You are on page 1of 4

MAPPING AGRICULTURAL CROPS WITH EO-l HYPERION DATA

Konstantinos D. Ntouros', Ioannis Z Gitas', Georgios N Silleos 3

1. Lab ofRS and GIS, School of Agriculture, Aristotle University of Thessaloniki, Greece.
2. Lab ofRS and Forest Management, School of Forestry and Natural Environment, Aristotle University of Thessaloniki,
Greece.
3. University of Macedonia, Department of Applied Informatics, Thessaloniki, Greece

ABSTRACT spectral resolution of airborne instruments with the practicality of


satellite remote sensing. Previous studies have indicated that forest
Hyperspectral data acquired by the Hyperion instrument, on board species were classified with greater accuracy by the use of
the Earth Observing - 1 Satellite (EO-I), were evaluated for the Hyperion images than multispectral data [13]-[14]. Furthermore,
classification of five agricultural crops (maize (Zea mays), cotton Hyperion's data were used for crop diseases detection [15],
(Gossypium hirsutum L.), rice (Oryza Sativa), tobacco tNicotiana discrimination between sugarcane varieties [16], land use/land
Tabacum) and tomato (Lycopersicon esculentumi in Greece and cover classification [17] and estimation of leaf nitrogen content
the results were compared to classification of Landsat 5 TM data. [18].
In addition, was investigated the contribution of Hyperion SWIR The aim of the study was to investigate the potential of
bands on crops classification. The research was conducted in an Hyperion data in discriminating and mapping agricultural crops.
agricultural area located in the North-Eastern Greece. The The specific objectives were i) to map the main agricultural crops
Hyperion radiance values, from the 196 bands, were converted into appearing in the study area with Hyperion imagery, ii) to evaluate
surface reflectance values using the Fast Line-of-sight the contribution of Hyperion SWIR bands on crop classification
Atmospheric Analysis of Spectral Hypercubes (FLAASH) model and iii) to compare the results with Landsat 5 TM imagery, in
which is embedded in ENVI software. The data dimensionality terms of classification accuracy.
reduction of Hyperion's image bands was achieved by using MNF
transformation whereas the Maximum Likelihood algorithm was 2. METHODS
used in order to perform image data classification. The results
showed that an overall accuracy of 91% was obtained from the 2.1. Study site
classification of Hyperion image, while the overall accuracy
resulted from the classification of Landsat 5 TM image was 81%. The study was conducted in an agricultural area in the North-East
Also the Hyperion SWIR bands provide additional information on part of Greece, between Kavala and Xanthi Prefectures, in about
crops classification, not available by VNIR bands. 200 km distance from the city of Thessaloniki and about 40 Km
away from Kavala city (Fig.1).
Index Terms - Hyperion, Crops classification, Hyperspectral The study area presents a flat terrain and an average land
Remote Sensing, agriculture, Minimum Noise Fraction elevation of about 70m above the sea level. The dominant crop in
the area is maize followed by rice and cotton. In a lesser extent,
1. INTRODUCTION tobacco and tomato crops also co-exist, whereas, other agricultural
crops such as asparagus, sugar beet, wheat and alfalfa can also be
The significance of crop mapping at regional as well as at global found, but in a non-significant proportion
scale, for planning and management of agriculture production, has Following the phenological stages of the five spring
been recognized by international organizations such as the World agricultural crops of main interest (maize, cotton, rice, tobacco and
Bank and the Food and Agriculture Organization (FAO) of the tomato), an image acquisition during August proved to be the most
United Nations [1]. Over the last 30 years the multispectral satellite suitable period, due to great profound spectral differentiations
data has been extensively used for monitoring agricultural between the crops [19].
production and [2] to support European policies such as the For that reason, an archived Hyperion image of the study area,
Common Agriculture Policy (CAP) [3]. acquired on August 1, 2003 was selected along with an archived
Airborne imaging spectrometry data have been used by Landsat 5 TM image acquired one day later (August 2, 2003).
geologists since the early 1980's [4] while airborne hyperspectral
data were used in vegetation studies, such as crop classification 2.2. Data preprocessing
under different stress conditions [5]-[6], weeds detection [7]-[8]-
[9], crop residues mapping [10]-[11] and crop diseases detection Since Hyperion operates from a space platform, the effects of
[12]. However, the high cost and the limited coverage of airborne atmospheric absorption results in a lower surface signal level at
sensors constituting restrictive factors for their use in wider scale orbital altitude. As a result, before using Hyperion data a
studies. preprocessing stage is required, dealing with problems of
Hyperion, the first spaceborne hyperspectral sensor, miscalibration and noise [20].
overcomes the limitations mentioned above, since merges the

