An Efficient On-Board Lossless Compression Design for Remote Sensing Image Data

An Efficient On-Board Lossless Compression Design for Remote Sensing Image Data Guoxia Yu, Tanya Vladimirova, Martin Sweeting Surrey Space Centre University of Surrey, Guildford, UK, GU2 7XH {g.yu, t.vladimirova, m.sweeting}@surrey.ac.uk Keywords: Earth observation, On-Board, Lossless image compression Introduction Remote sensing tasks require transmission to ground of an extensive amount of imaging data with onboard image compression being the solution to the “Bandwidth versus Data Volume” dilemma of modern spacecraft [1]. The Consultative Committee for Space Data Systems - Lossless Data Compression recommendation (CCSDS-LDC) is a very low complex algorithm while having low memory and power usage. With consideration of hostile space environment, its important feature is error-resilience. To stop error propagation, a sample is periodically kept uncompressed as a reference. So data before and after a reference sample are compressed independently. Therefore error is constrained in a small region, called Independent Compression Region (ICR) [2]. Being a 2-D type of data, image can be compressed further, by using a 2-D prediction scheme, instead of the default 1-D scheme. However compression of a current pixel will depend on neighbour pixels of previous lines, which is contrary to the featured coding independency, assuming one line of data is fairly larger than one ICR, which is true for most Earth observation remote sensing tasks. In this paper, we introduce a new design, which is a combination of 2-D prediction and independency coding by using a scan scheme beforehand. This new design could increase the compression ratio around 93%, with being only slightly more complex. Its performance is better than the state-of-the-art JPEG-LS, under the same circumstance.

Proposed Scan Scheme and 2-D Prediction Normally image data are read in raster scan order as shown in Figure 1-a. On each line the first pixel is taken as the reference sample. Hence ICR is just one horizontal line of data. The scan method, shown in Figure 1-b, named Peanno-Hilbert (PH scan), is believed to be the optimum scan to reduce 2-D spatial correlation to 1-D correlation. To enable a 2-D prediction without affecting the coding independency, a new vertical scan (VS) is proposed, which has a “V” shape as shown in Figure 1-c. This scan goes down vertically, and turns from the start again after N pixels. N is 16, as it is the number of samples in the smallest compression unit. Therefore a 2-D prediction can be made using previous vertical line(s), while inside one ICR. Here Gradient-Adjusted Predictor (GAP) [3] prediction is applied in the vertical mode as shown in Figure 2. The value of the current pixel marked as a star is predicted by using two pixels above and some pixels of two previous vertical lines. The definition of vertical and horizontal gradients of intensity and the pseudo-code for GAP prediction are given in Figure 3. A 3-D extension for multispectral images is designed as well, and will be included in the full paper with its compression performance evaluation. The linear CCD image sensor in on-board push-broom imaging payloads, has different offset and shift registers, and hence difference in brightness for the even and odd column pixels. Thus less correlation between odd and even column will suppress the compression performance. In [4], a method called Brightness Difference Compensation (BDC) is reported, which is be able to bring 5.5% further data reduction on JPEG-LS. BDC is applied to images tile by tile with tile size of 512 by 512 pixels. The proposed scan scheme needs to buffer 16 lines of image data, while BDC needs 512 lines. Here by adapting BDC to the proposed scan scheme, we apply an embedded BDC, which means that it is inserted into the GAP prediction. So in order to get a better prediction, the WN, W, and WS pixel values in Figure 3, are compensated through the embedded BDC using equation 1.

Ei = Ei + (mean(Oi ) + mean(Oi −1 )) / 2 − mean( Ei −1 ) (1) E E O O where i is the current even column line, accordingly i −1 is the previous even column line, i and i −1 are the two previous odd column lines. They all consist of 16 pixels, as defined in proposed scan scheme.

Compression performance The state-of-the-art lossless compression algorithm, JPEG-LS, is compared with the CCSDS-LDC based algorithms. Here one ICR consists of 128 times 16 pixels, which applies to JPEG-LS as well. Compression results in terms of compression ratio are given in Table 1. The results show that the different scan scheme

give similar performance, but an extra 2-D GAP processing available exclusively to proposed VS, could bring significant better performance, which exceeds that of JPEG-LS. NN

… … … (a) Raster Scan

WW

W

WSW

WS

Figure 2. Vertical GAP Casual Neighbors

…

dv = abs(NN-N)+ abs(W-WN)+ abs(W-WS) dh = abs(N-WN)+ abs(W-WW)+ abs(WS-WSW) if dh-dv > 80 pre = N;

(b) Peano-Hilbert Scan

elseif d h-dv < -80 else

. . .

