J Indian Soc Remote Sens (March 2013) 41(1):5770 DOI 10.

1007/s12524-012-0211-x

RESEARCH ARTICLE

Understanding the Unique Spectral Signature of Winter Rape

Rama Rao Nidamanuri & Bernd Zbell

Received: 24 November 2011 / Accepted: 8 March 2012 / Published online: 4 April 2012 # Indian Society of Remote Sensing 2012

Abstract Driven by significant technological developments in the hyperspectral imaging, material mapping using reference spectra has received renewed interest of the remote sensing community. The applicability of reference spectral signatures in image classification depends mainly on the material type and its spectral signature behaviour. Identification and spectral characterization of materials which exhibit unique spectral behaviour is the first step in this approach. Consequently there have been active researches for the identification of surface materials which exhibit unique spectral signatures. The uniqueness of reflectance signature of winter rape relative to its cooccurring crop species was reported in this study. Reflectance spectral libraries constructed from field spectral reflectance measurements collected over five agricultural crops (alfalfa, winter barley, winter rape,
R. R. Nidamanuri (*) Department of Earth and Space Sciences, Indian Institute of Space Science and Technology, Valiyamala, Thiruvananthapuram 695 547 Kerala, India e-mail: ramarao.iit@gmail.com R. R. Nidamanuri e-mail: rao@iist.ac.in B. Zbell Institute for Landscape Systems Analysis, Leibniz Centre for Agricultural Landscape Research (ZALF), 15374 Mncheberg, Germany

winter rye, and winter wheat) during four subsequent growing seasons were classified by the linear discriminant analysis (LDA). Further, the reference field spectral database was used for the spectral feature fitting and classification of a historical HyMAP airborne hyperspectral imagery acquired at a separate site, by spectral library search. Results indicate the existence of a meaningful spectral matching between image and field spectra for winter rape and demonstrate the potential for transferring spectral library for hyperspectral image classification. The observed consistency in the discrimination of winter rape demonstrates experimentally the fundamental principle of remote sensing which suggests the theoretical existence of unique spectral signatures for materials which can be incorporated as reference spectral signatures for hyperspectral image classification. Keywords Hyperspectral remote sensing . Spectral signatures . Winter rape . Spectral angle mapper . Spectral library search

Introduction Hyperspectral imaging has enhanced the capabilities of optical remote sensing many folds for the identification and mapping of various earth surface materials. Sophisticated ground and airborne hyperspectral sensors are available with capabilities of acquiring reflectance measurements in hundreds of bands throughout

58

J Indian Soc Remote Sens (March 2013) 41(1):5770

the 3202,500 nm wavelength range typical to optical remote sensing. For example, latest version of the ASD Fieldspec Spectroradiometer (ASDI Inc. USA) captures reflected electromagnetic radiation in more than 1,500 bands with a sampling interval of up to 2 nm. The reference reflectance measurements collected using field spectroradiometers are successfully used in the pre-processing and calibration of aerial and space borne hyperspectral imaging sensors. Building upon this scientific development, researchers are trying to evaluate the theoretical possibility of the existence of unique spectral signatures for materials. The premise behind this interest is that unique spectral signatures can be complied as a reference spectral library that can be used for material identification in hyperspectral image by searching through the spectral library. The process of material identification by spectral library search does not require analyst with spectral expertise, thus helping non-hyperspectral remote sensing specialists handle identification of various surface materials. Further, the spectral signatures, if they are unique, can be used for automatic identification and mapping of those materials across space and time (Campbell 2006). Previous studies report the complex scenario of the existence of unique spectral signatures for minerals and, in general, non uniqueness of vegetation spectral signatures. Although the method of material identification by spectral signatures were first developed for mineral mapping (Clark et al. 1990a, b, Van der meer 2005; Clark et al. 2007; Kokaly et al. 2008; Kruse and Perry 2009; Baldridge et al. 2009), they have also been evaluated to a modest level for soil (Shepherd and Walsh 2002; Bojinski et al. 2003; Brown et al. 2006; Brown 2007), and urban material classfication (Gomez 2002; Herold et al. 2004) with various degress of accuracy. However, unlike minerals which exhibit characteristic absorption features, reflectance signatures of vegetation species are influenced by soil and environmental factors. Hence the possibility of having a characteristic absorption feature for vegetation is rare, if not non-existent (Price 1994; Cochrane 2000). Therefore, the more likely case is the existence of an outstanding reflectance pattern (e.g. unique shape of the reflectance spectrum) that behaves like a unique spectral signature for a particular vegetation species (Asner and Martin 2009). In hyperspectral remote sensing terminology it is often called exemplar spectra (Gomez 2002; Nidamanuri and Zbell 2010). These exemplar spectra permit integration of in situ