978-1-4244-4687-2/09/$25.00 ©2009 IEEE


geocorrected image had an estimated root mean square error of
0.1811 pixels or about 5.43m providing adequate spatial accuracy.
Landsat 5 'I'M image (six non thermal bands) was atmospheric
corrected using FLAASH model; thereafter georectification was
performed in an image subset. As base map, a mosaic of
panchromatic orthorecti fied aerial-photomaps was used again and
4 12 ground control points were located throughout the image. A l SI
',. ,. .-
.
.: + ~- '".:= ,-= =~ .:
order polynomial transformation was applied and the resulting
transformation had an estimated total root mean square error of
t " .... 0.2118 or about 6.35m.
'>
I t,"

0, .. ~ . / . Finally, I-Iyperion image as well as Landsat 'I'M image was


oJ- .. i ll! .... • 4

~ ) . . -"f masked out in order to eliminated the non agricultural land cover
types such as settlements, roads, riparian vegetation and rocky non
. . . . .., j arable areas.

2.3. Image analysis


Fig. I. Study site
Although, hyperspectral data provides significant advantages when
For many applications such as multi- temporal studies, multi- compared with convent ional remote sensor systems in various
sensor studies and use of spectral libraries, atmospheric effects applicat ions, narrowband s image analysis appears to be more
have to be removed prior to image analysis [20] - [21]. In this complicate due to its data dimensionality [23].
study, Hyperion bands radiance values were converted into surface In this study, in order to minimize the uncorrelated spatial
reflectance values using Fast Line-of-sight Atmospheric Analysis noise of post-atmospheric reflectance image and subsequently to
of Spectral Hypercubes (FLAASI-I) model which is embedded in overpass the data dimensionality the minimum noise fraction
ENYI software. FLAASH model is based on MODTRAN 4 (MNF) transformation was used [15]-[21]-[24]. The transformation
radiation transfer code and a standard atmospheric model was used was performed separately for VNIR and SWIR bands of Hyperion
(Mid - Latitude Summer) in order to minimize scattering and image due to its differentiated noise structure in those two
absorption effects of several atmosphere components (e.g., water arrays[20]. The analysis was implemented with ENYI software
vapor, carbon dioxide, ozone, oxygen). Also the atmospheric water [25] and resulted in II eigenimage s for VNIR and 8 eigenimages
vapor technique was used for the calculat ion of column water for SWIR bands respectively, after examination of eigenvalue plots
vapor content for each pixel in the image [22]. as well as visual inspection of eigenimages . The 19 eigenimages
Out of 242 total I-Iyperion bands only 135 (bands : 8 to 57, 84 were 'stacked ' in a single image file for the subsequent
to 96, 103 to 118, 135 to 163 and 191 to 217) were used in the classi fication analysis.
image analysis because many of the bands have low signal to noise For the classification comparison between the two sensors,
ratio or other problems. Hyperion images Level I R have 44 supervised maximum likelihood classification was selected for
uncalibrated bands (I to 7, 58 to 76 and 225 to 242) and arc set to 0 each sensor. Additionally maximum likelihood classification was
values by USGS. Four out of 198 calibrated bands arc in the performed on II eigenimages (YNIR) in order to evaluate the
overlap bandwidth between VNIR and SWIR spectrometers. These contribution of Hyperion SWIR bands on the crops classification.
arc YNIR 56 (915.23nm) and 57(925.40nm) and SWIR Ground reference data of the study area for the year of the
77(912.45nm) and 78(922.54nm). The band 77 and 78 were image acquisition were obtained by the Payment and Control
eliminated because have higher level of base noise [20]. An Agency for Guidance and Guarantee Community Aids
additional 61 bands that are strongly affected by atmospher ic water (OPEKEPE) of Greece. These are cadastral GIS polygon layers
vapor absorption, were identified and excluded prior to image which arc created annually in the framework of CAP commitment
analysis process. agreement, regarding the inspection of EC subsidiary benefits,
Hyperion images also display vertical stripes as an outcome containing among other information in the attribute table, the crop
of miscalibrated or dead detectors in either the YNIR or SWIR type of each arable property. Therefore; the GIS cadastral layer of
arrays, since Hyperion is a pushbroom technology sensor, the study area was overlaid on Hyperion image and regions of
collecting data simultaneously across 256 detectors along the track interest (ROI) for each crop type of interest were picked out
of the satellite's orbit [20] - [21]. These striping (abnormal) pixel manually by selecting only those pixels that were coinciding within
have lower DN values compared to their neighbor pixels. The the parcels boundaries. At the end, the ROI' s were divided to the
approach that was followed in order to normalize this effect was training and testing ROI's with the use of stratified random
the replacement of DN values of these striping pixels with the approach, leaving out the 30% of ROI's for testing of the
average DN values of their immediate left and right adjacent supervis ed classification.
pixels, since neighboring pixels have the highest spatial Supervised classification for Hyperion image as well as for
autocorrelation [13]. Landsat 5 'I'M image was applied using ENYI software and
A mosaic of panchromat ic orthorect ified aerial-photomaps classification accuracy assessments were performed using the error
granted by Hellenic Ministry of Rural Development and Food was matrices generated from the testing ROI' s through post
used as a base map for georectification of Hyperion image. Twelve classification ENYI' s module [26].
ground control points which were well distributed throughout the
image were located. A I st order polynomial transformat ion was
performed and nearest neighbor resampling method was used in
order to retain as much as possible spectral integrity. The
Table I
Classification accuracies for Hyperion - Landsat 5 TM

