Main Processes for OVS-1A & OVS-1B: From Manufacturer to User

Commercial remote sensing has boosted a new revolution in traditional processing chain. During the development of OVS-1A and OVS-1B, we construct the main processing pipeline for ground and calibration system. Since these two satellites utilize colorful video imaging pattern, the underlying video stabilization and color adjustment is vital for end user. Besides that, a full explanation is given for researchers to shed light on how to promote the imagery quality from manufacturing satellite camera to generate video products. From processing system, the demo cases demonstrate its potential to satisfy end user. Our team also releases the possible improvement for video imaging satellite in the coming future.


Introduction
The OVS-1A and OVS-1B video satellites (also known as ZHUHAI-1 01 and ZHUHAI-1 02 [1]) were launched on June 15, 2017, and the two equipped cameras can capture 10-bit image or 8-bit video sequence at 20 fps both converting from Bayer formatted data.
After one year running around, we feel honorable to reveal the main processing methods used, which may be different from general close range community, for end user.
In the past few years, mini satellite for commercial user is largely developed from overhead video imaging which can be recently dated back to typical SkySat [2]. Nowadays, the surveillance need for local interesting event is boosted in Given a tight budget, Beijing Institute of Space Mechanics & Electricity (shown as BISME hereinafter) undertook the contract in 2016 from Zhuhai Orbita Aerospace Science & Technology Co, Ltd. (available at http://www.myorbita.net/), which must be finished in one year. Our engineers inherit some effective design from industrial applications and furthermore proposed some techniques to promote higher quality for standard product. Figure 1 and Table 1 show the appearance and system characteristics of these two satellite cameras. Opposite to the combined imaging plane of SkySat, the platform chooses to rotate to hold the nadir viewport all the time, and a common 90 seconds duration video sequences are stored on board or directly transferred to ground station.  From the working mode in Figure 1, one thing should be noted the viewing angle always varies, and bad effect also comes from the platform vibration at 550 km, even its absolute value is small, this is why we carefully designed the stabilization method.
To the radiometric calibration, a classical and uniform region is not practical any more [4], and statistical analysis for line sensor fails the area plane detector, and the real value of each pixel should be computed individually.
A subset of real image covering Cape Town is given in Figure 2. As a first impression, we list all the techniques embedded in this paper: 1) How to get high level radiometric calibration result for area device; 2) How to make change for Bayer formatted pixel to get an uniform image; 3) How to calculate the color transformation for true RGB; 4) How to implement video stabilization and approximately still video; 5) Geometrical processes for video sequences.

Radiometric Model
Before going to the main process, the linear model should be given for imaging ( ) r λ -transmittance of optical system; 1 λ and 2 λ -range of wave length.
In most cases, the non-uniform noise in image mainly comes from the gain and offset differences of every pixel.

Main Processing
From the calibration images, the normal no-uniformity coefficient can be calculated. Then stack the surface image at high and low digital value to generate two virtual images, a linear fitness of these two images modifies the initial NU coefficients. After that, we utilize the radiometric parameter to enhance image, generate L1 to L2 and video product. Figure 3 shows the main pipeline of following steps.

No-Uniformity Correction
Using the initial calibration data, the top of atmospheric no-uniformity coefficient for raw data can be drawn as:   (6) Above processes fit response of every pixel to the statistical mean value, and generate the initial gain and offset coefficients Figure 4 shows the correction result in lab.
Specially, for Bayer image, there should be two normal approaches for RGB channels calibration separately: 1) interpolate Bayer to virtual RGB, calibrate RGB separately; 2) pick real RGB from Bayer, calibrate them and demosaic to get final result.
We found the latter one is a physical and strict data reconstruction method which generates better result. In one word, Bayer image cannot be interpolated to RGB before radiometric calibration. Figure 5 illustrates the specific operation.  After the launch, non-uniform noise still residues in real image, and it can be easily attributed to atmosphere or some other transfer jitter. Actually, with the help of our former work [7], we stacked the high DN image from mid latitude cloud and low level image covering nightly ocean, and got near 3% radiometric result after a second calibration. Figure 6 outlines the strengthened result from our method.

Color Adjustment and Image Dehaze
Excellent results can usually be achieved with the inexpensive, widely-available 24-patch X-Rite color checker. The CCM (color correction matrix) can be applied to images to achieve optimum color reproduction, which is defined as the minimum mean squared color error between a corrected test chart patches and the corresponding reference values, which can be the standard (published) values.
From the radiometric model, the R, G, and B values of each output (corrected) pixel is mathematically a linear combination of the three input color channels, Equation (7) shows the recommended 3 3 × CCM A and its usage, where the P is processed pixel and O is the original color value: And also, 4 3 × CCM usually works fine, please refer to [8] for more details.

Video Stabilization and Registration
Due to the mechanical vibrations caused by the engine and the effect of turbulent    We noticed a parametric strategy in IPOL (www.ipol.im) with open source implementation [10], and used feature-based motion (homography) estimation with local matrix-based motion smoothing, and Figure 9 shows the basic framework.
For further approximately still video, we choose a registration method based on feature point to generate such product. SIFT [11] or ASIFT [12] feature descriptor can be used directly after receiving the stabilized sequences. Figure 10 shows the processed result for Marseille.

Video Compression
Although the compression standard is common in close range community, we got response from end user the compression always damage the expert analysis for small object, especially the high profile H.264 standard (it will incur small blocks near the edge and is of various compression ratios).
Hence, we limited the compression ration no more than 10 for interested region (or quality test video) and no more than 30 in other regions. Figure 9. Illustration of the motion smoothing and imge registration.

Geometric Processing
After building RPC file using attitude and GPS data, we can construct the L2 product mapped into a cartographic projection and corrected from sensor and terrain distortions. The mean value of geometrical location error is 500m after system correction and calibration (small satellite without high level star sensor).
Mosaic products will be automatically processed in order to provide end user with an ortho-image of larger size, this is vital for multi footprint mode.

Module Transfer Function Value and SNR
We do not pave the traditional knife edge pattern (high cost) but find some natural object such as road or building to test the MTF value, showing in Figure 11.

Video Stabilization Level
To quantify our video stabilization method, we carefully detect the matched feature points and calculate the residual error of every frame. In Figure12, it can be seen the average of error distance is reduced from largest 5 pixels to no more than 1 pixel in every frame after our processing, which prove its effectiveness (please go to Table 3). Journal of Computer and Communications   One thing should be kept in mind, this approach will not be used in video sequences with rare feature point, such as small ocean island, desert region.

Conclusions
OVS system is designed hoping to optimize the end user's need. Its main image quality performances are ensured by appropriate processing line in end user's production center (we worked on it for months).
Despite the success or progress, we noticed: 1) There are potential drawbacks to using color filter array (CFA) sensor such as Bayer filter, since CFA inherently lowers the resolution of the underlying staring sensor relative to a single spectrum image [2]. We are now designing more practical and effective system for overhead video system; 2) Video mode remote sensing lacks universal framework to guide how to construct the product, and end user usually cannot get the longitude or latitude from video product from any video player. We anticipate a full analysis to get an engineering product definition in the coming soon future.
After OVS-1A and OVS-1B, we successfully designed the TDI mode for area detector and extend the data picturing mode to traditional long strip.
Last but not least, Orbita has launched its newest video satellite with 0.9m resolution in 2018 July, and it can also be seen as another symbol for developing consumer remote sensing.