reflectance measurements with airborne hyperspectral measurements for material identification and mapping. Zomer et al. (2009) report successful classification of PROBA-1 hyperspectral imagery for wetland vegetation species mapping by transferring field reflectance measurements collected from five locations in the California, USA. Despite the growing interest in using spectral measurments to integrate from the field-level to at-sensor level, material identifcation by spectral analyses has not yet received much attention from remote sensing specialists. One reason for this may be that even after nearly two decades of hyperspectral remote sensing research, we are still at the stage where we cannot yet make a quantitative, or qualitative mapping of the raw spectral pattern with the corresponding material type without calibration using ground reference data apriori. However, given the number of terrestraial and planetary materials that have been successfully identified with reference spectral measurements, it is clear that hyperspectral reflectance spectra could provide the information necessary for spectral identifcation of surface materials. This provides the motivation for the present work. During our analyses of the spatiotemporal variability of reflectance characteristics of five agricultural crops as part of our ongoing research, a distinct spectral behaviour was observed with winter rape. Further experiments on the spectral signatures of winter rape evaluating the potential for using in HyMAP hyperspectral image classification revealed the existence of a unique spectral signature for winter rape. Therefore, the objective of this work was evaluating the spectral uniqueness of winter rape relative to its co-occurring crops from field spectral crop discrimination perspective.

Materials and Methods Study Area Spectral datasets used in this work were collected from two sites located in the Northeast Germany. Field reflectance measurements were collected from experimental plots of the Leibniz-Centre for Agricultural Landscape Research (ZALF), Muencheberg, Germany. Airborne hyperspectral image was acquired for the Dedelow research station of the ZALF, which is located approximately 100 km north of the site of

J Indian Soc Remote Sens (March 2013) 41(1):5770

59

field reflectance measurements. Figure 1 illustrates the location of the study area. The land use categories in Dedelow are agricultural crops, namely, alfalfa (Medicago sativa), winter barley (Hordeum vulgare), winter rape (Brassica napus), winter wheat (Triticum spp.), and winter rye (Secale cereale). The phenology of these crops during the period of image acquisition was quite variable ranging from main shoot development to reproductive development. Spectral Data Used Field Hyperspectral Measurements The crops chosen for reflectance spectral measurements were alfalfa, winter barley, winter rape, winter rye, and winter wheat. Canopy level spectral measurements were collected using an ASD FieldSpec JR Spectroradiometer during four subsequent growing seasons 2002 through 2005. The Spectroradiometer collects data at 3 nm spectral resolution in the visible-near infrared (VNIR) and at 10 nm in the shortwave-infrared (SWIR) wavelength and stores data resampled to 1 nm in the VNIR and 2 nm in the SWIR wavelength region. Spectral measurements were collected at discrete point locations at regular
Fig. 1 Location map of the study site (satellite image: True colour image of the Landsat ETM+acquired on 16 June, 2003 obtained from USGS)

intervals of 7 to10 days during the entire growth cycle of each crop. At each measurement location, 50 reflectance measurements were consistently taken with a nadir view using an 18o FOVat a height of about 1 m above the canopy and averaged to form a reflectance spectrum. In this way, overall 1,350 reflectance spectra were collected for all the selected crops. The spectral data has been calibrated with the readings taken simultaneously with a Spectralan barium sulphate (BaSO4) white reference panel. To avoid the impact of diurnal illumination changes, all the spectral measurements were collected during 11:0013:00 noontime. To facilitate a direct comparison of wavelengths and reflectance values, the field reflectance spectra were converted to the HyMAP spectral radiometric resolutions using the spectral response function of the HyMAP imaging sensor. The wavelengths (band centres) and full width half maxima (FWHM) of the HyMAP imaging sensor were given as inputs and the spectral resampling was done using a Gaussian model with an FWHM equal to the band spacing. Airborne Hyperspectral Data Airborne hyperspectral image was acquired on May 6, 1999 using the HyMAP sensor with an area extent of

Dede elow site of image acqu uisition

* Berlin

Muenc cheberg site s of spec ctral measu urements

60

J Indian Soc Remote Sens (March 2013) 41(1):5770

610 km along eastwest flight lines. The HYMAP image used in this study was captured with a spatial resolution of 5 m and 128 spectral bands with spectral resolution up to 20 nm. Spectral band configuration of the HyMAP imaging sensor is shown in Table 1. This real-time georeferenced image was further processed to account for comprehensive radiometric and atmospheric corrections using the FLAASH atmospheric correction code (ENVI User Guide; Felde et al. 2003). A False Colour Composite (FCC) of the HyMAP image acquired is shown in Fig. 2. As our main objective was spectral discrimination of agricultural crops, all non-vegetation classes such as bare soil, roads, built-up area, and wetlands were eliminated from the classification process by preparing a crop mask based on the reference Land Use Planning Map available for the study area (Wegehenkel et al. 1999). Further, the within field non-crop features such as water body and built up were masked out from the analyses based on a vegetation mask generated by NDVI thresholding. Field Spectral Crop Classification Using Linear Discriminant Analysis (LDA) In the spectral crop discrimination, we are operating under the assumption that there is a direct relationship between spectral reflectance pattern and crop type, and thus crops have different spectral reflectance characteristics. Our hypothesis is that this spectral reflectance pattern is a signature that can be interpreted and used as a succinct, but quantitative, description of the crop type. Even without conducting detailed spectral chemical analyses, it should therefore be possible to determine the degree to which crops are spectrally similar or dissimilar on the basis of their spectral signatures. In this regard, we must

