This is the accepted version of the article: García-Sobrino, Joaquín; Pinho, Armando J.; Serra Sagristà, Joan. «Competitive Segmentation Performance on Near-lossless and Lossy Compressed Remote Sensing Images». IEEE geoscience and remote sensing letters, (August 2019). DOI 10.1109/LGRS.2019.2934997 This version is avaible at https://ddd.uab.cat/record/215774 under the terms of the license IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. XX, NO. Y, MONTH Z 2019 1 Competitive Segmentation Performance on Near-lossless and Lossy Compressed Remote Sensing Images Joaquı́n Garcı́a-Sobrino, Armando J. Pinho, Member, IEEE, and Joan Serra-Sagristà, Senior Member, IEEE Abstract—Image segmentation lies at the heart of multiple to the rich information contained in the images. Several image processing chains, and achieving accurate segmentation of the existing segmentation algorithms rely on two main is of utmost importance as it impacts later processing. Image traditional strategies: 1) region-based methods, which depend segmentation has recently gained interest in the field of remote sensing, mostly due to the widespread availability of remote on region information, and 2) edge-based methods, which look sensing data. This increased availability poses the problem of for boundaries between objects to find them [5], [6]. Both transmitting and storing large volumes of data. Compression is strategies have proven to achieve competitive segmentation a common strategy to alleviate this problem. However, lossy or results. near-lossless compression prevents a perfect reconstruction of the However, improving the quality of the acquired data has recovered data. This letter investigates the image segmentation performance in data reconstructed after a near-lossless or a also led to increase the size of the collected data, requiring lossy compression. Two image segmentation algorithms and more means for transmission and storage. Data compression two compression standards are evaluated on data from sev- is an effective way to reduce the volume of the collected eral instruments. Experimental results reveal that segmentation data. Nonetheless, achieving high compression ratios usually performance over previously near-lossless and lossy compressed entails some level of distortion in the recovered scene, which images is not markedly reduced at low and moderate compression ratios. In some scenarios, accurate segmentation performance can may compromise the usefulness of the reconstructed data and be achieved even for high compression ratios. prevent achieving good segmentation results. Texture blurring and unclear edges between objects can be a consequence of Index Terms—Remote sensing data, image segmentation, lossy compression, near-lossless compression, maximum likelihood, near-lossless and lossy compression. These effects can make successive band merging, JPEG 2000, JPEG-LS. image segmentation a major challenge, because of difficulties in detecting homogeneous regions or boundaries [7]. The impact of a near-lossless or lossy compression stage on I. I NTRODUCTION subsequent processes is lately receiving a lot of attention: in the scope of classification applications, different strategies to I MAGE segmentation is generally defined as the process of partitioning an image into regions (or segments). The un- derlying idea consists of dividing an image into homogeneous perform classification on previously compressed images were presented in [8]–[11]. Results revealed that the classification segments by grouping neighbouring pixels that follow a sim- accuracy was still reliable after the compression stage, even at ilarity criteria. The main purpose is to estimate the structure high compression ratios. Similar conclusions were yielded for of the scene and to extract image objects and boundaries for anomaly detection [12], [13] for linear spectral unmixing [14], subsequent object-based analysis. [15], and for statistical retrieval algorithms [16]–[18]. To the In the field of remote sensing, image segmentation is an best of our