Remote Sensing of Crop Residue Using Hyperion (EO-1) Data A. Bannari 1, K. Staenz 2 and K. S. Khurshid 3 1
Remote Sensing and Geomatics of Environment Laboratory
Department of Geography, University of Ottawa, Ottawa (Ontario) Canada K1N 6N5 Phone (613) 562-5800 (Ext. 1042), Fax (613) 562-5145; Email:
[email protected] 2 Alberta Terrestrial Imaging Centre (ATIC)/University of Lethbridge 401, 817 - 4th Avenue South, Lethbridge, Alberta, Canada T1J 0P 3 Agriculture and Agri-food Canada, 960 Carling Avenue, Ottawa (Ontario) K1A 0C6 Canada Abstract - The goal of this research was to investigate the
potential of hyperspectral Hyperion (EO-1) data and constrained linear spectral unmixing analysis (CLSMA) for percent crop residue cover estimation and mapping. The Hyperion image data was acquired at the beginning of the agricultural season, May 20 2002, as well as ground reference measurements for validation purposes. In this period, there is mainly only the presence of bare soil and crop residue before any crop cover development. The image data were corrected for the sensor artifacts: a spatial misregistartion between the VNIR and SWIR data, and striping, and, in addition, the noise was reduced. The data were atmospherically corrected and then transformed to surface reflectance and, subsequently, corrected for sensor smile/frown and post-processed to remove residual errors. In order to extract the crop residue fraction (percent cover), the image was unmixed using the pure spectra (endmember) collected in the field simultaneously with Hyperion data from different targets (dry and wet wheat residue, and bright and dark soil) with a GER3700 spectroradiometer. In order to represent the existing trees and all the photosynthetically targets in the scene, a vegetation endmember was selected from our spectral library. Correlation between ground references measurements and extracted fractions from Hyperion data using CLSMA showed that the method satisfactorily predicts crop residues percent cover (D of 0.94, R2 of 0.73 and RMSE of 8.7%) and soil percent cover (D of 0.91, R2 of 0.70 and RMSE of 10.03%).
Keywords: crop residue, hyperspectral remote sensing, Hyperion, Soil and Erosion, Unmixing, precision agriculture.
I. INTRODUCTION On agricultural lands, conservation of post-harvest crop residues plays an important role in the protection of the soil surface against water and wind erosion [1]. As for the environmental aspect, crop residue improves soil structure, increases organic matter content in the soil, positively influences water infiltration, evaporation and soil temperature [2] and plays an important role in fixing CO2 in the soil [3]. Accordingly, good residue management practices on agricultural lands have many positive impacts on soil quality and crop production. In the literature, several spectral indices for quantitative evaluation of the crop residue cover were developed [4]. However, these methods are not sufficiently rigorous and accurate for discriminating residue from bare soil
and for estimating the fraction of residue cover. They do not consider the spectral mixture of different materials in the same pixel [5]. Moreover, the broad spectral and spatial resolutions of multispectral sensors do not allow an accurate analysis of the composition of the target materials on the ground [6]. Using hyperspectral (Probe-1) data and constrained LSMA approach; the estimation of percent residue cover was improved. However, the results were not entirely satisfactory, since the image data were acquired when the crop vigour was very high [7], causing the crop residue be shaded and/or completely covered by the plants Acquisition of imagery postharvest or prior to spring seeding would simplify the extraction of percent residue cover by eliminating confusion with the crop canopy. The objective of this research is to investigate the potential of the Hyperion (EO-1) hyperspectral 30-m data and CLSMA for crop residue fraction cover estimation and mapping. The Hyperion image data were acquired at the beginning of the agricultural season, May 20 2002, as well as ground-based reference measurements for validation purposes. In this period, there is only the presence of bare soil and residue and no crop cover. In order to extract the crop residue fraction, the image was unmixed using the spectral range from 427 nm to 2355 nm. Endmembers were collected in the field, simultaneously with Hyperion data from different targets using a GER-3700 spectroradiometer. Crop residue and soil fractions for each sample point were extracted from unmixing results and compared to ground-based reference measurements. II. MATERIAL AND METHODS A. Study Site The data were collected in an agricultural region near Indian Head (50oN, 104oW), approximately 70 km east of Regina, Saskatchewan, Canada. The selected site is an intensively cultivated agricultural area dominated by Black Chernozemic soils, which were developed on neutral to slightly alkaline and uniform clayey lacustrine deposits. Principal economic activities of this area are based on agricultural practices. Major crops grown are wheat, pea, canola and corn. The test fields are located on a precision agricultural farm near the Indian Head Agricultural Research Foundation (IHARF). This experimental farm was used in the context of a large project to investigate the potential of hyperspectral remote sensing in precision agriculture.