Fig. 2 FCC of the HyMAP image (R: 30, G: 20, B: 10) of the study area

Table 1 Spectral band configuration of the HyMAP sensor Module Spectral range (m) 0.450.89 0.891.35 1.401.80 1.952.48 Bandwidth across module (nm) 1516 1516 1516 1820 Average spectral sampling interval (nm) 15 15 13 17

VIS NIR SWIR1 SWIR2

take a holistic view, and focus on the entire spectrum, rather than individual wavelengths. To this end, we used linear discriminant analysis (LDA) in combination with principal component analysis (PCA) as method to investigate and evaluate the spectral discrimination of a set of samples. The LDA is a popular statistical tool that permits quantiative discrimination of a range of variables. The objective of LDA is to establish relationships between a dependent grouping variable and the original independent discriminating variables. It involves deriving a variate (or canonical function), the linear combination of the independent variables that will discriminate best between a priori defined groups (Hair et al. 1995). Discrimination is achieved by setting the variates weights for each variable to maximize the between group variance relative to the within-group variance. Various researchers have used discriminant analysis for the classification of various agricultural and land use categoeiries using in situ reflectance measurements (Kemsley et al. 1995; Atkinson et al. 1997; Lister et al. 2000). We used the LDA on the first 10 eigen vectors derived from the PCA rather than individual wavelengths, in order to reduce the likelihood of overfitting the data (Bolster et al. 1996; Hallett et al. 1997). Crop discrimination using LDA was performed with various combinations of reflectance databases collected during the four growing seasons categorized as calibration, and test data. The various combinations of spectral databases used in the LDA are shown in Table 2. The classification results were validated by cross-validation and manual inspection.

J Indian Soc Remote Sens (March 2013) 41(1):5770 Table 2 List of various combinations of spectral databases used for crop discrimination by Linear Discriminant Analysis (SD1 spectral data collected during 2002, SD2spectral data collected during 2003, SD3spectral data collected during 2004, SD4 spectral data collected during 2005) S.No 1 2 3 4 Calibration data SD1 SD2 SD3 SD4 Test data SD1(partly), SD2, SD3, SD4 SD1, SD2 (partly), SD3, SD4 SD1, SD2, SD3(partly), SD4 SD1, SD2, SD3, SD4 (partly)

61

1990b). The SFF method compares library and image spectra by exploiting the spectral absorption features using a least-squares technique. The library spectra are matched with the image spectra after the continuum is removed from both datasets. The output from the SFF is a set of gray level image pairs known as Scale and RMS for each crop type. The brighter pixels in the Scale image and darker pixels in the RMS image indicate a better match to the reference material. Scale values exceeding 1 indicate the possibility of incorrect reference endmembers or wavelength range used. Spectral Library Construction

Spectral Crop Identification Experiment with HyMAP Hyperspectral Image If the spectral signatures of crops are characteristic indeed, it should be then possible to identify and map crops in hyperspectral image by quantitative comparison of a limited number of field spectra with the corresponding image spectra. In other words, the field spectra can be transferred across for classifying hyperspectral image. This can be accomplished by formulating the classification of image spectra as a spectral library search problem. In this method, field or laboratory spectra of known materials are characterized and organized as a reference spectral database known as spectral library. This spectral database is searched through by an appropriate library search algorithm, also called as comparison metric. While the library search of materials which have relatively stable reflectance spectra (e.g. minerals) is straight forward and a simple Euclidean distance metric can be employed, the dynamic nature of crop spectral features require the library search algorithm to be insensitive to the brightness differences caused by illumination changes. Hence, the Spectral Angle Mapper (SAM) which performs hyperspectral image classification in spectral library search mode and is insensitive to illumination changes was selected as the method of evaluating reference library spectra for HyMAP image classification. The SAM labels pixels based on the lowest spectral angle difference between image and library spectra. Pixels further away than the user specified maximum angle threshold are left unclassified. In addition to this, quantitative spectral matching of the image and library spectra with emphasis on the presence of spectral absorption features was evaluated by the Spectral Feature Fitting (SFF) method (Clark et al.

In order to assess the spectral matching of field and image spectra and the impact of number of samples for each crop type, the entire field spectral database was compiled into two spectral libraries after all the spectral measurements were successively averaged to generate a limited number of representative spectra for each crop. The first library, named as SL1 contained 16 reference spectra for each crop, four spectra per growing season. The second one, named as SL2 contained the entire spectral database of the crops without any spectral averaging. In the spectral library SL2, the number of entries for each crop varied from 230 (winter rye) to 295 (winter wheat).

