The Unique Uses of IFSAR Data in Difficult Agricultural ProjectsThomas M. Carson, PhD (Presenter) Steve Shaffer Technologies Division Fugro EarthData, Inc. Introduction The National Agricultural Imagery Program (NAIP) of the United States Department of Agriculture (USDA) annually acquires 1 and 2 meter resolution digital orthorectified imagery for agricultural regions in the United States during the summer crop growing season. About twenty percent of the annual coverage is 1 meter resolution, and eighty percent 2 meter resolution. The one meter product is intended to provide an updated orthorectified base procured on a 3-5 year cycle and is accurate to within 5 meters of Digital Ortho Quarter Quads (DOQQ’s). The two meter product is accurate within 10 meters of DOQQ’s and is intended for aerial compliance and other programs not requiring high spatial accuracy. One issue with the NAIP program is the need for continuous cloud-free days in the height of the growing season in order to complete the state-wide imagery collection projects. This is often a problem, particularly in the southeastern Unites States. One way to relieve this issue is to collect NIAP quality imagery with radar sensors, which can image day or night, and through cloud cover. Under USDA contract AG-8447-C-06-0006, Fugro EarthData, Inc. (EarthData) is conducting a pilot project to demonstrate the capabilities of the GeoSAR airborne Interferometric Synthetic Aperture Radar (IFSAR) to augment the USDA National Agricultural Photography Program (NAIP). Under the scope of this contract, the EarthData GeoSAR system collected sufficient X-band and P-band interferometric SAR data using its standard imaging modes to fully map the county of Yazoo, Mississippi. GeoSAR collected a total of 16 lines of data in 5 sorties between August 29, 2007 and August 31, 2007. EarthData processed this data in its Frederick, Maryland facility to produce standard products, including 3-meter resolution X-band elevation models and orthorectified mosaics and 5-meter resolution P-band elevation models and orthorectified mosaics. Additionally, a subset of the X-band data was processed into a 1.25 meter resolution orthophoto mosaic to demonstrate a resolution product similar to NAIP imagery. This paper will compare and contrast features evident in airborne radar data with those in standard NAIP imagery. In particular, the paper will present results of how the unique characteristics of the X- and P-band SAR images can facilitate the collection of certain feature types, such as fence lines, that are not readily visible in standard NAIP imagery. The paper will also present initial results of how the common crop types present in the data sets (corn, soy, wheat, sweet potatoes etc.) are separable within the feature space defined by the GeoSAR X-band image data. Geosar technology overview GeoSAR is a dual-sided, dual-frequency, interferometric synthetic aperture radar (IFSAR) mapping system. The system is integrated onto a Gulfstream II business jet (Figure 1), and is wholly owned by Fugro EarthData Incorporated. It is capable of collecting data from 40,000 ft above ground level at an airspeed of over 400kts yielding a net collection rate of over 280 sq km per minute. The SAR operates at 2 frequencies simultaneously, X-Band, with a center frequency of 9700MHz, and P-Band, with a center frequency of 350MHZ. The system is designed to produce high accuracy digital elevation models (DEM) through the process of radar interferometry, as well as SAR orthophoto mosaics. GeoSAR is a commercialization of IFSAR technology developed by NASA Jet Propulsion Laboratory for wide area airborne mapping applications. EarthData has been commercially operating the GeoSAR aircraft since 2002 generating large cover area maps for the US National Oceanic and Atmospheric Administration in Southern California and for the National Geospatial-Intelligence Agency in South America. Image and DEM quality is excellent and is independent of cloud cover and sun illumination, yielding nearly all weather collection capability.  Figure 1 GeoSAR IFSAR collection aircraft. The P-Band antennas are contained in the tip-tanks on either side of the aircraft. The X-Band antennas are in the fairing visible near the fuselage on the underside of the aircraft. (Photo courtesy Fugro EarthData Inc.) Description of Study area Yazoo County Mississippi is a predominantly rural county in west central Mississippi, with an area of 2,419 km2 (934 mi2). The major crops in the county, in terms of economic value, are corn, cotton, rice, wheat, hay, sorghum grain, soybeans and sweet potatoes. The major livestock commodities are cattle and catfish. The physiography of the county can be divided into two regions. To the west of the loess-capped bluff bisecting the county the landscape is part of the Yazoo-Mississippi Basin or delta. This terrain is level to very gently