320 likes | 477 Views
Processing of Multichannel RS Data for Environment Monitoring. Processing of Multichannel RS Data for Environment Monitoring VLADIMIR LUKIN Dept of Transmitters, Receivers and Signal Processing, National Aerospace University, 17 Chkalova St., Kharkov, 61070, Ukraine,
E N D
Processing of Multichannel RS Data for Environment Monitoring Processing of Multichannel RS Data for Environment Monitoring VLADIMIR LUKIN Dept of Transmitters, Receivers and Signal Processing, National Aerospace University, 17 Chkalova St., Kharkov, 61070, Ukraine, tel. +380577074841, e-mail lukin@xai.kharkov.ua Presentation outline 1. Applications of multichannel remote sensing 2. Problems of data offering to potential customers 3. Some aspects of automatic image pre-processing 4. Possible Strategies of On-board/on-land Processing and Compression 5. Peculiarities of noisy image lossy compression 6. Comparison of strategies’ performance 7. Classification accuracy of processed multichannel data 8. Conclusions Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
Potential customers of RS data are: governmental boards, nature protection organizations, space agencies, marine traffic services, meteorologists, agriculture and forestry, etc. All require more reliable information its more operative providing to them offering of data in the most convenient form More reliable information can be provided by multichannel (dual and full polarization radar, multi and hyperspectral) RS systems. But how to meet two other requirements? Applications of multichannel remote sensing Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
Information content of RS data and effectiveness of solving final tasks of environment monitoring depend upon many factors: information content of original (raw) data determined by RS operation mode, range of wavelengths covered by an imaging system, number of its channels and spatial resolution, noise level and statistical characteristics of the formed images; adequateness of noise models and/or availability of a priori information about model parameters; effectiveness of the methods used for RS data processing where by processing here we mean a wide set of operations that, depending upon application, might include evaluation of noise characteristics, filtering, compression, registration, geo-referencing, calibration, classification, interpretation, etc. General information Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
Block structure of research Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
It is practically impossible to give one strict recommendation what strategy of data pre-processing and offering is the best and to give a full description of possible approaches. Below we concentrate on possible strategies and stages of multichannel RS data processing. First, processing can be carried out on-board, on-land, or, in general, both. Second, before transmission data are to be compressed and we focus on lossy compression since even the most powerful techniques of lossless coding are nowadays unable to provide a compression ratio (CR) larger than 3.5…4 and this is often not enough for practical applications due to downlink channel limitations. Third, we insist that automation of data processing should be applied as possible at all stages (surely on-board and desirably on-land). Problems of data offering to potential customers Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
Some aspects of automatic image pre-processing The methods for blind determination of noise type and parameters (variance of additive noise, variance of multiplicative noise, impulse noise probability) that provide appropriate accuracy have been designed. X-band SLAR image (add=8; =0.09) Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
Some aspects of automatic image pre-processing We have also proposed blind methods for estimation of spatially correlated noise characteristics. The corresponding DCT based filters have been designed. This allows increasing PSNR of filtered images by 2...3 dB. а b Original L-band SAR image (а) and the obtained output image taking into account the estimated σ2аdd=14 andσ2μ=0,15 (b) Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
Pre-requisites of joint image processingPreliminary co-registration and geo-referencing Registration and correction of multichannel radar images: а) Ka-band HH SLAR image with marked control points; b) Ka-band VV SLAR image with marked control points; c) Co-registered images using affine transform without geometric correction; d) Images co-registered using nonlinear slant range correction and geometric transform b) a) d) c) Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
Pre-requisites of joint image processingPreliminary co-registration and geo-referencing a) b) X-band SLAR image of large size (more than 3000х3000 pixels) before (а) and after (b) slant range correction and nonlinear registration to topology map using 10 control points Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
Properties of hyperspectral (AVIRIS) data Noisy sub-band image Denoised sub-band image Almost noise-free image Estimated SD of noise in sub-band images Inter-channel correlation factor Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
There are, at least, three possible strategies for on-board/on-land processing and lossy compression of multichannel RS data. Each of them can exploit component-wise (sub-band, each channel separately) and 3D (grouped, vector-like) processing (pre- or post-filtering) and compression. Strategy 1: a multichannel image is a subject to lossy compression without pre- and post-filtering. The first (on-board) stage is blind evaluation of noise variance. It is applied component-wise with obtaining a set of estimates of noise standard deviations (SDs) where N denotes a number of components of multichannel image. Then the component (sub-band) images can be either grouped or compressed separately. Possible Strategies of On-board/on-land Processing and Compression Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
