CC BY
18Journal of Siberian Federal University. Engineering & Technologies, 2017, 10(6), 805-811
УДК 528.854.2; 004.93'11
Comparison of Spectral-Spatial Classification Methods for Hyperspectral Images of High Spatial Resolution
Pavel V. Melnikov, Igor A. Pestunov and S.A. Rylov*
Institute of Computational Technologies SB RAS 6 Akademika Lavrentieva, Novosibirsk, 630090, Russia
Received 30.04.2017, received in revised form 04.06.2017, accepted 18.08.2017
This paper reviews three methods of spectral-spatial classification for hyperspectral images of high spatial resolution: 1) pixelwise classification with post-filtering of resulting class map; 2) spectral-spatial classification based on geometric moments; 3) spectral-spatial classification based on segmentation. The paper provides the results of experimental comparison of these methods. The experiments are based on classification of images obtained by airborne hyperspectral sensor.
Keywords: hyperspectral images, local context, spectral-spatial classification.
Citation: Melnikov P.V., Pestunov I.A., Rylov S.A. Comparison of spectral-spatial classification methods for hyperspectral images of high spatial resolution, J. Sib. Fed. Univ. Eng. technol., 2017, 10(6), 805-811. DOI: 10.17516/1999-494X-2017-10-6-805-811.
© Siberian Federal University. All rights reserved
* Corresponding author E-mail address: [email protected], [email protected]
Сравнение методов классификации гиперспектральных изображений высокого пространственного разрешения по спектральным и пространственным признакам
П.В. Мельников, И.А. Пестунов, С.А. Рылов
Институт вычислительных технологий СО РАН Россия, 630090, Новосибирск, пр. Академика Лаврентьева, 6
В работе рассматриваются три метода классификации гиперспектральных изображений высокого пространственного разрешения: 1) попиксельная классификация с последующей фильтрацией получаемой картосхемы, 2) спектрально-текстурная классификация на основе геометрических моментов и 3) спектрально-текстурная классификация на основе предварительной сегментации. Приводятся результаты экспериментального сравнения указанных методов на данных, полученных с помощью авиационного гиперспектрометра.
Ключевые слова: гиперспектральное изображение, локальный контекст, спектрально-текстурная классификация.
Introduction
In the field of aerospace remote sensing there is active development of hyperspectral systems, providing images in visible and infrared regions of the spectrum [1]. Now there is a potential to use highly informative hyperspectral images (HSI) for a wide range of scientific and practical problems. However, a significant limitation to such usage is the lack of suitable tools for automated analysis of hyperspectral images.
Among the main features of the HSI are high spectral resolution (of the order of a few nanometers) and a large number (up to several hundreds) of spectral channels, which raises the problem of the so-called "curse of dimensionality", due to which many of the traditional classifiers become unusable. In addition, pixelwise classification of high spatial resolution HSI often results in fragmented noisy maps, which are difficult to interpret and to use [2].
This report presents the results of experimental comparison of three HSI classification methods that take into account both spectral and spatial characteristics: 1) pixelwise classification followed by spatial filtering of a resulting classified image, 2) spectral-textural classification based on geometric moments and 3) spectral-textural classification based on preliminary segmentation. For the experiments, we used two images taken in 2011 by aerial hyperspectrometer developed by NPO "Lepton" (Zelenograd-based company) [3]. Before the classification a selection of uncorrelated systems of spectral features as created by applying Principal Component Analysis (PCA) method and its modification, Minimum Noise Fraction (MNF) method. These methods are well established in the area of HSI processing and allow to reduce the number of spectral features by an order of magnitude without compromising the quality of the classification [4].
Description of the HSI classification methods
The first classification method used in the experiments is described indetail in [4]. The method consists of pixelwise classification of the HSI and then spatial filtering of the resulting classified image with Majority Filter (MF). Each pixel is assigned a class to which the majority of pixels in a predetermined surrounding area belong. For pixelwise classification, method ofMnximumlikelitood (ML) and Support Vector Machine (SVM) were used.