undulating near the Yazoo River and around abandoned and extinct river channels. The land cover in the delta consists of swampy forests, agricultural fields, and occasional catfish ponds. East of the bluff there are low, wooded Loess Bluffs or Brown Loam Hills. Among the hills nearest the bluff, valleys are often deep and steep-walled. Geosar data acquisition GeoSAR collected a total of 16 lines of data in 5 sorties between August 29, 2007 and August 31, 2007. The data was acquired both day and night, and through scattered cloud cover. The lines were collected to cover the entire county with NS looking and EW looking data. The average line collected was 122 km long. Ground Truth data collection A ground campaign was conducted in conjunction with the radar collection. A team of three EarthData scientists, accompanied by the local USDA extension officer, visited approximately 25 sites in Yazoo County from August 26 to August 29, 2007. At each site, data was collected on individual fields regarding the crop type, local soil roughness, row spacing, plant height, width, spacing, and predominant compass direction of rows. Approximately 220 ground-based photographs were taken on August 27 and August 29, 2007. The photographs assist in the evaluation of field and crop conditions. Soil cores were collected in 5 major soil types for later laboratory analysis for bulk density, soil moisture and electrical conductivity. The team conducted an airborne reconnaissance for 2.5 hours of flight time in a light aircraft on August 26, 2007 for the acquisition of approximately 550 oblique, natural color images from an altitude of 150-200 meters. Two trihedral X-band corner reflectors were placed in the study area, with their locations precisely measured through GPS. The corner reflectors will be used for absolute geometric and radiometric control of the radar data. GeoSAR derived products The basic products derived from GeoSAR are orthorectified backscatter image mosaics derived from the amplitude information from the X-Band and P-Band data, and digital elevation models derived from the phase information through radar interferometry. The image resolution is nominally 3 meters for X-Band and 5-meters for P-Band; however 1.25 meter products have been successfully produced. The digital elevation models have a point spacing of 5 meters. When the radar signal impinges on a resolution element in a scene, the signal is scattered in all directions. The amount of energy that is scattered back towards the receiver is thus the backscatter energy. The backscatter strength is a function of the structure of the surface and its composition. In order to make quantitative analysis of radar imagery data across an image, it is desired that the backscatter energy measured by related solely to the target characteristics. In a radar backscatter image, however, the measured backscatter is also related to the position of the target in near to far range, and the local slope characteristics. For this reason, sigma naught (?0), the scattering coefficient, is estimated from the image data. Sigma naught (?0) is a measure of the average reflectivity of a horizontal sample normalized to a unit cross sectional area and is defined as the amount of backscattered power compared to the power of the incident wave (Henderson and Lewis, 1998). The GeoSAR X-Band radar produces single polarized data (VV, or vertical transmit and vertical receive), while the P-Band radar produces a cross polarized data set (HH and HV, or Horizontal transmit, and both horizontal and vertical receive). For most applications, only the HH P-Band data is processed. However, for the purpose of this study, the HV P-Band data was processed as well. The cross-polarized element of the P-Band radar image provides an additional data layer to support image interpretation. Because the GeoSAR system is a single pass interferometer, an additional data element is available, that of interferometric correlation. The correlation layer is a measure of the phase correlation between the two signals received at each antenna, with values of 0 to 1. A value of 1 represents a perfect correlation between the two received signals. If the antennas were in the exact same position, the received phase information would conceivably be perfectly correlated. Values of less than 1 indicate some level of decorrelation between the received signals. Decorrelation can be related to several factors, the one of interest in this application being due to the vertical distribution of scattering elements within the scene. Because of this, the correlation layer can be used to differentiate between forested regions, regions of crops with varying vertical structure (for example corn versus soy beans), and open areas with no vertical scattering distribution. SAR image interpretability compared to NAIP Figure 2 juxtaposes a NAIP air photo from 2006 with the