HYPER ION (19 HYPER ION (11


~ 5 T M
MNF) MNF)
Accu racy (%) Accurac y (%) Accuracy (%)
Class es PA UA PA UA PA UA
maize 94.81 96.97 87.76 88.66 92.59 97.4
cotton 85 8793 69.49 65.08 7917 8796
rice 100 96.77 8609 94.24 99.52 96.31
t obacco 78.18 84.31 67.14 73.44 74.55 78.85
tomato 64.91 5522 465 1 29.41 6667 44.71
Overall
91.62% 81.21% 89.49%
accuracy
KHA T 0.8775 0.7307 0.8478

individual error matrices is another measure of agreement or


accuracy . So KHAT value of Hyperion image classification
(KHAT: 0.88) indicates strong agreement (KHAT > 0.80) in
contrast to the Landsat TM image classification (KHAT: 0.73)
which exhibit moderate agreement with the reference data (0.40 <
KHAT < 0.80) [26].
Table I summarizes the classification results for the two
sensors . Highest accuracies were achieved for Hyperion data for all
agricultural crop classes.
The classification of Hyperion data (19 MNF eigenimages)
showed that three (rice, maize and cotton) out of five crops were
classified with accuracies greater than 85% while the lowest
accuracy was obtained for the tomato crop. On the other hand,
Landsat 5 TM image classification showed that only two (rice and
com) out of the five crops were classified with accuracies greater
than 85%. Also, the lowest accuracy was produced for the tomato
crop. Probably, one of the reasons that tomato was classified with
that low accuracy in both cases short-range area of land properties
LEGEND employed for the tomato crop in the study area in relation to the
Class_Names spatial resolution of the sensors.
Maize
The classification images are depicted in Fig 3(a) (I-Iyperion -
19 MNF eigenimages), Fig 3(b) (Landsat 5 TM) and Fig 3(c)
• Cotton
(Hyperion -II MNF eigenimages). The classification legend is
• Rice shown in Fig 4(c).
• Tobacco The comparison of the Hyperion classified images (table I)
• Tomalo which were derived from using of 19 MNF and II MNF
eigenimages respectively, showed that the overall accuracy in the
first case was slightly higher (- 2%) but this minimal difference it
is not statistically significant as it was proven after applying Z test
statistics test (Z = 1.48, Z eritiea! = 1.96 for 95% confidence level)
[26].Nevertheless, as it shown on table I, the cotton crop among
the other ones was classified with higher accuracy (6%) when 19
(c) (d) MNF eigenimages were used. So the using of SWIR bands
provides additional information not available by VNIR bands [23].
With MNF transformation data compaction of 86% (19 MNF
Fig. 2. (a) Hyperion classification (19 MNF). (b) Landsat 5 TM eigenimages) or greater, almost 92% (II MNF eigenimages) was
classification. (c) Hyperion classification (IIMNF). (d) achieved, since classification accuracy was marginally increased
Classification legend (2%) when the 8 additional MNF eigenimages were used.
Hyperspectral data dimensionality reduction is a key issue for data
3. RESULTS AND DISCUSSION processing; i.e. traditional image classification techniques can not
be performed in particular cases. Specifically, while dimensionality
The comparison between error matrices of Hyperion and Landsat of the image data increases, the minimum number of training
TM classified images has shown that the Overall Accuracy of pixels per class must be increased, as well, so that the accuracy of