knowledge, a similar study on the impact of essential step of low-level image analysis and is usually per- compression on image segmentation has not been presented. formed in the first stages of a processing chain [1]. Efficiently The aim of this work is to investigate the performance of im- segmented images are required in a wide variety of applica- age segmentation in remote sensing images reconstructed after tions, such as object detection [2], land cover classification [3], a near-lossless or a lossy compression. This analysis, carried or landscape change detection [4]. out in an early stage of the processing chain, can determine the Advances in remote sensing technology have led to im- usefulness of the data for subsequent processing (classification, prove the quality of the harvested data in terms of spectral detection, ...). Segmentations on the reconstructed scenes are and spatial resolution, and allowed to reveal more ground evaluated with respect to segmentations on the original scenes details. Segmentation results have thus been improved due to determine the impact of compression. For near-lossless compression, we employ the JPEG-LS Manuscript received August 10, 2019. standard [19], which has been widely used in compression JGS and JSS are with the Group on Interactive Coding of Images, of high-resolution remote sensing data due to the speed and Universitat Autònoma de Barcelona, Spain. Web: http://www.gici.uab.cat. E- mail: joaquin.garcia.sobrino@deic.uab.cat; Joan.Serra@uab.cat. efficiency provided. For lossy compression, the JPEG 2000 AJP is with the Information Systems and Processing Group (ISP), Institute standard [20] is employed, because it supplies a set of fea- of Electronics and Informatics Engineering of Aveiro (IEETA), Universidade tures that are important in the field of remote sensing, such de Aveiro, Aveiro, Portugal. Web: http://www.ieeta.pt. E-mail: ap@ua.pt. This work was supported in part by the Spanish Ministry of Economy as competitive low bit-rate performance, progressive quality and Competitiveness and by the European Regional Development Fund image transmission, random codestream access, etc. These under Grant RTI2018-095287-B-I00, by the Catalan Government under Grant features make both JPEG-LS and JPEG 2000 standards highly 2017SGR-463, by Portuguese Funds through the FCT - Foundation for Science and Technology, under the project UID/CEC/00127/2013, and by the attractive in the context of remote sensing applications. Santander Bank under Grant Program ”Ibero-America Santander Research”. For the segmentation stage, two segmentation algorithms are IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. XX, NO. Y, MONTH Z 2019 2 employed: Maximum Likelihood (ML) [21] and Successive TABLE I: Characteristics of the datasets. Number of scenes Band Merging (SBM) [22]. ML is a simple segmentation per instrument and sizes are reported. technique, whose low computational complexity is suitable for Average size remote sensing imagery. It has become an accepted estimate Number of Instrument (spectral channels × used as the baseline for more complex algorithms [23], [24]. scenes rows × columns) SBM is a recent segmentation algorithm specifically designed AVIRIS 2 212 × 160 × 160 for multispectral remote sensing images. Landsat 8 1 8 × 800 × 500 The remainder of the letter is organized as follows. Sec- QuickBird 20 4 × 1052 × 1182 tion II describes the compression techniques and the segmen- ROSIS 2 103 × 642 × 416 tation algorithms used in the experiments. Section III presents the data collection and the parameter setting. Section IV reports the experimental results. Conclusions are drawn in A. Data Collection Section V. To conduct the experiments, 25 images from four different remote sensing instruments were studied. We selected images II. M ETHODS and ground truth data devised for the testing of segmentation This section describes the segmentation techniques and the algorithms. In order to study a set of representative data, compression standards used in the experiments. the