Results and Analysis Crop Discrimination Using the LDA Spectral discrimination of the crops using LDA was carried out for various combinations of the spectral datasets collected during the four growing years. Summary of the accuracy assessment are presented in Table 3. When the calibration and test data are drawn from the same growing year, all the five crops exhibit excellent inter-crop discrimination resulting in classification accuracies exceed 90 %. The significant decrease in the crop classification accuracy when using the calibration and test spectra drawn from different growing years indicates the susceptibility of crops spectral discrimination to seasonal and environmental factors. However, this decrease in the classification accuracy is quite variable and is crop specific. Winter barley, winter wheat, and winter rye exhibit significant reduction in the crop discrimination as evident from

62 Table 3 Classification accuracy (%) obtained for various combinations of calibration and testing spectral datasets (SD1 spectral data collected during 2002, SD2 spectral data Class Alfalfa Winter barley

J Indian Soc Remote Sens (March 2013) 41(1):5770 collected during 2003, SD3spectral data collected during 2004, SD4spectral data collected during 2005) Winter rape Winter rye Winter wheat

(a) when using spectral database SD1 as calibration data Test data SD1 SD2 SD3 SD4 Test data SD1 SD2 SD3 SD4 Test data SD1 SD2 SD3 SD4 Test data SD1 SD2 SD3 SD4 100 93 92 89 89 99 86 84 87 96 95 88 81 85 88 96 99 72 81 69 82 95 76 65 61 68 91 78 73 66 71 92 100 100 97 99 99 100 97 99 99 100 97 99 97 100 97 98 99 67 70 55 72 97 60 52 54 63 93 69 64 53 69 89 98 82 71 69 79 96 81 77 69 76 92 77 55 61 72 91

(b) when using spectral database SD2 as calibration data

(c) when using spectral database SD3 as calibration data

(d) when using spectral database SD4 as calibration data

the reduced classification accuracies. The comparatively lower classification accuracy obtained for winter barley (61 % to 82 %), winter wheat (55 % to 82 %) and winter rye (53 % to 72 %) indicates (i) the influence of seasonal and micro-management factors on the crops spectral discrimination, and (ii) the lack of robust spectral features. However, comparatively higher and stable classification accuracy of alfalfa and winter rape indicates the presence of distinct spectral behaviour which is stable across the growing seasons. Especially, the inter-growing season classification accuracy of winter rape (i.e. classification using calibration and test spectra from different growing seasons) is comparable with that of same-season classification accuracy indicating the presence of consistent and distinct spectral behaviour for winter rape. Overall, significant inter-growing season spectral discriminations of the crops were obtained using in situ hyperspectral reflectance measurements. However, transferability of the in situ canopy spectral differences to airborne image pixel level is crucial to assessing the existence of spectral finger prints for crop species.

Spectral Crop Identification in HyMAP Image by the SAM Method Spectral discrimination of the crops in the HyMAP image was carried out by varying the number of input reference spectra as per the spectral libraries SL1 and SL2. Varying the number of reference spectra iteratively for each class changed the composition and distribution of classified pixels across the growing seasons. Figure 3 shows the classified images obtained from using the spectral libraries constructed. When only two reference spectra per crop type were used for library search (i.e. searching through SL1 selecting only two spectra per crop per growing at a time), the majority of the image pixels are left unclassified (results are not shown). However, when increasing the number of reference spectra per crop type to four and repeating the image classification, most of the image pixels were classified (Fig. 3(a)).The majority of the pixels left unclassified (60 %) are winter rape. The accuracy of the classification ranged from negligible (1 %) to significant (71 %) (Table 4). As evident

J Indian Soc Remote Sens (March 2013) 41(1):5770

63

Fig. 3 Classified image showing the crops selected (HyMAP image was classified by library search through spectral library SL1 (a) including four reference spectra per crop, (b) including

eight reference spectra per crop, and (c) using all the 16 reference spectra per crop

from Fig. 3(a) and Table 4, with the exception of winter rape, majority pixels of all the crops are misclassified (as belonging to one crop or the other). For instance, most of the winter wheat pixels are labelled

as winter rye, while the winter barley pixels are systematically misclassified as winter wheat. Remarkably, classified pixels of winter rape are accurate despite nearly one half of the area of winter rape left

Table 4 Contingency matrix showing accuracy assessment results obtained from the classification of HyMAP image by spectral library search through SL1 using only four reference spectra per crop Ground truth (pixels) Class Classified image (pixels) Unclassified Alfalfa Winter barley Winter rape Winter rye Winter wheat Total Overall accuracy 0 32 % Kappa coefficient 0 0.18 Alfalfa 3 135 12 0 95 505 750 Winter barley 79 11 13 12 156 982 1253 Winter rape 595 0 0 711 9 23 1338 Winter rye 2 0 2 0 189 73 266 Winter wheat 49 0 8 0 299 339 695 Total 337 146 35 1114 748 1922 4302