If grouping is applied, it is carried out with taking into account two rules. First, each group should contain either 4, or 8, or 16 sub-band images (for hyperspectral data like AVIRIS). Second, less strict rule, is that images with indices are grouped if standard deviation estimates for these component images do not differ a lot, for example, if (1) If grouping is not used, each component image is compressed with setting a 2D coder quantization step equal to where C1 is a parameter. If grouping is used, then a 3D coder is used where quantization step for it is determined as for each q-th group. Details of Strategy 1 Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
As a 2D coder we propose to apply the coder AGU (http://ponomarenko.info) . This coder is based on discrete cosine transform (DCT) in 32x32 pixel blocks, more advanced probability models, and image deblocking after decompression. This coder outperforms most wavelet-based coders and it has been modified to 3D case. One more advantage of this coder is its relative simplicity, a parameter controlling CR is quantization step. To provide fast implementation the aforementioned condition of sub-band grouping by 4, 8, or 16 channels has been introduced. An advantage of the strategy described above is its simplicity. The only operation to be done before lossy compression is automatic estimation of noise standard deviations in sub-bands. Another advantage is that this strategy provides CR from 4.5 to 9 for component-wise compression and from 8 to 25 for compression with adaptive sub-band grouping for 224-channel AVIRIS data due to incorporating inter-channel correlation inherent for hyperspectral data. Details of Strategy 1 Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
Consider a test 8-bit image corrupted by additive noise (added artificially) with variance 200. It has been compressed by three coders: JPEG, JPEG2000, and AGU. Since we had the noise free image, it was possible to calculate PSNRnf of decompressed image with respect to the noise free image. Analysis of these curves shows that there exists optimal operation point (OOP) – such bppOOP for which PSNRnf reaches maximum. This means that for OOP neighborhood lossy compression provides image enhancement due to filtering effect. Background of Strategy 1 Dependences PSNRnf vs bpp for different coders without and with pre-filtering Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
Visual Example Helsinki area noisy satellite map (σ2=100) and decompressed image for our compression technique (bpp=0.75) Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
Lossy compression of 224-channel AVIRIS data MSE dependences for 3-D and component-wise compression. Quantization steps for component-wise (М1) and 3-D compression (М2) in groups The designed automatic methods of analysis and compression allow providing CR=15…30 (2 times larger than for component-wise compression)with less distortions than for component-wise compression. Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
Hyperspectral AVIRIS data compression 125-th channel of Lunar Lake: original image (left) and compressed image with QS=500 (PSNR=36,43 dB), distortions are not seen. Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
Lossy compression of hyperspectral AVIRIS data 113-the channel of Lunar Lake: original image (left) and compressed images with QS=500 (center, PSNR=16,51 dB, huge distortions) and with QS=20 (right, PSNR=33.42 dB, distortions are appropriate). Thus, QS should be adapted to channel image noise SD and dynamic range Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
Although such lossy compression performs some denoising, such denoising is not perfect in the sense that advanced filtering techniques are able to carry out noise removal considerably better. It can be difficult to further improve decompressed data quality on-land by post-filtering. The strategy described above is based on assumption of additive noise model although recent investigations (Barducci et al., 2005) show that noise model can be more complex. Drawbacks of Strategy 1 ) noise model Dependencies MSEnf vs QSn for the mixed ( Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
Strategy 2: a multichannel data are compressed in near-lossless manner with accounting noise characteristics of component images. The first stage of RS data processing is blind evaluation of . Then the data can be either compressed component-wise or grouped as described above and compressed using 3D version of the AGU coder where for component-wise compression and for each q-th group for adaptive grouping based compression. C2 is considerably smaller than C1 for the strategy 1. We recommend using C2≈1.3. Two possible alternatives to Strategy 1 Typical Dependences PSNRnf vs bpp and PSNRnf vs QS/σ(n) for lossy compressed/decompressed and then filtered images Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
Drawback: The strategy 2 produces sufficiently smaller CRs than the strategy 1, namely, from 3.0 to 4.6 for component-wise compression and from 4.6 to 7.6 for compression with grouping (for conventional test images Moffett Field, Cuprite Mine, Lunar Lake, Jasper Ridge). The main advantage of the strategy 2: it provides practically full potential for consequent effective filtering of decompressed images on-land where resources are not so limited as on-board and filtering can be carried out in a better way. Upon user’s request, both almost original and filtered RS data can be offered to Customers. Properties of Strategy 2 Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