tested datasets included images with different spectral and spatial resolution, dissimilar number of spectral channels, obtained from different instruments, and scanned on distinct A. Segmentation Algorithms scenarios. The scenes were captured over rural and urban areas For image segmentation we use two methods: an algorithm and contain buildings, roads, grass, trees, crops, water, and based on the well-known ML approach and a novel method railways, among other objects. Table I provides the charac- based on Markov Random Fields (MRF), the SBM algorithm. teristics for the datasets. Most of the images are public and Both methods have shown good performance on several re- available at [26], [27]. mote sensing datasets [22], [25]. The ML algorithm computes the probability that a specific pixel belongs to a particular region. The algorithm assumes B. Experimental Settings and Software that the statistics for each region in the scene follow a normal For near-lossless compression, the maximum absolute error distribution. Each pixel will be part of the region that has the per sample introduced in the data was controlled using seven highest probability, considering that, if the highest probability Peak Absolute Errors (PAE), δ ∈ {1, 3, 7, 15, 31, 63, 127}. falls below a given threshold, then the pixel will not be Higher values of the PAE originate higher compression ratios, assigned to any region. SBM is a segmentation algorithm but the level of distortion introduced is also increased. For devised to work with multispectral remote sensing images. The lossy compression, the overall bit-rate was controlled using six segmentation process is carried out through three stages. In the target bit-rates, distributed between 4 and 0.1 bits per pixel per initial step, the maximizer of the posterior marginals (MPM) component (bpppc). It is expected that providing an analysis is estimated. In the second stage, contextual information of the signal in terms of frequency components benefits the is included in a nonparametric way. Finally, each pixel is coding. Therefore, the wavelet transform was used to provide a assigned to a region in the scene [22]. more decorrelated and compact representation of the signal. In the experiments, six levels of the Discrete Wavelet Transform (DWT) were applied in the spatial dimension to exploit the B. Compression Standards spatial redundancy present in the scene. For image compression we consider two paradigms: near- The segmentation stage was carried out on the original lossless and lossy compression. On the one hand, near-lossless and on the reconstructed data. Both the ML and the SBM compression is used to control the maximum absolute error algorithms were evaluated on the four datasets. To start the per sample introduced in the reconstructed data. On the other segmentation process, the selected algorithms needed an initial hand, lossy compression is employed to control the overall segmentation. All experiments started from an initial segmen- bit-rate. Two standard coding techniques are employed in tation obtained from k-means [28]. Default options were used the experiments: JPEG-LS for near-lossless compression and for the remaining parameters. JPEG 2000 for lossy compression. JPEG-LS is a lossless and To achieve a fair comparison, the parameter configuration near-lossless compression standard that provides simplicity, was kept constant for all compression and segmentation ex- low computational complexity and memory requirements. periments, in such a way that we used suitable settings for the JPEG 2000 is an international standard that provides an characteristics of all the images. The software implementations extensive range of features in a single compressed bit-stream employed for compression were: JPEG-LS software [29] for for a large amount of applications, such as remote sensing JPEG-LS and Kakadu software [30] for JPEG 2000. The imagery, medical diagnostic imaging, and mobile communi- segmentation stage (ML and SBM algorithms) was carried out cations, among others. using the