bloom occurred in the Yellow Sea (YS) just before the Olympic Sailing Regatta and draw much attention of the world. Then Ulva bloom annually occurred in the YS each summer and attracted the interest of the ocean remote sensing and marine biologist community. Wide swath and short re-visit time optical satellite data were first utilized to monitor the Ulva bloom and estimate their coverage area using methods such as NDVI and FAI. However, the estimated Ulva bloom coverage area were very different in the published literature mainly due to adopted different thresholds T₀ and T₁. T₀ was used to differentiate the algae-containing pixel with seawater pixel while T₁ was used to differentiate algae full-covered pixel with partial-covered pixel. In this study, those two thresholds were further studied.

Threshold T₀ was derived as follows: (1) First, generate the FAI and FAI gradient image for each optical satellite imagery; (2) classify all pixels in the satellite image as seawater pixel and non-seawater pixels based on the FAI gradient image;(3) replace non-seawater FAI values with median FAI of the neighboring seawater pixels to construct the seawater FAI background image; (4) generate seawater-removed FAI image by subtracting the seawater FAI background image from the FAI image, where pixels greater then zero will be algae pixels.

In the recent study, the author developed the Ulva bloom biomass estimation model and found that (1) FAI threshold T₁ was significantly underestimated in the former study (2) FAI linearly increased with the increase of algae coverage but increased slowly after algae completely covered, thus the turn-point in the FAI-biomass curve was the FAI threshold T₁. In this study, the lookup table of threshold T₁ for specific satellite sensor were simulated in different sun/sensor viewing geometry and atmospheric conditions based on the filed measured algae spectrum.

The high-resolution satellite data were used to validate the results in this study. (1) The Ulva bloom coverage estimated from MODIS data and concurrent high-spatial resolution Chinese GF-1 WFV data (16m) were compared with the relative difference 10%. (2) The threshold T₁ was validated with high-spatial Worldview-2 satellite data.

This method can be implemented to OLCI/Sentinel-3 data and the subsequent estimation Ulva bloom coverage from OLCI data will be conducted in this summer.

Synthetic aperture radar (SAR) imagery provides an effective source of data for observing, measuring and quantifying oceanographic phenomena. The ability of SAR sensors in retrieving data in almost all weather conditions, independently of sunlight illumination, is an extremely useful aspect, important for oceanographic applications. The quality of SAR imagery is dependent on the mode of acquisition and raw data processing. SAR data acquisition techniques introduce a significant backscattering trend in the range direction of the received signal. In Wide Swath and Extra Wide Swath modes, results in a progressive reduction of brightness over images from near to far range, introducing errors on the detection and classification of dark features to oil spills and lookalikes. The present research, aims to examine normalization techniques previously applied to ENVISAT SAR Wide Swath data. We investigate possible methods for limiting the issue of Normalized Radar Cross-Section (NRCS or σ°) variation, due to the incidence angle variations in Sentinel-1 Wide and Extra Wide Swath and data. NRCS depends on the relative azimuth angle between the radar look direction and wind direction. At low incidence angles over a certain wind speed and direction NRCS values are different from these at high incidence angles. In order to eliminate errors during image processing and analysis for oceanographic feature extraction, a normalization is required to limit the NRCS variation over the various incidence angles is suggested. Based on previous studies, the most widely used incidence angle correction technique is the square cosine correction. However, the square cosine correction is valid for surfaces with Lambertian reflectance properties and is not expected to perform in a satisfactory way over the sea. Here we apply two sensor independent functions aiming to improve the dark object detection in SAR imagery over ocean: a theoretical backscattering shape function which is derived from a minimum wind speed and an empirical range fit of NRCS against incidence angle θ. The former method exploits only the modelled NRCS values, while the latter only the image content. The aim of this paper is to apply a simple but consistent scheme for reducing the dynamic range of SAR images by removing the mean incidence angle dependence. The approach targets to normalize the Wide and Extra Wide Swath SAR image to a fixed reference angle. From the approach a single global threshold is applied to detect dark objects in SAR imagery.

ABSTRACT

The pollution of the marine environment due to deliberate discharge of untreated ballast waters is not only an operational procedure done by ships but also by offshore platforms. Although, the amount of oil released into the ocean by the offshore platforms using this procedure on one-time basis is derisory, if compared with the massive spill event in case of platform accidents like the Montara and Deepwater Horizon, the long term damages on the flora and fauna are severe.

As the North Sea is characterized by fairly shallow water bathymetry, it has been extensively exploited in the past years for oil extraction and production. By now, it hosts a significant number of offshore installations and therefore the probability of minor leaks is quite high. A large number of Synthetic Aperture Radar (SAR) images are being acquired and processed operationally over North Sea platform installations for potential oil pollution using the TerraSAR-X satellite [1]. The ScanSAR and WideScanSAR mode in VV polarization imagery have been preferred due to large coverage and higher oil-water contrast. Among the dataset collected, a constant leakage from the platforms belonging to the Forties oil field has been observed.

Thanks the low TerraSAR-X orbit altitude and relatively high latitude location of the Forties oil field, leaks from the same platforms have been observed with temporal interval of less than 13 h. While most previous studies on tracking oil spills assume that the observed slicks by consecutively acquired SAR images are the same and spatial displaced by the drift effect [2], a completely different situation is outlined by the analysis in [3]. By model simulation it is shown that leaks were not start-stop but continuous with only part of the old oil being drifted.

REFERENCES

[1] S. Singha, D. Velotto, and S. Lehner, “Near real time monitoring of platform sourced pollution using TerraSAR-X over the North Sea,” Mar. Pollut. Bull., vol. 86, no. 1–2, pp. 379–390, Sep. 2014.

[2] Y. Cheng et al., “Monitoring of Oil Spill Trajectories With COSMO-SkyMed X-Band SAR Images and Model Simulation,” IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., vol. 7, no. 7, pp. 2895–2901, Jul. 2014.

[3] X. M. Li, T. Jia, and D. Velotto, “Spatial and Temporal Variations of Oil Spills in the North Sea Observed by the Satellite Constellation of TerraSAR-X and TanDEM-X,” IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens., vol. 9, no. 11, pp. 4941–4947, Nov. 2016.