Satellite data lift the veil on offshore platforms in the South China Sea

Oil and gas exploration in the South China Sea (SCS) has garnered global attention recently; however, uncertainty regarding the accurate number of offshore platforms in the SCS, let alone their detailed spatial distribution and dynamic change, may lead to significant misjudgment of the true status of offshore hydrocarbon production in the region. Using both fresh and archived space-borne images with multiple resolutions, we enumerated the number, distribution, and annual rate of increase of offshore platforms across the SCS. Our results show that: (1) a total of 1082 platforms are present in the SCS, mainly located in shallow-water; and (2) offshore oil/gas exploitation in the SCS is increasing in intensity and advancing from shallow to deep water, and even to ultra-deep-water. Nevertheless, our findings suggest that oil and gas exploration in the SCS may have been over-estimated by one-third in previous reports. However, this overestimation does not imply any amelioration of the potential for future maritime disputes, since the rate of increase of platforms in disputed waters is twice that in undisputed waters.


Supplementary: the South China Sea
The South China Sea (SCS) is crowded, environmentally sensitive and strategically important. It is the third largest marginal sea in the world, bordered by the Chinese mainland and Taiwan to the north, the Philippines to the east, Vietnam, Thailand, and Cambodia to the west, and Brunei, Singapore, Indonesia, and Malaysia to the south ( Supplementary Fig. S9). Although it is ecologically a distinct large marine ecosystem, the Gulf of Thailand is often included in the reference made to the SCS. The SCS has an area of 3.3 million km 2 excluding the Gulfs of Thailand and Tonkin, and up to 3.8 million km 2 if these gulfs are   The geometric accuracy of Landsat-8 OLI L1T can reach to 12 meters (i.e., approximately a half pixel in OLI images) 22 . Moreover, the South China Sea is a tropical sea, and is often affected by clouds -hence, most of the images are of poor quality. The cloud cover ranges from 0.04 to 61.33%, with a mean of 15.49% (n = 1035 images).