Markov-Models toolbox [31] employed in [22]. III. DATA AND E XPERIMENTAL S ETTINGS IV. E XPERIMENTAL R ESULTS This section presents the datasets and the parameter config- This section presents the experimental results. The proposed urations used in the experiments. sequential chain includes first a compression stage. At the IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. XX, NO. Y, MONTH Z 2019 3 TABLE II: Region and content of the scenes. Instrument Identifier Region Content AVIRIS Indian Pines Rural area (Indiana) Alfalfa, corn, grass, wheat, woods, and buildings, among others Landsat 8 Humid Pampas Rural area (Argentina) Wildfire, corn, fallow land, and water, among others QuickBird zh11 Urban area (Zurich town) Roads, buildings, trees, grass, water, and railways, among others ROSIS Pavia University Urban area (Pavia town) Asphalt, meadows, trees, buildings, and shadows, among others Indian Pines Humid Pampas zh11 Pavia University 1 1 1 1 ML ML ML ML SBM SBM SBM SBM 0.8 0.8 0.8 0.8 Near-lossless 0.6 0.6 0.6 0.6 Kappa 0.4 0.4 0.4 0.4 0.2 0.2 0.2 0.2 0 0 0 0 5 25 5 25 5 25 5 25 Compression ratio Compression ratio Compression ratio Compression ratio 1 1 1 1 ML ML ML ML SBM SBM SBM SBM 0.8 0.8 0.8 0.8 0.6 0.6 0.6 0.6 Kappa Lossy 0.4 0.4 0.4 0.4 0.2 0.2 0.2 0.2 0 0 0 0 10 100 10 100 10 100 10 100 compression ratio Compression ratio Compression ratio Compression ratio Fig. 1: Segmentation performance in κ b (higher is better) for near-lossless (upper row) and lossy compression (lower row). In the plots, the vertical axis represents the segmentation performance (b κ) and the horizontal axis represents the compression ratio. Results using reconstructed data are plotted in solid lines and results using uncompressed data (original image) are plotted in dashed lines. Each mark represents the segmentation performance achieved when the scene has been previously compressed at a particular compression ratio. Results for the ML algorithm are plotted in green lines and results for the SBM algorithm are plotted in blue lines. Each column shows results for a particular scene. receiver side the scene is decoded, then image segmentation Table III. One can see that, for near-lossless, compression is carried out on the reconstructed data. ratios of about 10:1 allow to achieve at least the same segmentation performance as with the original data. Only the Indian Pines scene requires lower compression ratios of, A. Segmentation Assessment approximately, 4:1 to keep the segmentation results unchanged The segmentation performance was assessed with respect to when the SBM algorithm is employed. SBM uses the spectral handmade ground truth segmentations. Following the strategy information to obtain the most probable state of each pixel. used in [22], we used Cohen’s kappa coefficients (b κ) [32] to When the compression ratio increases, too much distortion compare the performance of the different schemes adopted. may be introduced in the spectral dimension, which is reflected The κb coefficient provides a measure of agreement between in the performance of the SBM algorithm. the machine segmentation and the human ground truth. Com- In a lossy scenario, compression ratios of about 25:1 allow plete agreement corresponds to κ b = 1, and lack of agreement to retain enough features to achieve the same segmentation corresponds to κb = 0. performance as with uncompressed data. As for the case of Here, segmentation results are reported for four scenes (one near-lossless compression, the Indian Pines scene begins to from each dataset). Results for the remaining images from decrease the segmentation performance at lower compression each dataset are similar. Table II provides characteristics for ratios of, approximately, 8:1 when the SBM algorithm is the scenes reported. Detailed experimental results for the 25 employed. It is interesting to note that the ML algorithm scenes described in Table I are provided as supplementary achieves higher κb coefficients compared to the SBM method material at http://gici.uab.cat/pub/coding