91.03% was obtained for Hyperion 's image classification, which the statistical estimations is preserved [27].
additionally provided a further increase in accuracy of about 10%
over six non thermal bands of Landsat 5 TM (Overall Accuracy:
81.21%). Apart from overall accuracy, KHAT value of the
4. CONCLUSION [11J A.Bannari, A.Pacheco, K.Staenz, H.McNairn, K. Omari,
"Estimating and mapping crop residues cover on agricultural lands
In this study, we evaluated the implementation of Hyperion image using hyperspectral and IKONOS data", Remote Sensing of
data in the discrimination and mapping of agricultural crops and Environment, 104, pp.447--45, 2006.
compared with broadband sensor data such as Landsat 5 'I'M. [12J M. Zhang, Z.Qin, X.Liu, S. L. Ustin, "Detection of stress in
The analysis of data showed that the use of I-Iyperion data is tomatoes induced by late blight disease in California, USA, using
more complex than the use of multispectral data and a pre - hyperspectral remote sensing", International Journal of Applied
processing is needed before image analysis. Additionally, the Earth Observation and Geoinformation, 4, pp. 295-310, 2003.
' problem' of data dimensionality has to be addressed especially [13J D.G.Goodenough, ADyk, K. O.Niemann, I. S.Pearlman,
when traditional image analysis techniques are employed. H.Chen, 'I' Han, M.Murdoch, C. West, " Processing Hyperion and
I-lyperion image classification exhibits of 10% greater ALI for Forest Classification", IEEE Transactions on Geoscience
accuracy in comparison with Landsat 5 'I'M image classification. and Remote Sensing, 41,6, pp. 1321-1331,2003.
All agricultural crop classes were better classified using Hyperion [14J P.S.Thenkabail , E.AEnclona, M. S. Ashton, C.Legg, M. Jean
image especially cotton, tobacco and tomato. De Dieu, "Hyperion, IKONOS, ALI, and ETM+ sensors in the
Also the results indicate that the Hyperion SWIR bands study of African rainforests ", Remote Sensing of Environment, 90,
provide additional information useful for the mapping of the five pp. 23--43, 2004.
main crops and especially for the cotton . [15J A.Apan, A Held, S. Phinn, J. Markley , "Detecting sugarcane
Although, the classification of the Hyperion image resulted in 'orange rust' disease using EO- I I-Iyperion hyperspectral imagery",
very high overall accuracy (91%) further evaluation is needed on International Journal of Remote Sensing, 25, 2, pp. 489--498,
other agricultural crops and of different phenological stages. 2004 .
[16J S.Galvao, A R.Formaggio, D. ATisot, "Discrimination of
5. REFERENCES sugarcane varieties in Southeastern Brazil with EO-l Hyperion
data" , Remote Sensing of Environment, 94, pp523-534, 2005.
[I] Y.Cohen, M.Shoshany, "A national knowledge-based crop [17J Bing Xu and Peng Gong, "Land use/Land cover classification
recognition in Mediterranean environment", International Journal with multispectraland hyperspectral EO- I data", Photogrammetric
of Applied Earth Observation and Geoinformation, 4, pp. 75--87, Engineering & Remote Sensing, 73, 8, pp. 955-965, 2007
2002. [18J N.C.Coops, M.L.Smith, M.E.Martin, S.V Ollinger,"Prediction