64

J Indian Soc Remote Sens (March 2013) 41(1):5770

unclassified (Fig. 3(a)). Further increase in the number of reference spectra per crop to eight and repetition of the library search produced results almost similar to those obtained from using four reference spectra for each crop (see Fig. 3(b)). But, there is a decrease in the number of unclassified pixels of winter rape. Repetition of the classification by searching through SL1 using all the reference spectra produced the results (see Fig. 3(c) and Table 5) which are almost similar to those obtained from searching through SL1 using eight reference spectra per crop. However, major changes are the labelling of the previous unclassified pixels and the completeness of the classification performance by the winter rape. Post classification processing of screening out isolated single pixels and running a 33 majority filter on the classification results (obtained from searching through the SL1 using all the reference spectra) yielded a classified image which is smooth and coherent (Fig. 3(c)) indicating its potential for further studies to long-term reference spectral signature development. On the other hand, the spectral confusion as evident from the misclassifications of winter rye, winter barley and winter wheat enhanced to being mutually inclusive. This is evident from the completeness in the mislabelling of winter barley and winter wheat plots (Fig. 3(c)). The final case of repeating the library search by using all the 1,350 entries in the spectral library SL2 produced no additional gains in the classification performance but impractical post classification processing (Yang and Tien 2010).

Several hundreds of tiny and single pixel classes were accumulated in the classified image with no meaningful relation to the ground truth image. Therefore results from the SL2 were discarded. Classification Using the SFF Method Because of the complexity of the output grey level images, the SFF method was applied only on the various combinations of reference spectra in SL1. Examination of the numerous scale and RMS error images of the crops across the growing seasons and number of entries in the spectral libraries revealed the quantitative basis for the classification results obtained by the SAM method. Results from the SFF method confirm the randomness observed in the classification of winter barley, winter rye, and winter wheat by the SAM method. The Scale and RMS values of winter barley, winter rye, and winter wheat range from 1.1 to 1.6 and 0.28 to 0.59 respectively. Typical Scale and RMS image pairs of the crops obtained by the SFF are shown in Fig. 4. From Fig. 4, it is evident that the Scale values for winter barley (Fig. 4(c)), winter rye (Fig. 4(g)), and winter wheat (Fig. 4(i)) exceed the theoretical upper limit of one for a valid spectral comparison (Clark et al. 1990a) indicating the inability of the field spectra of these crops for spectral matching. This invalid spectral matching is also evident by the comparatively higher RMS error values. In contrast to that, winter rape exhibited an outstanding spectral matching with the image spectra while maintaining distinct spectral separation from the other

Table 5 Contingency matrix showing accuracy assessment results obtained from the classification of HyMAP image by spectral library search through SL1 using 16 reference spectra per crop Ground truth (pixels) Class Classified image (pixels) Unclassified Alfalfa Winter barley Winter rape Winter rye Winter wheat Total Overall accuracy 0 43 % Kappa coefficient 0 0.22 Alfalfa 0 52 1 0 42 655 750 Winter barley 3 1 32 0 89 1128 1253 Winter rape 2 0 0 1327 2 7 1338 Winter rye 0 0 1 0 140 125 266 Winter wheat 4 0 35 0 346 310 695 Total 9 53 69 1327 619 2225 4302

J Indian Soc Remote Sens (March 2013) 41(1):5770 Fig. 4 Scale and RMS images of alfalfa (a, b), winter barley (c, d), winter rape (e, f), winter rye (g, h), and winter wheat (i, j) from the SFF method

65

66

J Indian Soc Remote Sens (March 2013) 41(1):5770

crops. This is clearly evident from the higher scale values but less than 1 (Fig. 4(e), (f)) and the lowest RMS error values in the corresponding Scale and RMS images for winter rape. The maxima and minima of the Scale and RMS values obtained for winter rape ranged from 0.8 to 1 and 0.013 to 0.028. This observation supports validity of the results obtained from the SFF method for winter rape (Fig. 4(e), (f)) and the exemplar spectral matching observed by using the SAM method.

Discussion State-of-the-art hyperspectral imaging systems provide reflectance data with spectral resolutions that are required for the purposes of studying material chemistry. Material identification by reference library search is an emerging approach of hyperspectral image classification. Leveraging on the scientific understanding that many minerals exhibit unique spectral absorption features, previous studies have demonstrated the proof of the concept by automatic classification of airborne hyperspectral images for mineral identification mapping (Clark et al. 1990a). In these studies, an unknown spectrum is identified by simply matching the dominant absorption feature in its reflectance spectrum with those present in a spectral library. However, unlike minerals which exhibit characteristic absorption features, vegetation species show distinct spectral reflectance patterns driven by plant type, structure, phenology, and the presence and orientation of individual plant components (Price 1994; Cochrane 2000; Kumar 2007; Jungho and Jensen 2008). Therefore, vegetation species identification by spectral library search will more likely be a result of the existence of distinct shape of reflectance spectrum, often called exemplar spectra (Gomez 2002; Nidamanuri and Zbell 2010). These exemplar spectra permit integration of in situ reflectance measurements with airborne hyperspectral measurements for material identification and mapping. Our experimental setup of collecting multi-season field spectral measurements independent of the time and location of the hyperspectral image acquisition made it possible to assess the existence of unique spectral signatures with potential for hyperspectral image classification. While an excellent intragrowing season spectral discrimination is exhibited by all the five crops considered, winter rape has shown