segmentation/. when the spatial resolution of the scene is high and vice Figure 1 illustrates the segmentation performance for the versa. While ML does not take into account relations between near-lossless and lossy compression settings proposed. Plots neighboring pixels, SBM does. If there are no relations, the show the κb coefficients for different compression ratios when performance might decrease, which is likely to occur when the ML and SBM algorithms are used in the segmentation the spatial resolution is high. stage. Results using original data (uncompressed) are also The reported results reveal that the segmentation perfor- provided for comparison purposes. The relationship between mance over previously near-lossless and lossy compressed PAEs/target bit-rates and compression ratios and the SNR scenes is not markedly reduced at low and moderate compres- Energy achieved by each compression setting are reported in sion ratios. Nonetheless, lossy compression allows to keep the IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. XX, NO. Y, MONTH Z 2019 4 TABLE III: For Near-lossless (Peak Absolute Error, PAE) and Lossy(target bit-rate, BR) compression, compression ratio (CR), Bpppc and SNR Energy (in dB) are reported. The dynamic range of the original image is provided in parenthesis. The corresponding experiments for the segmentation results reported in Fig.2 are here marked in italic font. Indian Pines (14 bpppc) Humid Pampas (15 bpppc) zh11 (12 bpppc) Pavia University (13 bpppc) PAE CR bpppc SNR CR bpppc SNR CR bpppc SNR CR bpppc SNR 1 2.4:1 5.8 71.6 2.0:1 7.5 85.1 3.1:1 3.9 50.5 1.7:1 7.6 66.1 3 2.9:1 4.8 63.8 2.3:1 6.5 77.3 4.3:1 2.8 42.8 2.0:1 6.5 58.4 7 3.7:1 3.8 57.1 2.8:1 5.4 70.6 6.2:1 1.9 36.3 2.4:1 5.4 51.7 15 4.8:1 2.9 50.9 3.4:1 4.4 64.3 8.6:1 1.4 30.2 3.0:1 4.3 45.3 31 6.6:1 2.1 45.0 4.2:1 3.6 58.1 12.1:1 1.0 24.3 3.9:1 3.3 39.2 63 9.4:1 1.5 39.1 5.4:1 2.8 52.1 18.1:1 0.7 18.5 5.2:1 2.5 33.1 127 14.0:1 1.0 33.3 7.3:1 2.1 46.2 32.0:1 0.4 12.8 7.2:1 1.8 27.3 BR 4 3.8:1 3.7 56.9 4.0:1 3.8 59.8 3.0:1 4.0 53.2 3.3:1 3.9 41.5 2 7.6:1 1.8 44.5 8.1:1 1.9 46.5 6.0:1 2.0 41.5 6.6:1 2.0 30.5 1 15.0:1 0.9 36.7 16.0:1 0.9 39.8 12.1:1 1.0 33.5 13.2:1 1.0 24.4 0.5 30.0:1 0.5 31.4 32.5:1 0.5 35.2 24.2:1 0.5 27.7 26.2:1 0.5 20.3 0.25 60.9:1 0.2 28.0 65.3:1 0.2 32.7 48.4:1 0.3 23.3 52.0:1 0.25 17.3 0.1 150.5:1 0.1 25.1 161.5:1 0.1 29.5 120.0:1 0.1 18.7 130.4:1 0.1 14.3 segmentation results unchanged at higher compression ratios instruments, image processing chains, and image analysis compared to near-lossless compression. This is due to the applications. different nature of the distortion introduced by near-lossless and lossy compression. The impact of compression is the same R EFERENCES for both the ML and SBM algorithms. Only in the case of [1] V. Dey, Y. Zhang, and M. Zhong, “A review on image segmentation tech- the Indian Pines scene, the SBM algorithm is affected more niques with remote sensing perspective,” in ISPRS TC VII Symposium 100 Years ISPRS, 2010, pp. 31–42. significantly by compression than the ML algorithm. [2] R. G. Cinbis, J. Verbeek, and C. Schmid, “Segmentation Driven Object Figure 2 reports the handmade ground truth employed to Detection with Fisher Vectors,” in Proc. of the IEEE International assess the segmentation performance and some segmentation Conference on Computer Vision, 2013, pp. 2968–2975. [3] L. Zhang, K. Jia, X. Li, Q. Yuan, and X. Zhao, “Multi-scale segmentation results for the ML and SBM algorithms. This figure shows approach for object-based land-cover classification using high-resolution the segmented images when the original scene and the re- imagery,” Remote Sensing Letters, vol. 5, no. 1, pp. 73–82, 2014. constructed data from near-lossless and lossy compression are [4] L. Ma, M. Li, T. Blaschke, X. Ma, D. Tiede, L. Cheng, Z. Chen, and D. Chen, “Object-Based Change Detection in Urban Areas: The Effects used in the segmentation stage. For reconstructed data, the of Segmentation Strategy, Scale, and Feature Space on Unsupervised higher compression ratio that