[2] M. S.Moran, Y.lnoue and E. M. Barnes, "Opportunities and of Eucalypt foliage nitrogen content from satellite- derived
Limitations for Image-Based Remote Sensing in Precision Crop hyperspectral data", IEEE transactions on Geoscience and Remote
Management", Remote Sensing of Environment, 61, pp.319 -346, sensing, 4, 6,ppl-9, 2003.
1997. [19J N. Silleos, N. Misopolinos and K. Perakis, "Relationships
[3] http://mars.jrc.ec.europa.eu/mars between remote sensing spectral indices and crops discrimination.
[4] F. A.Kruse, J. W.Boardman and J. F.I-Iuntington, "Comparison Geocarto International ", 7:2, pp. 41 - 51, 1992
of Airborne Hyperspectral Data and EO- I I-Iyperion for Mineral [20J B. Datt, T. R.McVicar, T. G.Van Niel, D. L. B Jupp,
Mapping", IEEE Transactionson on Geoscience and Remote J.S.Pearlman , "Preprocessing EO-I Hyperion J-1yperspectral Data
Sensing,41, 6, pp.1388-1399, 2003. to Support the Application of Agricultural Indexes", IEEE
[5] P.K Goel, S.O Parsher, R.M.Patel, J.A. Landry, R.B. Bonnel, transactions on Geoscience and Remote sensing" 41, 6, pp.1246-
A.A. Viau, "Classification of hyper spectral data by decision trees 1259,2003
and artificial neural networks to identify weed stress and nitrogen [21J B.W.Pengra,C.A Jonhston and T.R.Loveland, "Mapping an
status of corn", Computers and Electronics in Agriculture, 39 ,pp. invasive plant, Phragmites australis, in coastal wetlands using the
67-9,2003 . EO-l Hyperion hyperspectral sensor", Remote Sensing of
[6] C.C.D Lelong, P.C.Pinet, H.Poilve, "Hyperspectral Imaging Environment, 108:1, pp.74 -81, 2007.
and Stress Mapping in Agriculture : A Case Study on Wheat in [22J ENVI FLAASH Module User's Guide
Beauce (France)" , Remote Sensing of Environment, 66, pp. 179- [23J P.S.Thenkabail , , E.A.Enclona, M.S .Ashton, B.Van Der
191,1998. Meer, "Accuracy assessments of hyperspectral waveband
[7] A.P.WilIiams, R.E.Hunt, "Estimation of leafy spurge cover performance for vegetation analysis applications", Remote Sensing
from hyperspectral imagery using mixture tuned matched of Environment, 91,pp. 354-376. 2004
filtering", Remote Sensing of Environment, 82, pp. 446----456, 2002. [24J E.Underwood, S. Ustin, D. DiPietro,"Mapping nonnative
[8] J. T.Mundt, N. F.Glenn, K.T.Weber, T.S. Prather, L.W. Lass, I. plants using hyperspectral imagery", Remote Sensing of
Pettingill, "Discrimination of hoary cress and determination of its Environment, 86, pp. 150-16, 2003
detection limits via hyperspectral image processing and accuracy [25J ENVI User's Guide
assessment techniques", Remote Sensing of Environment, 96, pp. [26J R.G.Congalton, K.Green, "Assessing the Accuracy of
509 - 517,2005. Remotely Sensed Data: Principles and Practices". Lewis
[9] N.Glenn, J. T.Mundt, K.T.Weber, T. S.Prather, Publishers, pp 43-53, 1999
I.Pettingill,"l-lyperspectral data processing for repeat detection of [27J J. A.Richards, J. Xiuping , "Remote Sensing Digital Image
small infestations of leary spurge",Remote Sensing of Analysis", Springer-Verlag Berlin Heidelberg, pp 364-365,2005
Environment, 95, pp. 399--412, 2005
[10] C.S.T. Daughtry, P.C.Doraiswamy, Jr., E.R.Hunt, AJ .Sterna,
I.E.McMurtrey, J.H.Prueger, "Remote sensing of crop residue
cover and soil tillage intensity", Soil & Tillage Research, 91,pp.
10 1-1 08, 2006.

You might also like