The second and third classification methods are based on ode teoe of cnformation about image texture. There is no universally accepted definition of texture but in theaioaof maOoiifiid hyporspoctal imagery, texture of an object can be interpreted as the characteriaticoC tCidiotaibuiionifspcoteal brightness vectors of the image region occupied by an object, which is caused by the regular arrangement of non-uniform elements of the object.
The second method of classification consists of extracting aexioralfeaiurestiy uimggeomutrio moments and subsequent classification of obtained feature vectors.Geometric momenisorewido lyused to determine the textural characteristics of the objects on monochromoimagec[5]. Geome trical moment
M N
mpq of the orderp, q of the digital image I(i, j) (with size M x N) is defined os mpq = ^^ ip jaI(i, j).
i=i j=i
In the area of texture analysis geometric moments are calculated for a window (of size l x l) surrounding the pixel in question. For hyperspectral images calculation of moments for multiple sets of p, q significantly increases the number of features, thereby only intensifying the "curse of dimensionality". Therefore, in this study only one moment m00 was used, which is the sum of the spectral brightnesses of pixels in the window of size l o I (in this case for each pixel the feature value is determined by the formula Avg(i, j) = m0 0 (i, j)/l2). As in the first method, the classification of resulting feature vectors was performed by ML and SVM methods.
The third method of classification is described in [6] and is based on the pre-segmentation of HSI based on spectral features. The basic idea of this method is as follows. Using only spectral features for texture classification will lead to a fragmented noisy classified image. However, in a given area of the image covered by one object the percentage of pixels of different clusters will approximately be the same while for different objects this characteristic will differ. This pattern holds for most of the textures corresponding to objects of natural origin (e.g. forest, swamp, tundra). This approach has been successfully used for textural segmentation of multispectral images based on grid clustering algorithms [7]. The advantage of this method is that it does not require a large amount of training samples; it is sufficient to provide only a few samples for each class.
Experimentalresults
In the experiments two images with sizes of 600x420 and 1000 x350 pixels were used. Each image contained 87 spectral channels in the range 404-1016 nm. RGB composites of the images are shown in Fig. 1a and 2a. The spatial resolution was around 1 m. The images show areas of Savvatevskoe forestry in Tver Oblast region.
Ground-truth reference maps obtained from the surveys of forest taxation were available for areas that are presented on these images. The ground-truth maps contained classes corresponding to species and age composition of forest stands. However, reference maps were several decades older than
trth'
1 100%, SO years
2 ^|Pine 100%, 75 years
3 Pine 100%, 70 years
4 j_Pine 100%, 50 years
5 ^^|Birch 70%, pine 30%, 60 years
6 ^^jBirch60%, pine 40%, 60 years
7 ^Bcirch 70%, pine 30%, 50 years S | Birch 60%, pine 40%, 70 years
9 ^|Pine 100% (sparse), 50 years
10 Pine 80%, birch20%, 50 years 111 Soil
Fig. 1. RGB-composite (channels81, 19,10)(a) and referencemap(b)of image 1
Fig. 2. RGB composite (channels 81,19, 10)(a) andreferencemap(b) of image 2
the images, so reinterpretation by visual analysis of the images was performed by experts (resulting reference maps are presented in Fig. 1b and 2b). Doubtful pixels on the boundaries of the classes were assigned to the background (shown as black) and were not taken into account in the assessment of classification accuracy.
Control samples were used to assess the quality of the classification. 1000 randomly selected points of each class were used for training of classifiers. The classification results were averaged over five independent runs (with different training sets). We used the majority filter (MF) with window size of 5*5 pixels, and for calculatingAvg(i, j) texture features we used a window of 21*21 pixels.
The results of classification using different sets of features and classification methods are shown in Fig. 3 and 4. For comparison, the figures also include the accuracy of pixelwise classification based only on spectral features. The accuracy of the spectral-textural segmentation-based classification of image 2 is shown in Figure 5. First 4 principal components were used as spectral features in this experiment.