allows to achieve competitive Methods,” Remote Sensing, vol. 8, no. 9, p. 761, 2016. segmentation performance for both the ML and SBM algo- [5] M. Mueller, K. Segl, and H. Kaufmann, “Edge- and region-based segmentation technique for the extraction of large, man-made objects rithms is displayed. in high-resolution satellite imagery,” Pattern Recognition, vol. 37, no. 8, pp. 1619–1628, 2004. V. C ONCLUSIONS [6] H. G. Kaganami and Z. Beiji, “Region-based segmentation versus edge detection,” in IEEE International Conference on In Intelligent In this letter, the performance of image segmentation on Information Hiding and Multimedia Signal Processing (IIH-MSP’09), scenes that had gone through a compression and decom- September 2009, pp. 1217–1221. [7] S. Deng, Y. Zhang, and S. Tian, “Building surface texture segmentation pression stage was investigated. In the compression stage, in urban remote sensing image using improved ORTSEG algorithm,” in the JPEG-LS and the JPEG 2000 standards were used to IEEE Geoscience and Remote Sensing Symposium (IGARSS), July 2015, perform near-lossless and lossy compression, respectively. pp. 4348–4351. [8] I. Blanes and J. Serra-Sagristà, “Quality evaluation of progressive lossy- Image segmentation was carried out using the ML and SBM to-lossless remote-sensing image coding,” in 16th IEEE International algorithms. Conference in Image Processing (ICIP), 2009, pp. 3709–3712. Experimental results revealed that low and moderate com- [9] A. Zabala and X. Pons, “Effects of lossy compression on remote sensing image classification of forest areas,” International Journal of Applied pression ratios do not reduce the performance of the segmenta- Earth Observation and Geoinformation, vol. 13, no. 1, pp. 43–51, 2011. tion algorithms when reconstructed data is used. Near-lossless [10] F. Garcı́a-Vı́lchez, J. Muñoz-Marı́, M. Zortea, I. Blanes, V. González- compression allowed to produce the same segmentation per- Ruiz, G. Camps-Valls, A. Plaza, and J. Serra-Sagristà, “On the impact of lossy compression on hyperspectral image classification and unmixing,” formance as with the original data at compression ratios of, ap- IEEE Geoscience and Remote Sensing Letters, vol. 8, no. 2, pp. 253– proximately, 10:1. For lossy compression, unchanged segmen- 257, 2011. tations were achieved at compression ratios of, approximately, [11] Z. Chen, Y. Hu, and Y. Zhang, “Effects of compression on remote sens- ing image classification based on fractal analysis,” IEEE Transactions on 25:1. The impact of compression was similar for both the Geoscience and Remote Sensing, vol. 57, no. 7, pp. 4577–4590, 2019. ML and SBM algorithms and the different scenarios analyzed. [12] G. Carvajal, B. Penna, and E. Magli, “Unified lossy and near-lossless hy- Only in the case of the Indian Pines scene, segmentation results perspectral image compression based on JPEG 2000,” IEEE Geoscience and Remote Sensing Letters, vol. 5, no. 4, pp. 593–597, 2008. remained unchanged at compression ratios of 4:1 and 8:1 for [13] Q. Du and J. E. Fowler, “Hyperspectral image compression using JPEG near-lossless and lossy compression, respectively, when the 2000 and principal component analysis,” IEEE Geoscience and Remote SBM algorithm was used. These observations are consistent Sensing Letters, vol. 4, no. 2, pp. 201–205, 2007. [14] Q. Du and C. I. Chang, “Linear mixture analysis-based compression for with similar results reported for other scientific areas and hyperspectral image analysis,” IEEE Transactions on Geoscience and may benefit the development of upcoming remote sensing Remote Sensing, vol. 42, no. 4, pp. 875–891, 2004. IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. XX, NO. Y, MONTH Z 2019 5 Ground Original scene Recovered scene from near-lossless Recovered scene from lossy truth Scene ML SBM Scene ML SBM Scene ML SBM Indian Pines κ b = 0.25 κ b = 0.41 PAE = 7 κ b = 0.25 κ b = 0.41 BR = 2 κ b = 0.25 κ b = 0.38 CR = 4:1 CR = 8:1 Humid Pampas κ b = 0.69 κ b = 0.74 PAE = 127 κ b = 0.69 κ b = 0.74 BR = 0.1 κ b = 0.71 κ b = 0.70 