1
B. Image data acquisition Image data were acquired on May 20, 2002 with the Hyperion hyperspectral sensor on NASA’s Earth Observer-1 (EO-1) platform. The launch of the Hyperion sensor in November 2000 marked the first test of a space-borne hyperspectral sensor covering both the VNIR and SWIR spectral regions [8]. Hyperion is a pushbroom imaging spectrometer that collects data in the along-track direction. This sensor collects the upwelling radiance in 242 spectral bands, each 10 nm wide with an average spectral sampling interval of 10 nm. Hyperion has a single telescope and consists of two spectrographs, one covering the VNIR wavelength range from 357 nm to 1055 nm, the other, the SWIR from 851 nm to 2576 nm. Since Hyperion is a pushbroom sensor, the entire swath is obtained in a single frame. Its telescope images the Earth onto a slit with a field-of-view (FOV) of 0.624°, resulting in a swath width of 7.65 km from a 705 km altitude. Each data set acquired by this sensor covers a nominal alongtrack length of 40 km. C. Ground data acquisition Ground measurements were acquired simultaneously with Hyperion data. Spectroradiometric measurements were acquired in the field from different targets (dry and wet wheat residue, and bright and dark soil) using a GER-3700 spectroradiometer [9]. This instrument measures radiance over a spectral range from 400 to 2500 nm in 704 spectral bands varying from 1.5 nm to ∼20 nm at full-width half-maximum (FWHM). Ground reflectance was calculated by rationing target radiance to the radiance obtained from a calibrated 25cm by 25-cm white Spectralon panel [10]. Corrections were made for the wavelength dependence and non-lambertien behaviour of the panel. The Spectralon radiance was acquired immediately before the target radiance. The average of twenty spectra was convolved with Gaussian response profiles to match the bandwidths and the band centres of the Hyperion EO-1 sensor. For the validation of the results, with ground-based reference measurements, multiple vertical photographs, covering more than one image pixel, were taken above several sampling points from a 2-m height above the ground. Five photos were collected at each sampling point. These photos were classified using an unsupervised Isodata classifier in PCI Image Works [11] and aggregated into three classes representing crop residue, soil and vegetation. Fractions of each class were converted to percentages of total area. Percentages derived from each photo were averaged per site and compared to the extracted soil and residue fractions for the 77 sites in the 10 fields. Ground sample locations were referenced in the field using a GPS and located on an orthorectified Quickbird image collected over the area in July 2002. The Quickbird image was then referenced to the Hyperion data with sub-pixel accuracy using GCP works in PCI Geomatica, and pixel/line locations for each sample point were recorded. This way the hyperspectral data is not affected by resampling. D. Image data pre-processing The Hyperion EO-1 data were pre-processed with an aim to correct for sensor artifacts and atmospheric and geometric effects [12] using the Imaging Spectrometer Data Analysis System (ISDAS) developed at the Canada Centre for Remote Sensing [13]. The procedure begins with a spatial misregistartion between the VNIR and SWIR data. The rotation and associated spatial resampling of the SWIR data was carried out after removal of the striping in the data. Prior to destriping the data, the column dropouts caused by dead pixels
were then replaced with values interpolated between the closest valid pixels in the same or different bands [14]. The data cube was subsequently analyzed to characterize the distortions of keystone and spectral smile [15]. At this step, the data were cropped to exclude noisy bands resulting in a final data set that spans the spectral range from 426.82 nm to 2355.20 nm with a total of 192 bands (excluding the overlap bands between the VNIR and SWIR spectrographs). For our data, the keystone distortion varied from a minimum of - 0.075 pixels to a maximum of 0.3 pixels for VNIR and was about 0.1 pixels for SWIR [12]. A method developed by Neville et al. [15] was applied to determine the smile/frown