. . .

. . .

. . .

. . .

. . .

. . .

. . .

N

WN

…

pr e = W; pre = (N + W)/ 2 + (WS-WN)/ 4;

if d h-dv > 32

pre = (pre + N)/ 2;

else if dh-dv > 8

pre = ( 3*pre + N)/ 4;

else if dh-dv < -32 else if dh-dv < -8

pre = (pre + W)/ 2; pre = ( 3*pre + W)/ 4;

end

(c) V Scan Figure 1. Different scan scheme in the preprocessing part

end Figure 3. Vertical GAP Prediction

Table 1. Compression Ratio on Standard Test Images (RS: Raster Scan; PH: Peano-Hilbert; VS: Proposed) CR Goldhill Lena Mandrill Peppers AVE

JPEG-LS 1.57 1.74 1.17 1.62 1.53

RS+CCSDS-LDC 1.53 1.6 1.26 1.56 1.49

PH+CCSDS-LDC 1.52 1.67 1.24 1.6 1.51

VS+CCSDS_LDC 1.5 1.71 1.21 1.59 1.5

VS+GAP+CCSDS-LDC 1.64 1.77 1.3 1.67 1.6

Five panchromatic images (4m GSD, 6144 by 6144 pixels) captured from the Surrey Satellite Technology Ltd. (SSTL) ‘Beijing-1’ small satellite are selected, containing different features. The Beijing-1 panchromatic imager is of push-broom type, so we could compare performance of the proposed embedded BDC with that of its complex rival BDC. Different combinations of JPEG-LS, BDC, EmbeddedBDC, RS, VS, GAP and CCSDS-LDC are tested. Table 2 shows that the results in column 6 are comparable to the results in column 2, which are much better than JPEG-LS (column 1). By comparing the results in columns 6 and 7 it can be seen that embedded BDC only slightly underperforms BDC, however it reduces the buffer memory size 32 times. Table 2 also shows that the results of the proposed solution (column 7) are much better than the results achieved by RS and CCSDS-LS (column 3), nearly doubling the compression ratio only with an extra memory buffer of 16 lines image data and a combination of simple scan scheme and GAP prediction. It can be concluded that the proposed approach is the most efficient scheme for lossless data compression with constrained error propagation functionality. Table 2. Compression Ratio on Satellite Test Images CR

Kuwait Gloucester De Aar Bahrain Tehran AVE

JPEG-LS 1 3.48 4.49 2.84 3.47 2.71 3.4

BDC+JPEGLS 2 3.64 5.32 2.94 3.68 2.78 3.67

RS+CCSDSLDC 3 1.93 1.73 1.87 1.88 1.83 1.85

BDC+RS+C CSDS-LDC 4 3.2 4.69 2.6 3.2 2.39 3.21

BDC+VS+C CSDS_LDC 5 3.13 4.87 2.52 3.29 2.29 3.22

BDC+VS+GAP+ CCSDS-LDC 6 3.68 5.07 2.98 3.63 2.73 3.62

VS+EmbeddedBDCGAP +CCSDS-LDC 7 3.63 4.94 2.96 3.61 2.74 3.58

1. T. Vladimirova, M. J. Meerman, A. da Silva Curiel. “On-Board Compression of Multispectral Images For Small Satellites”, Proceedings of IEEE International Geoscience & Remote Sensing Symposium (IGARSS06’06), Vol. VII, pp. 3533-3536. 2. P.-S. Yeh, G. A. Moury, and P. Armbruster, "CCSDS Data Compression Recommendation: Development and Status," in Applications of Digital Image Processing XXV, Seattle, WA, USA, 2002, pp. 302-313. 3. X. Wu and N. Memon, "Context-based, adaptive, lossless image coding," Communications, IEEE Transactions on, vol. 45, pp. 437-444, 1997. 4. G. Yu, T. Vladimirova, and M. Sweeting, "A New Automatic On-Board Multispectral Image Compression System for Leo Earth Observation Satellites," in Digital Signal Processing, 2007 15th International Conference on, 2007, pp. 395-398.