consistently higher classification accuracy across the growing seasons. Alfalfa, which exhibited fairly good classification accuracy across growing seasons from in situ spectral measurements has shown marginal spectral discrimination at the pixel level. The observed consistency in the discrimination of winter rape demonstrates experimentally the fundamental principle of remote sensing which suggests the incorporation of reference spectral signatures for image classification (Gong et al. 1997; Campbell 2006). A visual examination of the reflectance profiles of the crops selected indicates the presence of spectral similarity among cereals and a unique reflectance pattern for winter rape (Fig. 5). Especially the visible reflectance pattern of winter rape is unique in its entirety, explaining the higher classification accuracy obtained from both the SFF and SAM methods. Adding to that, in situ reflectance spectrum of winter rape closely approximates the spectrum extracted from airborne HyMAP image by spectral variation. These outstanding spectral features may be responsible for the discrimination of winter rape by the SFF and SAM methods. The remarkably higher reflectance of winter rape in the red and green wavelengths indicate the reflectance of the characteristic yellow flowers of winter rape, which are the dominant canopy components for mature winter rape. However, caution must be exercised in generalizing the outstanding reflectance signature of winter rape across other independent sites and phenological stages because of the fact that winter rape flowers are the dominant plant canopy components which may not be optimal for uniquely characterizing the reflectance properties. On the other hand, the finer spectral differences existing among the field spectra of cereal crops (Fig. 5) are evident mainly by the apparent differential brightness values rather than the spectral variation, thus leading to negligible discrimination of winter barley, winter rye, and winter wheat by the SFF and SAM methods. This is because both the SFF and SAM methods use shape of the reflectance spectrum for endmember identification (Mazer et al. 1988; Kruse et al. 1993). However, the SFF method quantifies similarity based on the position and depth of spectral absorption features, whereas the SAM method considers the overall shape of reflectance spectra as measured by vector dot product (Price 1994). In a similar study Nidamanuri et al. (2007) reported an overall accuracy of about 90 % in an experiment of

J Indian Soc Remote Sens (March 2013) 41(1):5770 Fig. 5 Field and image spectral signatures of the crops selected: a alfalfa, b winter barley, c winter rape, d winter rye, e winter wheat. Winter rape and alfalfa show an outstanding reflectance pattern in the Visible and NIR wavelengths. Unique visible reflectance feature of winter rape is clearly evident. Solid line indicates image spectrum and dotted indicates field spectrum

67

supervised classification of Hyperion hyperspectral imagery by the SAM method for mapping four

agricultural crops (paddy rice, cotton, sugarcane, and chillies) using field spectra. As against our case of

68

J Indian Soc Remote Sens (March 2013) 41(1):5770

independent locations of field measurements and image acquisition, however, their study used field spectra that are collected from the same crop fields and concurrent to the image acquisition. The observed consistency in the lack of matching between field and image spectra for winter barley, winter rye, and winter wheat by any of the methods closely follows the observations of Price (1994) and Cochrane (2000). After posing the question How unique are spectral signatures? Price concludes that the spectral signatures of various vegetation species are not unique and that there is no oneto-one relationship between vegetation species and spectral signatures. By using leaf and branch level reflectance spectra for forest species previous studies report the lack of unique spectral signatures for vegetation. The plant specific outstanding spectral feature of winter rape observed in this study supports the study reported by Andrew and Ustin (2006). They report unique identification and discrimination of perennial pepperweed from its co-occurring species using field reflectance measurements and airborne hyperspectral imagery. Furthermore, studies of Zomer et al. (2009), Ciraolo et al. (2006), Kutser et al. (2006), and Purkisand and Pasterkamp (2004) report successful automatic classification of airborne hyperspectral data for various wetland and submerged aquatic vegetation species with spectral libraries developed using in situ reflectance spectra. However, most of these studies used in situ reflectance spectra that were collected concurrent with the date of image acquisition and from the same locations. Contrary to this, we used spatially and temporally independent image and spectral measurements. The wider spatio-temporal differences maintained in our experiment set the conditions necessary for assessing uniqueness of spectral signatures and ensure free of being site-specific at least on a regional basis. The observed consistency in the identification and classification of winter rape across growing seasons and independent of the location suggests the existence of a unique spectral signature, aka spectral finger print for winter rape. This is an important observation in the context of the subjectivity in the current notion that unique spectral signatures may not exist for vegetation (Cochrane 2000), despite the theoretical possibility. Interestingly, the observed randomness in the spectral matching of winter barley, winter rye, and winter wheat contribute to the notion of non-existence of

unique spectral signatures. Based on these results, it can be inferred that the existence or non-existence of spectral fingerprints cannot be generalized across vegetation but is case specific. The crops (or natural vegetation species) which exhibit this feature of spectral uniqueness may become potential candidates for automated mapping using hyperspectral imagery. We, therefore, recommend for review the of on-going efforts of building spectral libraries (e.g. Clark et al. 2007; Baldridge et al. 2009; Huenia et al. 2009) to maintain a collaborating open-end vegetation reference spectral library whose entries exhibit unique spectral characteristics. The results of this study indicate that winter rape is one such potential candidate.