using known atmospheric absorption features. Wavelength shifts (smile/frown effect) were calculated, approximately between 1 and 3 nm in the VNIR and SWIR, and applied after the atmospheric correction process. The calibrated at-sensor radiance data were converted to surface reflectance using a look-up table (LUT) approach to correct for the atmosphere [16]. Subsequently, the reflectance data were corrected for smile/frown effects and the bandwidth using a spectral resampling technique as described in the reference [15] to achieve a set of common band wavelength centres and bandwidths for the entire data cube. Finally, the post-processing concluded the corrections by removing residual errors that still remained after the correction of sensor artifacts and atmospheric effects [17]. E. Image data processing In order to extract the crop residue fraction, a CLSMA [18] was performed on Hyperion imagery using an algorithm implemented in ISDAS [19], and was carried out using the full spectral range from 426.82 nm to 2355.20 nm. In spite of the omission of the multiple scattering effects among different targets in the pixel [20], the errors associated with the linear assumptions have been found insignificant [21]. LSMA assumes that the source of the spectral signature from a pixel is more than one spectrally distinct component [22]. The fundamental assumption of LSMA is that generally each pixel is a physical mixture of several constituents weighted by surface abundance, and the spectrum of the mixture is a linear combination of the endmember reflectance spectra. Endmember Spectra were collected in the field from different targets (dry and wet wheat residue, and bright and dark soil), simultaneously with the Hyperion data using a GER spectroradiometer. In order to represent the existing trees and all the photosynthetically targets in the scene, a vegetation endmember was selected from our spectral library. These endmembers were measured from “pure” patches, which were considered representative of the wheat residues and soils in the study site. It is important to mention that a crop shadow endmember was not included in the endmember selection process given that the study area is relatively very flat and crop cover was absent or at the beginning of the development and, therefore, no shadow component exists in the considered fields. As a result, the crop residues fraction map was derived from the Hyperion data for each individual field. The values range from 0 to 100%, where 0% indicates a low abundance and 100% a high one. Finally, the percent crop residue cover (PCRC) and the percent soil cover (PSC) were validated with the ground vertical photographs. F. Statistical Analyses Statistical analyses were computed with the Statisticasoftware package [23]. Various statistics were computed for both ground measurements (observed values) and image data
2
(predicted values). Standard deviation statistics allowed the evaluation of data variability. This parameter was reported in all cases as an error percentage of the average extracted from ground measurements (vertical photographs) and image data. To validate PCRC and PSC, ground measurements and image data were compared using the 1:1 line. Ideally, observed and predicted values should have a correspondence of 1:1. An index of agreement reflects the degree to which the observed value is accurately estimated by the predicted value. This index provides a measure of the degree to which a model’s predictions are error free. The index ranges between 0 and 1, with 1 indicating a perfect match between observed and predicted values. The observed values were those calculated from the vertical photographs and the predicted values were from the abundance maps. The root mean square error (RMSE) was used as an additional measure as an overall error to supplement the index of agreement described above [23]. This error also quantifies the 1:1 relationship between observed and predicted values. The relationships between observed and predicted values were also analyzed using a linear regression model. The coefficient of determination (R2) of the regression model was used to evaluate the strength of the linear relationship between observed and predicted values.
