Ship extraction and categorization from ASTER VNIR imagery

We present a methodology for ship extraction and categorization from relatively low resolution multispectral ASTER imagery, corresponding to the sea region south east of Athens in Greece. At a first level, in the radiometrically corrected image, quad tree decomposition and bounding rectangular extraction automatically outline location of objects - possible ships, by statistically evaluating spectral responses throughout the segmented image. Subsequently, the object borders within the rectangular regions are extracted, while connected component labelling combined by size and shape filtering allows ship characterization. The ships’ spectral signature is determined in green, red and infrared bands while cluster analysis allows the identification of ship categories on the basis of their size and reflectance. Additional pixel- based measures reveal estimated ship orientation, direction, movement, stability and turning. The results are complemented with additional geographic information and inference tools are formed towards the determination of probable ship type and its destination.


INTRODUCTION
Among an extensive collection of applications, multispectral satellite imagery is used for recoding biophysical parameters and monitoring the ocean and coastal environment [1][2][3][4]. The detection and parametric representation of objects from satellite imagery within the sea environment is of utmost importance for recording both static (small rocky islands and reefs) and dynamic objects (ships). Other applications include ship related pollution, such as oil slicks, sea traffic [5], coastal management, fishery monitoring, and defense [6]. Nowadays, micro-satellite platforms are equipped with a combination of both active (microwave) and passive (multispectal) sensors [7]. Microwave remote sensing systems (Seasat, ERS-1, JERS-1, ERS-2 and RADARSAT-1) have been extensively used for ship detection [8][9][10].
In the multispectral area, ship detection and further classification based on type and size requires very high resolution imagery [11,12] and provides a relatively even sea reflectance according to the spectral band. Accordingly, pixel patches corresponding to ships provide large traces for ship categorization [13] given a pixel size of about 1m. The dynamic nature of sea traffic is most of the times addressed through multi-temporal imagery forming sea traffic maps [5]. Generally speaking, SAR oriented procedures are mainly used even though they are prone to noise and wave formations, while passive systems are vulnerable to various weather conditions. Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) sensor, emerged from a cooperative effort between NASA and Japan's Ministry of International Trade and Industry [14] that obtains high resolution (15 to 90 m) images of the Earth in the visible, near-infrared (VNIR), shortwave-infrared (SWIR), and thermal infrared (TIR) regions of the spectrum. ASTER data products include: spectral radiance and reflectance of the Earth's surface, surface temperature and emissivity, digital elevation maps from stereo images, surface composition and vegetation maps, cloud, sea ice, and polar ice products, etc [15,16]. The VNIR subsystem operates in three bands (green, red and near-infrared) with 15 m resolution and a 60 km swath width [17]. The revisiting period is about 20 days.
The aim of this research is to present a methodology for ship extraction, recognition, categorization and dynamic information inference from ASTER VNIR imagery exploring its applicability with respect to ship detection. Such an effort aids the ship related applications, such as sea traffic, expected behavior, ship type classification, etc. The challenges include low resolution processing according to the size of the ships, deduction of dynamic information from a static image and formation of inference tools.

ORIGINALITY AND CONTRIBUTION
The methodology presented includes ship extraction, recognition, categorization and dynamic information inference from ASTER VNIR imagery. The novelty of the proposed processes is described as follows: • To the best of our knowledge, this is the first attempt for ship extraction through ASTER imagery.
• The major challenge is to derive information from medium resolution images compared to ship size and shape.
• The image processing based methodology presented, involves algorithmic variations, including a first step generalized radiometric segmentation, automated thresholding criteria, and shifted search iterations, to compensate for regionally differentiating sea illumination and avoid misclassification from non-ship objects. • The combination of the image processing ship categorization with geographic information systems, provides inference results, including possible destination and ship type from a single snapshot.
Overall, the novelty is summarized as the deduction of dynamic information and inference from a static image.
Application-wise, image exploitation of modern, medium resolution, multispectral satellite imagery, impacts the fields of ship traffic monitoring with probable applications in environmental protection and safety. Such an effort also aids in expected behavior and type classification of ships. It could serve as a parallel process beyond SAR oriented ship extraction, which is prone to noise and wave formations.
Based on satellites' revisiting periods which vary according to their orbits, dynamic information from stable snapshots reveals tendencies and fills temporal gaps. The presented methodology proves readily applicable to satellite imagery of relevant resolution, and is customable to satisfy additional resolutions under minor alterations.