Fig. 3. Classification accuracyofimage l,basedondiffarent feature sets and clas sifioationmatheds,depending on number of features
Fig. 4. Classification accuracy of image 2, based on different feature sets and classification methods, depending on number of features
100 98 с S 96 3 * 94 ё % 92 S g 90 п « 88 86 84
/ ~~ ■ « — ■
5
15 25 35 45 55 С Fragment size A, pix
Fig. 5. Accuracyofspectral-textural classificationbasedonpre-segmentationfor image2
Conclusion
Analysis of the results shows that spatial information makes it possible to achieve a significant improvemeat inthe accuracy of classification (by 5-50%) in comparison to pixelwise spectral classification. For the test images used in this research the best results were achieved by classification based on geometric moments, the accuracy approached 100%.
Thisworkwas supportedby RFBR (grant No. 14-00-00249-a).
References
[(] Дворкин Б.А., Дудкин С.А. Новейшие и перспективные спутники дистанционного зондированияЗемли. Геоматико, 2010, 2,16-21. [DvorkinB.A,DudkinS.A Neweots-id->ersoective oemote se(osingn,&Ri0i(es.Gromatics1 2010,2,16-U- (1nRussion)(
[2] Ftuvel M., Benediktsson J.A., Chanussot J., Sveinsson J.R. Spectral and spatial classification of hypeBspectral data using SVMsand morphological profi.es. IEEE CaRdc. Goosei Remote Sens., 2008, 06(H1), 3804a^36^lit.
[P] KeoaaoravV.V.,K3ndsaninT.V.,E)mitrievE.V.]KamentsevV.P.Bayesian2lassifierKpplrcati5ns of airborne hyperspectral imagery processing for forested areas. Advances in Space Research, 2015, 55(11), 2657b2667.
[4] Грин ЖВТ-2016 Борзов С.М., Мельников П.В., Пестунов И.А., Потатуркин О.И., Федотов А.М. Комплексная обработка гиперспектральных изображений на основе спектральной и пространственной информации. Вычислительные технологии. 2016, 21(1), 25-39. [Borzov S.M., Milnitov R.V., Pestunov I.A., Potaturkin O.I., Fedotov A.M. Complex spectral-spatial processing of hyperspictraliioegas.Computotional 7fecÂwo/oic'Hr,2o16,2a(8),2a3,39(InRussmE)]
[5] KamarB., OnkarD.O. eper6ral-8pot(2l clantMcotion ор Ihyperopectral imaaery1t>atenon mtmsBntowKiiKBiAs.IEEEOoumvlofKrSecleU topi-srs apptteV earlUoboervotions ая-stmoto sensmg, 22oi,8in), R457-2o6i.
[6] VCыоое C.A., МелюшкавП.В., П.стубовИ.А.Впеупобльнагтепобурнаяклаеоифапзпия 58пepcпeктpdoo4ыrизs>Kpвуr6ий nbioeKoronpacopoHCoaohHCdo разрешения. Г^нте]иксгю Г,о-Cu6uFb- 2016i4Rr), 08-1I. [Rylov 0>.A., MelnikovP.V. Ueobunov 2A., bSR^cS^^l^-^iiatmiclRS-iKnaSnii^n ot hyper-pibtsoi images ofhigO rprtial resolution. IntetRx.o GEO-Sibtr-e016. 4UU6, 4.1), 78-84 (In Russian)]
[7] Пестунов И.А., Рылов С.А. Алгоритмы спектрально-текстурной сегментации спутниковых изображений высокого пространственного разрешения. Вестник КемГУ, 2012,
4/2 (52), 104-110. [Pestunov I.A., Rylov S.A. Algorithms of spectral-spatial segmentation of satellite images of high spatial resolution. Bulletin of Kemerovo State University, 2012, 4/2(52), 104-110 (In Russian)]