0%
100%
Figure 1: Percent Crop Residue Cover (PCRC) Maps derived from Hyperion cube using Unmixing. III. RESULTS AND DISCUSSION The used endmember spectra for the CLSMA were dray and wet wheat residue, bright and dark soil, and vegetation. The residue spectra are similar to the soil spectra in the VNIR region of the electromagnetic spectrum and differ only in amplitude, but they differ considerably in the SWIR region (1600-2400 nm). The moisture reduced the spectral signature of the residue across all wavelengths. The two considered soils have similar spectra with a little difference in amplitude, which is likely due to differences in their colour and moisture content. The dry and wet residue spectra show the typical cellulose absorption features around 2100, 2280 and 2340 nm; and lignin absorption features around 1680, 2270, 2330 and 2380 nm. Theses absorption features are absent in the soil and
vegetation spectra and offer the possibility to discriminate among residue, soil and green vegetation. Similar observations have also been made by [4 & 24]. In order to extract the crop residue fraction, the image reflectance cube was unmixed considering the entire spectral range from 427 to 2355 nm. The percent crop residue cover (PCRC) derived from the CLSMA was validated against the percent cover estimated from ground vertical photographs (PCRCG). The relationship between the two variables was established and statistics were generated. Figure 1 shows the derived fraction maps for residue from the Hyperion EO-1 data. Pixels in blue and red colour represent the very low (0%) and very high (100%) percent crop residue cover, respectively. This figure shows the spatial distribution and variability within the fields. Independent of the fraction cover, the spatial variability of the ground reference measurements was similar to the spatial variability found in the fractions derived from the hyperspectral data. The correlations between PCRC retrieved from the Hyperion data and PCRCG ground reference values calculated from ground vertical photographs show an excellent index of agreement (D of 0.94) with a relatively low RMSE (8.7%). The scatter plot as shown in Figure 2 reveals a satisfactory linear relationship between the two variables (R2 of 0.73) and an overestimated resulting in data points not very well fitting the 1:1 line. This is probably due to the consideration of only the endmembers of wet and dry wheat residue. Although this residue type is dominant in the study area, there are other crop residue types such as canola, barely, corn and bean present. However, there are very sparsely distributed and do not cover an area fully. The use of a specific endmember representing these crop residue types may improve the results. The contribution of the shadow component was not significant since the image data were acquired early in the beginning of the agricultural season without any significant crop development. Probably, the location of the ground sampling points in the Hyperion imagery (30 m by 30 m pixel size) and classifying the ground photos could be also a source of this overestimation of the crop residue. Nevertheless, despite this small overestimation, the relationship between the percent crop residue cover, estimated from Hyperion data and ground reference values, which were calculated from ground vertical photographs, is generally satisfactory modeled. Figure 3 shows the relationship between the percent soil cover derived from the Hyperion data and the ground-based photographs. Compared to the ground reference data, the model was overall a good predictor for soil percent cover with an index of agreement (D) of 0.91, a RMSE of 10.03 % and a R2 = 0.70. The scatter plots depicting the relationship between the PSC and PSCG, clearly illustrate in general a satisfactory fit to the 1:1 line. However, the percent soil cover between 50 and 85 % were slightly underestimated. This is likely due to the selected soil endmembers, which were probably too bright. The slope and the intercept corroborate these conclusions by expressing a week deviation from the 1:1 line. Selecting a soil endmember with a lower brightness would increase the fractions. The selection of the endmembers, which truly represent the crop residue and soils conditions, is crucial for proper unmixing. In order to minimize errors in selecting endmembers, “pure” patches were considered for residue and the dominant soil class. However, these endmembers were not perfectly and totally representative of the different soil types found in the test site. Indeed, small changes, which could affect the spectral signature of soil, might occur within the site. For example, changes in mineral composition, colour,
3
brightness, moisture, organic matter content and texture are factors, which influence the soil spectrum [25 & 26]. In addition, pre-processing of the Hyperion data is essential and will affect the integrity of the surface reflectance retrieval from at-sensor radiance data. The errors in the image data were reduced significantly by applying different techniques for the removal of sensor artifacts. However, residual errors probably remain and cause small reflectance differences between the measured field endmembers and their corresponding image endmembers.