same area collected in GeoSAR X-Band in 2007. The image shows that most, if not all of the features identifiable in the NAIP image are also present in the GeoSAR image. The forested areas, both in the wood-lots as well as along the stream bed are visible as blue, mottled features in the SAR image. The field boundaries are all clearly visible, including the fence-lines and wind breaks. The distinct patterns from successive river-meander deposits in the lower left quadrant of the images are clearly visible in the SAR as well as the NAIP imagery. The patterns are partially obscured in the SAR imagery because the crops are still present in some of the fields. While the circular field patterns created from center-pivot irrigation practices are not readily apparent in the SAR image, the irrigation structures themselves are easily visible.  Figure 2 NAIP imagery from 2006 (left, courtesy of the USDA Farm Service Agency Aerial Photography Field Office) with the same area of GeoSAR X-Band imagery (right, courtesy Fugro EarthData Inc.) from Yazoo Mississippi. The GeoSAR imagery combines the backscatter and elevation information to produce a colorized image. Crop and landcover classification This section will report on preliminary analysis of the utility of the SAR imagery for performing land-cover classification using standard ERDAS tools. The study reported here is based only on the GeoSAR X-Band sigma naught imagery. The study area had 5 distinct information classes, water, cotton, corn, sweet potatoes and void (due to radar shadow). Image Processing A single 3 square km image was subset from the GeoSAR data. The sigma naught image was processed as described above, and then further processed using a Lee speckle removal filter. The presence of speckle in SAR images degrades the interpretability of cover types. The Lee speckle reduction filter is designed to reduce speckle in radar imagery while simultaneously preserving texture information (Blumberg, 2007). Speckle in imaging radar can be mathematically modeled as multiplicative noise with a mean of 1. For this study, a series of four Lee-sigma filters were used for speckle suppression and image quality improvement (Lee, 1981). The Lee-sigma filter assumes that 95.5% of random samples are within a range of 2 standard deviations. It replaces the pixel of interest with the average of all DN values within the moving window that falls within the designated range. The coefficient of variation was computed for each scene separately and scaled kernel window values and standard deviations were used to run the filter. Table 1 The parameters used in each iteration of Lee filter for speckle suppression.
Imaging radar data noise follows a Rayleigh distribution. This yields a theoretical value for Standard Deviation (SD) of .52 for 1-look radar data. This agrees with the computed values of SD = .56 for the initial pass Table 1). The final SD = .26 is equivalent to 4-look radar data. Image Classification An unsupervised classification was initially performed on the data using the Iterative Self-Organizing Data Analysis Technique (ISODATA) (Tou and Gonzalez, 1974) available in ERDAS Imagine. ISODATA uses spectral distance and iteratively classifies the pixels, redefines the criteria for each class, and classifies again, so that the spectral distance patterns in the data gradually emerge (ERDAS 2007). A seed value of 8 for the number of spectral classes was chosen. The algorithm was set to run for a maximum of 15 iterations or a convergence threshold of 0.975 or greater. Once the initial unsupervised classification was performed, the 8 classes were combined into meaningful land-cover classes based on ground truth data. Figure 3 shows the color mapping of the 5 land cover classes (water, void, sweet potato, corm and cotton). The color mappings conform very closely to the field boundaries in the data, as shown in Figure 4.  Figure 3 Screen-shot from the ERDAS signature editor showing the color values of the 5 land-cover classes in the study area.  Figure 4 The juxtaposed images are split between the color classification map developed through supervised classification and the X-Band SAR image from which the classification was derived. Discussion and further work This initial evaluation of the GeoSAR IFSAR data collected in Yazoo County, Mississippi demonstrates its utility for agricultural analysis. The interpretability of the SAR image data was comparable to the NAIP imagery. The system also proved that it was capable of collecting imagery day and night and through cloud cover. The X-Band image data was shown to be usable in a supervised classification to identify major crop types in the image. Further work on this data set will be done to improve the classification results through the use of other radar features, including correlation data, elevation models, as well as the P-Band imaged data. References
|