CR = 7:1 CR = 161:1 zh11 κ b = 0.57 κ b = 0.44 PAE = 63 κ b = 0.54 κ b = 0.44 BR = 0.25 κ b = 0.57 κ b = 0.42 CR = 18:1 CR = 48:1 Pavia University κ b = 0.47 κ b = 0.38 PAE = 127 κ b = 0.47 κ b = 0.40 BR = 0.5 κ b = 0.57 κ b = 0.36 CR = 7:1 CR = 26:1 Fig. 2: Ground truth and segmentation results of the original and the reconstructed data. Each row shows a particular scene. Results are plotted for both the ML and SBM algorithms. The PAE/target bit-rate (BR) used in the compression stage and the compression ratio (CR) achieved are reported for each scene. The κb coefficient is reported for each segmentation. [15] M. Zortea and A. Plaza, “Spatial preprocessing for endmember extrac- approach to segmentation of multispectral remote sensing images based tion,” IEEE Transactions on Geoscience and Remote Sensing, vol. 47, on MRF,” IEEE Geoscience and Remote Sensing Letters, vol. 12, no. 8, no. 8, pp. 2679–2693, 2009. pp. 1720–1724, 2015. [16] J. Garcı́a-Sobrino, I. Blanes, V. Laparra, G. Camps-Valls, and J. Serra- [23] A. Sarti, C. Corsi, E. Mazzini, and C. Lamberti, “Maximum likelihood Sagristà, “Impact of near-lossless compression of IASI L1C data on segmentation of ultrasound images with Rayleigh distribution,” IEEE statistical retrieval of atmospheric profiles,” in Proc. of On-Board Transactions on Ultrasonics, Ferroelectrics, and Frequency Control, Payload Data Compression Workshop (OBPDC), Venice, Italy, October vol. 52, no. 6, pp. 947–960, 2005. 2014. [24] I. B. Ayed, A. Mitiche, and Z. Belhadj, “Polarimetric image segmen- [17] J. Garcı́a-Sobrino, J. Serra-Sagristà, V. Laparra, X. Calbet, and tation via maximum-likelihood approximation and efficient multiphase G. Camps-Valls, “Statistical atmospheric parameter retrieval largely level-sets,” IEEE Transactions on Pattern Analysis & Machine Intelli- benefits from spatial-spectral image compression,” IEEE Transactions gence, vol. 9, pp. 1493–1500, 2006. on Geoscience and Remote Sensing, vol. 55, no. 4, pp. 2213–2224, [25] J. M. Beaulieu and R. Touzi, “Segmentation of textured polarimetric sar 2017. scenes by likelihood approximation,” IEEE Transactions on Geoscience [18] J. Garcı́a-Sobrino, V. Laparra, J. Serra-Sagristà, X. Calbet, and and Remote Sensing, vol. 42, no. 10, pp. 2063–2072, 2004. G. Camps-Valls, “Improved Statistically Based Retrievals via Spatial- [26] GIC, “Hyperspectral remote sensing scenes,” Available: http://www.ehu. Spectral Data Compression for IASI Data,” IEEE Transactions on eus/ccwintco/index.php?title=Hyperspectral Remote Sensing Scenes. Geoscience and Remote Sensing, pp. 1–18, 2019. [27] M. Volpi and V. Ferrari, “Zurich summer dataset,” Available: https:// [19] M. J. Weinberger, G. Seroussi, , and G. Sapiro, “The LOCO-I lossless sites.google.com/site/michelevolpiresearch/data/zurich-dataset. image compression algorithm: Principles and standardization into JPEG- [28] T. Kanungo, D. M. Mount, N. S. Netanyahu, C. Piatko, R. Silverman, LS,” IEEE Transactions on Image Processing, vol. 9, no. 8, pp. 1309– and A. Y. Wu, “An efficient k-means clustering algorithm: Analysis and 1324, August 2000. implementation,” IEEE Transactions on Pattern Analysis & Machine [20] D. Taubman and M. Marcellin, JPEG 2000 image compression fun- Intelligence, vol. 7, pp. 881–892, 2002. damentals, standards and practice: image compression fundamentals, [29] D. A. Clunie, “JPEG-LS,” http://www.dclunie.com/jpegls.html. standards and practice. Springer Science & Business Media, 2012. [30] D. S. Taubman, “Kakadu software,” http://www.kakadusoftware.com/. [21] A. Dempster, N. M. Laird, and D. B. Rubin, “Maximum likelihood from [31] J. Baumgartner, “Markov-Models toolbox for segmentation of images incomplete data via the EM algorithm,” Journal of the Royal Statistical and local data,” 2015, LIMAC, FCEFyN. Society. Series B (methodological), pp. 1–38, 1977. [32] J. Cohen, “A coefficient of agreement for nominal scales,” Educational [22] J. Baumgartner, J. Gimenez, M. Scavuzzo, and J. Pucheta, “A new and psychological measurement, vol. 20, no. 1, pp. 37–46, 1960.