100
ACKNOWLEDGMENTS The authors would like to thank the National Science and Engineering Research Council (NSERC) and the University of Ottawa’s Faculty of Arts for their financial support. The authors would also like to acknowledge the Canada Centre for Remote Sensing (CCRS), Natural Resources Canada, which provided the hyperspectral Hyperion data. We would like to thank numerous people who were involved in this project and who provided their support and expertise: Lixin Sun (CCRS), Jean-Claude Deguise (Agriculture Canada), Dr. Robert Neville (CCRS), Rob Hitchcock (Prologic Systems) and all the field teams at Indian Head. We would also like to thank the anonymous reviewers for their constructive comments.
Residue Fraction (%)
80
60
40
20
0 0
20
40
60
80
100
Residue Cover From Ground Reference (%)
REFERENCES
Figure 2: Relationship between PCRC (percent crop residue cover) derived from Hyperion data and the ground reference measurements.
100
80 Soil Fraction (%)
that Hyperion performed well using data acquired in the beginning of the agricultural season. This performance of Hyperion is due to its spectral band characteristics, especially the availability of contiguous narrow bands in the SWIR region, which is sensitive to the residue (lignin and cellulose absorption features) and soil. Correlation between ground references measurements and extracted fractions from Hyperion data using constrained linear spectral mixture analysis showed that the model was overall a good predictor for crop residues percent cover (D of 94%, R2 of 0.73 and RMSE of 8.7 %) and soil percent cover (D of 0.91, R2 of 0.70 and RMSE of 10.3 %).
60
40
20
0 0
20
40
60
80
100
Soil Fraction From Ground Reference (%)
Figure 3: Relationship between PSC (percent soil cover) derived from Hyperion data and the ground reference measurements. IV. CONCLUSION The capability of Hyperion hyperspectral data was evaluated for estimating and mapping crop residues and soil fractions (percent cover) on agricultural lands using constrained linear spectral mixture analysis and endmembers acquired in the field with the exception of the vegetation endmember which was retrieved from our spectral library. The overall results indicate
[1] Daughtry, C.S.T., Hunt Jr., E.R., Doraiswamy, P.C. and McMurtrey, J.E. (2005) Remote Sensing the Spatial Distribution of Crop Residues. Agronomy Journal, 97(3): 864-871. [2] Aase, J.K. and Tanaka, D.L. (1991) Reflectance from four wheat residues cover densities as influenced by three soil backgrounds. Agronomy Journal, 83: 753-757. [3] Rice, C.W. (2002) Storing Carbon in Soil : Why and How ? Geotimes, pp. 1-5. [4] Bannari, A., Haboudane, D. and Bonn, F. (2000) Intérêt du moyen infrarouge pour la cartographie des résidus de cultures. Canadian Journal of Remote Sensing, 26(5): 384-393. [5] Bannari, A., Chevrier, M., Deguise, J.-C., McNairn, H. and Staenz, K (2003) Senescent Vegetation Mapping Using Artificial Neutral Networks and Hyperspectral Remote Sensing. IEEE-IGARSS, Toulouse, France, 21-25 July 2003. Proceedings on CD. [6] Daughtry, C.S.T., Doraiswamy, P.C., Hunt Jr., E.R., Stern, J.E., McMurtrey, J.E. and Prueger, J.H. (2006) Remote Sensing of Crop Residues Cover and Soil Tillage Intensity. Soil and Tillage Research, 91: 101-108. [7] Bannari, A., Pacheco, A., Staenz, K., McNairn, H. and Omari, K. (2006) Estimating and Mapping Crop Residue Cover in Agricultural Lands Using Hyperspectral and IKONOS images. Remote Sensing of Environment, 104: 447-459. [8] Beck, R. (2003) EO-1 User Guide, Version 2.3. http://eo1.usgs.gov and http://eo1.gsfc.nasa.gov. 74 pages. [9] Geophysical and Environmental Research Corporation (1990) GER 3700 Spectroradiometer User’s Manual, version 2.1, 55 pages. [10] Labsphere (2001) A Guide to Reflectance Coatings and Materials, North Sutton, New Hampshire, http://www.labsphere.com/tech_info/docs/Coating_&_Material_ Guide.pdf. [11] PCI Geomatics (2003) Using PCI Software, Richmond Hill, Ontario, Canada, 540 pages. [12] Khurshid, K.S., Staenz, K., Sun, L., Neville, R., White, H.P., Bannari, A., Champagne, C.M. and Hitchcock, R. (2006) Preprocessing of EO-1 Hyperion Data. Canadian Journal of Remote Sensing, 32(2): 84-97.