Conclusions The identification and spectral characterization of earth surface materials which have unique spectral signatures has tremendous potential for increasing the efficiency of material mapping using hyperspectral image. The existence of such spectral characteristics for winter rape has been identified and evaluated. In addition to the linear discriminant analysis of the field spectral database, the spatio-temporal spectral matching and classification of HyMAP image for five agricultural crop types suggests the existence of a unique spectral matching between image and field spectra for winter rape. The matching of the field reflectance measurements of winter rape with the image spectra, consistent across seasons, indicates the existence of physiologically unique spectral signature for winter rape. This distinct spectral behaviour can be used for automatic mapping of winter rape across space and time. However, the randomness observed in the classification of winter rye, winter wheat, and winter barley suggests that reflectance spectra of some crop species may closely approximate other similar group of crop species even by high resolution spectral data.

References
Andrew, M. E., & Ustin, S. L. (2006). Spectral and physiological uniqueness of perennial pepperweed (Lepidium latifolium). Weed Science, 54, 10511062. Asner, G. P., & Martin, R. E. (2009). Airborne spectranomics. mapping canopy chemical and taxonomic diversity in

J Indian Soc Remote Sens (March 2013) 41(1):5770 tropical forests. Frontiers in Ecology and the Environment, 7, 269276. Atkinson, M. D., Jervis, A. P., & Sangha, R. S. (1997). Discrimination between Betula pendula, Betula pubescens, and their hybrids using near-infrared reflectance spectroscopy. Canadian Journal of Forest Research, 27, 1896 1900. Baldridge, A. M., Hook, S. J., Grove, C. I., & Rivera, G. (2009). The ASTER spectral library version 2.0. Remote Sensing of Environment, 113, 711715. Bojinski, S., Schlpfer, D., & Itten, K. (2003). SPECCHIO. a spectrum database for remote sensing applications. Computers & Geosciences, 29, 2738. Bolster, K. L., Martin, M. E., & Aber, J. D. (1996). Determination of carbon fraction and nitrogen concentration in tree foliage by near infrared reflectance. a comparison of statistical methods. Canadian Journal of Forest Research, 26, 590600. Brown, D. J. (2007). Using a global VNIR soil-spectral library for local soil characterization and landscape modeling in a 2nd-order Uganda watershed. Geoderma, 140, 444453. Brown, D. J., Shepherd, K. D., Walsh, M. G., Dewayne, M. M., & Reinsch, T. G. (2006). Global soil characterization with VNIR diffuse reflectance spectroscopy. Geoderma, 132, 273290. Campbell, J. B. (2006). Introduction to remote sensing. 4th Edition, New York: Guilford Publications. Ciraolo, G., Cox, E., Loggia, G. L., & Maltese, A. (2006). The classification of submerged vegetation using hyperspectral MIVIS data. Annals of Geophysics, 49, 287294. Clark, R. N., King, T. V. V., Klejwa, M., Swayze, G., & Vergo, N. (1990a). High spectral resolution reflectance spectroscopy of minerals. Journal of Geophysical Research, 95, 1265312680. Clark, R. N., Gallagher, A. J., & Swayze, G. A. (1990b). Material absorption band depth mapping of imaging spectrometer data using a complete band shape least-squares fit with library reference spectra, Proceedings of the Second Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) Workshop. JPL Publication 9054, 176186. Clark, R. N., Swayze, G. A., Wise, R., Livo, E., Hoefen, T., Kokaly, R.,et al. (2007). USGS digital spectral library splib06a. U.S. Geological Survey, Digital Data Series 231. Cochrane, M. A. (2000). Using vegetation reflectance variability for species level classification of hyperspectral data. International Journal of Remote Sensing, 21, 21752087. Felde, G. W., Anderson, G. P., Cooley, T. W., Matthew, M. W., Adler-Golden, S. M., Berk, A., et al. (2003). Analysis of Hyperion data with the FLAASH atmospheric correction algorithm. Proceedings of the IEEE International Geoscience and Remote Sensing Symposium IGARSS '03, 90 92. Gomez, R. B. (2002). Hyperspectral imaging. a useful technology for transportation analysis. Optical Engineering, 41, 2137. doi:10.1117/1.1497985. Gong, P., Pu, R., & Yu, B. (1997). Conifer species recognition. An exploratory analysis of in situ hyperspectral data. Remote Sensing of Environment, 62, 189200. Hair, J. F., Anderson, R. E., Tatham, R. L., & Black, W. C. (1995). Multivariate data analysis with readings (4th ed.). Englewood Cliffs: Prentice-Hall.