Strategy 3: filtering is carried out on-board, then pre-processed data are compressed in a lossy manner, and transferred by downlink communication channel. On-land these data can be either disseminated or stored in compressed form, or decompressed and further processed. The first stage of processing is again blind evaluation of noise standard deviations in sub-bands for two purposes. The first one is the use of the obtained estimates for component-wise filtering of images. The second purpose is to set a coder quantization step as for component-wise compression or as for each q-th group for the adaptive grouping based compression. C3 is a parameter recommended to be approximately equal to 1.5. Two possible alternatives to Strategy 1 Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
The values of CR provided by the strategy 3 for AVIRIS images are from 3.2 to 5.4 for component-wise compression and from 5.1 to 9.6 for compression with adaptive grouping (larger than for Strategy 2 but smaller than for Strategy 1). Advantage of the strategy 3: it produces higher quality of images than the strategy 1. The simultaneous advantage and the drawback of the strategy 3 is that it provides already filtered multichannel images (an insight on this property depends upon user’s priority of requirements). One more drawback of the strategy 3 is that it requires more resources and time for on-board data processing than two other strategies since filtering of multichannel images is to be done. Properties of Strategy 3 Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
Experimental Results Table 1. CR for the methods of strategy 2 Table 2. CR for the methods of strategy 3 The provided CR varies depending upon a content of hyperspectral data and the automatically achieved CR is commonly smaller for images that contain more details and texture. Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
Experimental results for Strategy 1 Table 1. Performance characteristics of the compression methods M1 and M2, C1 =4.5. Table 2. Performance characteristics of the compression methods M1 and M2, C1 =5.5. Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
One can argue that providing minimal MSEdecnf (or maximal PSNRnf) does not guarantee the best solving of final tasks of RS data processing like classification, object and anomaly detection, etc. The studies have been carried out for a three channel Landsat image artificially corrupted by pure additive noise. There were five classes. Neural network (NN) and support vector machine (SVM) classifiers have been applied and provided similar results Classification of compressed multichannel images The test three-channel image in RGB representation and the corresponding classification map for noise-free multichannel data Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
Classification of compressed multichannel images Two examples of the partition schemes obtained for the DCT PS coder Examples of partition schemes obtained for the first channel for QS=4.5σa=45 and QS=2.5 σa=25 Performance of the coder DCT PS for noise variance 100 Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
Classification of compressed multichannel images Classification map for three-channel test image formed by LandSat: ground truth data for evaluation of classification accuracy; preliminary classified image fragments for classifier training Image classification is carried out on pixel by pixel basis. For the test image the feature vector of i-th and j-th pixel is composed of brightness values from R, G and B image components Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
Classification of compressed multichannel images Classification results (Pcorr for different classifiers, noise variances and coders) Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
Classification results Preliminary conclusion: RS data compression with providing OOP simultaneously produces an aggregate Pcorr close to maximum. The plots of probability of correct classification (in %) vs quantization step for the following classes: 1 – Bare Soil (red), 2 – Grass (green), 3 - Water (dark blue), 4- Roads and Buildings (yellow), 5 – Bushes (light blue), σa2=100 Conclusion: in aggregate, the choice QS = 5σa seems reasonable for providing close to largest possible Pcorr for all classes Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
Classification results (a) (b) The classification maps (a) for the noisy image and (b) the optimally compressed image σa2=100 Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841
Three different strategies for automatic processing and compression of multichannel RS images have been described. The recommendations concerning parameter selection for them have been given. The advantages and drawbacks of these strategies have been considered. The simplest strategy (without any filtering) is analyzed more in details. With the proposed modifications (the use of the corresponding HT) it can be applied if component images are corrupted not only by additive but also by multiplicative or signal-dependent (Poisson-like) noise. The relationship between compressed data quality in terms of PSNR and classification accuracy is considered. It is demonstrated that attaining of high PSNRnf results in providing probability of correct classification close to maximal. We plan to consider classification of hyperspectral images and to exploit more sophisticated models of noise. Conclusions and future work Vladimir Lukin lukin@xai.kharkov.ua +38 057 7074841