Supplementary: Method for offshore platform detection
Offshore platforms are essential infrastructure required to drill wells, extract, process, and temporarily store crude oil and natural gas. In the SCS, there are various types of offshore platforms, including the old model of small and fixed jackets in shallow coastal waters, large floating production storage and offloading units (FPSO) in deep water, jack-up, semi-submersible production units, stacked-led structures, spars, and others. Depending on the offshore oil/gas field water-depth and situation, platforms can be fixed to the ocean floor, or be moored. According to our previous study, the shift of position of a FPSO is no more than tens of meters 30 . In the SCS, most platforms are small and fixed. Usually, these platforms exhibit no more than 10 pixels in a 30 m resolution images, while the size of some hybrid platforms which are connected by several sub-platforms can reach dozens of pixels.
In general, the offshore oil and gas platforms are generally made of various grades of steel, from mild steel to high-strength steel, although some of the older structures were made of reinforced concrete. These metallic structures with a high degree of exposure to sunlight usually exhibit a high digital number (DN) in the short-wave infrared band of the optical images, and high radar cross section (RCS) in SAR images. In addition, some platforms burn off excess associated petroleum gas (APG), which make them rather significant in the nighttime.
Although offshore platforms may be the most significant manmade structures in the sea, detecting them and determining their attribute from space-borne images is challenging. The problems include the following ( Supplementary Fig. S11): (1) Small target. The detection of offshore platforms which usually exhibit one or several pixels in a moderate resolution images without sufficient supporting information (e.g., shape, structure, and etc.) is difficult. (2) Dim target. Many small platforms also present an inconspicuous difference compared with their surroundings, especially for optical images. (3) Noisy backgrounds. The quality of satellite images suffers from poor imaging conditions (e.g., clouds, mists), underlying surface (e.g., waves, ripples, turbidity), and the imaging device (e.g., stripe-missing, speckle noise); thus, offshore platforms are usually immerse in heavy noise. (4) Numerous false alarms. Offshore platform detection also suffers from numerous vessels which also present similar features. (5) Vast sea area. To pick out platforms over a vast sea area (e.g., the SCS) from satellite images (usually low signal-noise-ratio), automation and robustness need to be considered. To extract offshore platforms (usually tiny and dim) over a vast sea area (usually having a high noise and clutter background) from satellite images (usually influenced by poor weather), we improve the automated method for extracting offshore platforms (AMEOP) from time-series OLI images which we proposed previously 30 , and the following general technical framework was used.
(1) Sea surface target detection from mono-temporal satellite images according to spot-like and sparse principles. We assume the signature of sea surface targets in the satellite image is composed of background, sea surface targets (including offshore platforms, vessels, clouds, and others), and noise (Eq. 1). Hence, the signature of a sea surface target at t moment can be denoted by Eq. 2.
Where, f(x, y, t) is the DN of pixel (x, y) at t moment, while f s (x, y, t), f b (x, y, t), and f n (x, y, t) are DN of sea surface target, background, and noise, respectively.
In practice, an order-statistic filtering (OSF) was used to approximate the background (Eq. 3).
Specifically, the pixel in a sliding disk-like window were ascending sorted according to their values, and the r th element was selected to be the background value of the domain. In general, a higher order (r) will be capable of better suppressing non-uniform and complex backgrounds, since offshore platforms are (i) spotlike, (ii) have a higher DN value than those of their surroundings pixels, (iii) are tiny, and (iv) sparsely distributed. Note that the order (r) varies for different scales of the satellite images used: the order (r) for moderate resolution optical images should be set higher to resist the influence of the ripples generated by waves or currents ( Supplementary Fig. S11i and k), while the order (r) for the low resolution NTL products is relatively lower because the flaring of platforms often exhibit halos that are much larger than their actual size ( Supplementary Fig. S11b, c, and d). As a general rule, noise sources in satellite images are unknown.
Hereafter, the local standard deviation of the sliding window is used to approximate the noise. However, the standard deviation is usually subtle for a homogeneous background, especially using a large sliding window where the difference has been unexpectedly diluted, hence, a user-defined constant (ζ) is added to better suppress those noises in the smooth regions (Eq. 4).
(2) Noise removal and offshore platform detection from time-series satellite images according to the position invariance principle. Numerous false alarms, including vessels, pulse noises, and others, will be unexpectedly detected in the above process from mono-temporal satellite images. For two pairwise satellite images with high geometric accuracy, the position-invariance principle can be applied to discriminate between offshore platforms and moving vessels or random small clouds. In many cases, the over-density noise may not be completely removed by pairwise comparison. Moreover, optical images are highly susceptible to clouds, and thus offshore platforms may be covered by prevailing clouds in some specific phases' optical images. Subsequently, a time-series-image strategy was used to robustly eliminate these remnant errors (Eq. 5). In the accumulated image of coupled comparisons, offshore platforms usually had a higher occurrence frequency (high position-invariant consistency in time-series images) than the residual and randomly-distributed small clouds.
(3) Self-consistency check by time-series Landsat-8 OLI images. All targets identified by the aforementioned general platform detection method were overlapped on the time-series OLI images with high quality. Through careful visual examination, we excluded mis-extracted targets, such as small islands and clouds according to their varying size/shape. The detailed extraction process for various sensors is shown in Supplementary Fig. S12-16. Supplementary Fig. S17 give examples for determining types of offshore platforms from Landsat-8 OLI images. Supplementary Table S3 lists all detailed parameters for automated extraction from the used time-series images.
Supplementary Figure 11 | Challenges in offshore platforms detection from space-borne images. For NTL products, e.g., DMSP OLS products (a), and S-NPP VIIRS DNB products (b), the problem is how to robustly distinguish gas flaring/lights of offshore platforms from those of numerous fishing boats (c and d).
For optical images, e.g., Landsat-8 OLI band-6 images, the automation and robustness of platform extraction are challenged by the following factors: (e) the strip effect of the images, (f) tiny targets, (g) dim targets influenced by mist, (j) false alarms including clouds, (k) ripples, (l) noises, and (m) vessels. It should be noted that high-density clouds are prevalent in the SCS, e.g. sub-figure (e) (path/row: 124/053, January 30, 2014, the cloud cover ratio is 3.61 %). Usually such optical images with a cloud cover ratio of less than 5% are rare for the SCS. The coarse geo-location accuracy of Landsat-4/5 TM images in the sea (approximately 150 m) is another issue. Although the radar cross-section of man-made targets (e.g., offshore platforms and ships) is much higher because of the effect of multiple incoming radar waves from the targets' superstructure from the flat and calm sea surface, detecting platforms from SAR images suffers from poor geo-location accuracy (p), numerous false alarms (q), and heavy background clutter (r and s). Moreover, the intensity of backscatter signal may differ with the incident angles, e.g., subfigure (n) taken by ENVISAT ASAR in WSM mode (December 19, 2008), which means that it is difficult to effectively distinguish platforms using a fixed/global threshold. The DMSP OLS NTL data was downloaded from the NGDC Earth Observation Group (EOG) of NOAA (http://ngdc.noaa.gov/eog/download.html). The Landsat-8 OLI SWIR data were available from the Global Visualization Viewer of the U.

Supplementary: Validation
A total of 1311 targets with position-invariant characteristic were automatically extracted from Landsat-8 OLI images. By checking time-series OLI images with high-quality, we removed 13 errors (clouds) and 239 errors (small islands/reefs and docks/groynes, almost all located within 10 km from the coastline), with 1059 remaining. The robustness of the extractions was assessed using three capable spaceborne images: ALOS-1 PALSAR images, Sentinel-1 SAR images, and high resolution images.    S10f), covering an area more than 1.55 × 10 5 km 2 , were used. Among 768 platforms located in the regions covered by these high resolution images, 654 platforms were confirmed, remaining 114 platforms could not be confirmed: 98 cases were because of heavy cloud (among these, 78 cases were confirmed by ALOS-1 PALSAR images acquired in 2010) and 16 cases were because the platforms construction date lagged the high resolution image photograph date. In addition, a total of 23 platforms were not correctly identified (11 were also confirmed to be omitted by the ALOS/Setinel-1 validation), and 7 cases of noise were inadvertently retained (2 were also confirmed to be commissioned by the Setinel-1 validation).
Based on the above validation, a total of 42 offshore platforms were confirmed to be omitted and 19 were incorrectly retained, resulting in an omission error of 3.88 % and commission rate of 1.76 %. After