METHODOLOGY
Ship detection and categorization methodology is discussed in the remaining of the paper through the following steps: 1. Pre-processing of the satellite imagery including destriping and reflectance value estimation; 2. Quad tree decomposition -segmentation of the image compensating for varying sea reflectance and targeting for the subsequent image processing algorithms; 3. Possible ship region extraction through bounding rectangular patches based on background sea and ship reflectance, prone to non-ship object inclusion; 4. Ship detection and extraction from the rectangular regions and further categorization based on shape and reflectance; 5. Additional ship measures and classification of dynamic nature; 6. Inference logic through integration of additional geographic related information.

Study area and data
The satellite imagery was acquired on the 11th of December 2001 from the US Geological Survey during the one-year evaluation period of ASTER characterized by processing level 1A (no radiometric and geometric corrections were applied). The latitude and the longitude are within the range 37.4878 to 38.1428 and 23.1977 to 24.0596, respectively.

Pre-processing of ASTER imagery
First, radiometric corrections were applied. Each band was destriped using the histogram matching technique and digital values were converted to radiance on the basis of the corresponding band gain [18]. The ASTER imagery was georeferenced and geometrically corrected by the use of orbital data available in the image header ( Figure 1). The very near infrared band (band 3) was selected for ship detection due to the low reflectance of water bodies in the infrared portion of spectrum [19].

Sea region decomposition
Quad tree based decomposition [20,21] divides the sea region into sections which are tested for the uniformity of their spectral response (

Ship ext
The ship's bo transition fro computed fro degrees in the upper limit is nding rectangl ip size. The frequenc middle. Thre The pixels w ( Figure 5 righ of backgroun algorithm [2 parametrically 1, 2 and 3. In

Ship clu
The spectral for the identi analysis is a m of distance a represents the priori definiti      The results of F statistic ar comparing d differentiation ) e e e f r Table 3. Distances between cluster centroids (divergence).

Ship dynamic information categorization and inference
Based on the extracted ships from the previous step (figure 7a), a series of additional processes are performed towards ship categorization under various criteria. These procedures determine dynamic -time related information from a static image and are prone to subjectivity.
Beyond size, a general shape measure is derived from the elongation (eccentricity) of the object in question, in order to determine whether it is a ship or e.g. a platform. For the specific image scale, it is highly unlike to discern the rear from the front of the ship [13]. Based on the assumption that the ship's wake would be a thin tail of high reflectance (figure 7a and c), the ship's course line can be automatically determined. In addition, each ship's minimum bounding rectangle (MBR) is determined and orientation-azimuth is estimated through the principal eigenvector of the ship pixels (figure 7e).
Three different ways for the determination of possible orientation are devised. First, the displacement between the center of gravity and center of the MBR reveals the possible orientation of the ship's heading (figure 7f), since the wake's trace would be thinner than the actual ship (eccentricity). Second, two snapshots (figure b and d) of the ship using different grayscale thresholds, indicate the direction as seen in figure 7g. Finally, assuming a bright ship compared to a less bright wake, the concentration of bright pixels in one side implies its direction (figure 7h). These three direction measures enhance the overall determination confidence (figure 9a). In the case that the actual area of the ship-wake snapshot compared to the convex hull area is relatively small, turning of the ship is implied, as seen in figure 8.
Up to this point, ship related information is summarized as: general shape, size, stability, mobility, turning, course line, orientation. Inference procedures can be created in a GIS environment using the measures discussed so far, complemented by additional information. Towards this objective, we form the following inference tools: • Small static ships (size) close to the shore (distance to the shore) and of relatively high population (density measure), can be considered as fishing ships [30], as seen in figure 9c on the Northern area, pointed with red circles.
• The ship's course directions provide an estimation of the ports' expected traffic (figure 9 b) and may reveal natural or local ports, as seen on figure 9c.
• Lines of direction for non turning ships, away from the shore, and aligned to the digitized ship routes acquired from the ferries' companies (red dotted lines in figure 9c), along with the ship size can reveal possible destination, even if the destination is not included in the images. Categorization of destinations for the study area, may include the island of Crete (figure 9c) or Islands on Central Aegean or even local ports.

DISCUSSION OF RESULTS
Bounding rectangular selection provided an initial approximation of possible ship location and since SOBEL was applied in the bounding rectangles, the effect of local high frequency noise in seed selection was reduced. Additionally, the SOBEL operator allowed a data driven (thresholding of the slope frequency histogram on the basis of a trial and error