4
[13] Staenz, K., Szeredi, T. and Schwarz, J. (1998) ISDAS A system for processing and analyzing hyperspectral data. Canadian Journal of Remote Sensing, 42(2): 99-113. [14] Sun L., Neville R. A., Staenz, K. and White H.P. (2007) Automatic Destriping of Hyperion Imagery Based on Spectral Moment Matching. Canadian Journal of Remote Sensing (accepted for publication). [15] Neville R. A., Sun L. and Staenz K. (2007) Spectral Calibration of Imaging Spectrometers by Atmospheric Absorption Feature Matching. Canadian Journal of Remote Sensing (accepted for publication). [16] Staenz, K. and Williams, D.J. (1997) Retrieval of Surface Reflectance from Hyperspectral Data Using a Look-Up Table Approach. Canadian Journal of Remote Sensing, 23(4): 354368. [17] Staenz, K., Neville, R.A., Levesque, J., Szeredi, T., Singhroy, V., Borstad, G.A. and Hauff, P. (1999). Evaluation of casi and SFSI hyperspectral data for environmental and geological applications - two case studies. Canadian Journal of Remote Sensing, 25: 311-322. [18] Adams, J.B., Smith, M.O. and Johnson, P.E. (1986) Spectral mixture modeling: A new analysis of rock and soil types at Viking Lander 1. Journal of Geophysical Research, 91: 81138125. [19] Boardman, J.W. (1995) Analysis, understanding and vizualization of hyperspectral data convex sets in N-Space. Proceedings of SPIE’s International Conference on Imaging Spectrometry, Orlando, Florida. Vol. 2480, pp. 14-20.
[20] Myneni, R. B., Maggion, S., Jaquinta, J., Privette, J. L., Gordon, N., Pinty, B., Kimes, D.S., Verstraete, M. M. and Williams, D. L. (1995) Optical remote sensing of vegetation: modeling, caveats, and algorithms. Remote Sensing of Environment, 51 (1): 169–188. [21] Kerdiles, H. and Grondona, M. O. (1995) NOAA-AVHRR NDVI decomposition and subpixel classification using linear mixing in the Argentinean Pampa. International Journal of Remote Sensing, 16 (7): 1303–1325. [22] Schwarz, J. (1998) Classification of hyperspectral data. Masters Thesis, School of Computer Science, Carleton University, Ottawa (Ontario) Canada, 180 pages. [23] Willmott, C.J. (1982). Some comments on the evaluation of model performance. Bulletin American Meteorological Society. 63(11): 1309-1313. [24] Daughtry, C.S.T. (2001) Discriminating Crop Residues from Soil by Shortwave Infrared Reflectance. Agronomy Journal, 93(1): 125-131. [25] Irons, J.R., Weismiller, R.A. and Petersen, G.W. (1989) Soil reflectance. In: Asrar, G. Editor, , 1989. Theory and Applications of Optical Remote Sensing Wiley-Interscience, New York, pp. 66-106. [26] Bannari, A., Huete, A.R., Morin, D. and Zagolski, F. (1996) Effets de la couleur et de la brillance du sol sur les indices de végétation. International Journal of Remote Sensing, 17(10): 1885-1906.
5