69 Hallett, R. A., Hornbeck, J. W., & Martin, M. E. (1997). Predicting elements in white pine and red oak foliage with visible-near infrared reflectance spectroscopy. Journal of Near Infrared Spectroscopy, 5, 7782. Herold, M., Roberts, D. A., Gardner, M. E., & Dennison, P. E. (2004). Spectrometry for urban area remote sensing Development and analysis of a spectral library from 350 to 2400 nm. Remote Sensing of Environment, 91, 304319. Huenia, A., Niekeb, J., Schopfer, J., Kneubhler, M., & Itten, K. I. (2009). The spectral database SPECCHIO for improved long term usability and data sharing. Computers & Geosciences, 35, 557565. Jungho, I., & Jensen, J. R. (2008). Hyperspectral remote sensing of vegetation, geography compass. 2, 19431961. Kemsley, E. K., Ruault, S., & Wilson, R. H. (1995). Discrimination between Coffee arabica and Coffee canephora variant robusta beans using infrared spectroscopy. Food Chemistry, 54, 321326. Kokaly, R. F., King, T. V. V., & Livo, K. E. (2008). Airborne hyperspectral survey of Afghanistan 2007. flight line planning and HyMAP data collection, USGS Afghanistan Project Product No. 186, U.S. Geological Survey, Reston, Virginia. Kruse, F. A., & Perry, S. L. (2009). Improving multispectral mapping by spectral modeling with hyperspectral signatures. Journal of Applied Remote Sensing, 3, 033504. doi:10.1117/1.3081548. Kruse, F. A., Lefkoff, A. B., Boardman, J. B., Heidebrecht, K. B., Shapiro, A. T., Barloon, P. J., & Goetz, A. F. H. (1993). The Spectral Image Processing System (SIPS)interactive visualization and analysis of imaging spectrometer data. Remote Sensing of Environment, 44, 145163. Kumar, L. (2007). A comparison of reflectance characteristics of some Australian eucalyptus species based on high spectral resolution datadiscriminating using the visible and NIR regions. Journal of Spatial Science, 52, 5164. Kutser, T., Millerb, I., & Jupp, L. B. (2006). Mapping coral reef benthic substrates using hyperspectral space-borne images and spectral libraries. Estuarine, Coastal and Shelf Science, 70, 449460. Lister, S. J., Dhanoa, M. S., Stewart, J. L., & Gill, M. (2000). Classification and comparison of Gliricidia provenances using near infrared reflectance spectroscopy. Animal Feed Science and Technology, 86, 221238. Mazer, A. S., Martin, M., Lee, M., & Solomon, J. E. (1988). Image processing software for imaging spectrometry analysis. Remote Sensing of Environment, 24, 201210. Nidamanuri, R. R., & Zbell, B. (2010). A method for selecting optimal spectral resolution and comparison metric for material mapping by spectral library search. Progress in Physical Geography, 34, 4758. Nidamanuri, R. R., Garg, P. K., & Ghosh, S. K. (2007). Development of an agricultural crops spectral library and classification of crops at cultivar level using hyperspectral data. Precision Agriculture, 8, 173185. Price, J. C. (1994). How unique are spectral signatures? Remote Sensing of Environment, 49, 181186. Purkisand, S. J., & Pasterkamp, R. (2004). Integrating in situ reef-top reflectance spectra with Landsat TM imagery to aid shallowtropical benthic habitat mapping. Coral Reefs, 23, 520.

70 Shepherd, K. D., & Walsh, M. G. (2002). Development of reflectance spectral libraries for characterization of soil properties. Soil Society of America Journal, 66, 988998. Van der Meer, F. (2005). The effectiveness of spectral similarity measures for the analysis of hyperspectral imagery. International Journal of Applied Earth Observations and Geoinformation, 8, 317. Wegehenkel, M., Brozio, S., Mirschel, W., Prietzsch, C., & Zbell, B. (1999), Simulation tool for the evaluation of agricultural productivity (STEAP (Ertragspotential)).-in.ProSmart-Forschungs-Endbericht(DLR Frderkennzeichen. 50EE9816),

J Indian Soc Remote Sens (March 2013) 41(1):5770 DaimlerChrysler Aerospace, Dornier Satellittensysteme, Dok.No.. EB-DSS-REP-0001. S. 11/1-11/62. Yang, X., & Tien, D. (2010). An automated image analysis approach for classification and mapping of woody vegetation from digital aerial photograph. World Review of Science, Technology and Sustainable Development, 7, 13 23. Zomer, R. J., Trabucco, A., & Ustin, S. L. (2009). Building spectral libraries for wetlands land cover classification and hyperspectral remote sensing. Journal of Environmental Management, 90, 21702177.