1-KM AVHRR GLOBAL LAND DATASET: FIRST STAGES IN IMPLEMENTATION

J. C. Eidenshink, U.S. Geological Survey, EROS Data Center, Sioux Falls, South Dakota 57198, U.S.A.
J. L. Faundeen, Hughes STX Corporation*, EROS Data Center, Sioux Falls, South Dakota 57198, U.S.A.

Abstract. The global land 1-km data set project represents an international effort to acquire, archive, process, and distribute 1-km AVHRR data of the entire global land surface in order to meet the needs of the international science community. A network of 29 high resolution picture transmission (HRPT) stations, along with data recorded by the National Oceanic and Atmospheric Administration (NOAA), has been acquiring daily global land coverage since 1 April, 1992. A data set of over 40,000 AVHRR images has been archived and made available for distribution by the United States Geological Survey, EROS Data Center and the European Space Agency.

Under the guidance of the International Geosphere Biosphere Programme, processing standards for the AVHRR data have been developed for calibration, atmospheric correction, geometric registration, and the production of global 10-day maximum normalized difference vegetation index (NDVI) composites. The major uses of the composites are related to the study of surface vegetation cover. A prototype 10-day composite was produced for the period of June 21-30, 1992. Production of a 30 month time series of 10-day composites is underway.

* Work performed under U.S. Geological Survey contract 1434-92-C-40004

1.0 Introduction

1.1 Scientific Requirements

Scientific investigations indicate that global change information can be derived from the 1-kilometer (km) advanced very high resolution radiometer (AVHRR) data acquired by the National Oceanic and Atmospheric Administration's (NOAA) Television Infrared Observation Satellite (TIROS) (Townshend et al., 1994). Over the past 3 years various scientific organizations have identified the need for compiling a global 1- km resolution multi-temporal AVHRR data set.

The International Geosphere Biosphere Programme-Data and Information System (IGBP-DIS) land cover working group completed a study that stressed the need for a this data set (Townshend, 1992). The United Nations Food and Agriculture Organization's Forest Resources Assessment 1990 Project requires 1-km AVHRR data for all the forested lands on the planet, with emphasis on the tropical zones, for their global forest inventory mandate (FAO, 1990). NOAA researchers required a 1-km AVHRR data set of the northern hemisphere to study the urban heat island effects on surface observations of temperature data (Gallo et al., 1993). The Commission of the European Communities (CEC) and the European Space Agency (ESA) have a joint requirement for global, near daily, long-term, consistent optical and thermal satellite data for tropical environments in support of the joint Tropical Ecosystem Environment Observations by Satellite (1991) project.

The NASA Moderate Resolution Imaging Spectrometer land science team concluded that a global land 1-km AVHRR data set is crucial to develop algorithms for several land products for the Earth Observing System (EOS) (Running et al., 1993). The compilation of such a data set has received endorsement from the Committee on Earth Observations Satellites (CEOS) Plenary, through the recommendation of the CEOS Working Group on Data (CEOS-WGD). CEOS Plenary endorsement facilitates international cooperation ensuring participation and completion of the data set. NASA asked the U.S. Geological Survey (USGS), Earth Resources Observation Systems (EROS) Data Center (EDC) to coordinate the gathering and management of such a data set as a part of its role as the EOS Land Processes Distributed Active Archive Center (LPDAAC).

1.2 Feasibility of a Global Data Set

Analysis of existing 1-km AVHRR data collection capabilities of NOAA, the USGS, the ESA, Australia's Commonwealth Scientific and Industrial Research Organization (CSIRO), China's Satellite Meteorological Center (SMC), and a number of other foreign AVHRR receiving stations indicated the feasibility of collecting a global land 1-km AVHRR data set over all land surfaces. Each agency indicated a desire to participate in the project.

Beginning in May 1991, NASA, NOAA, ESA and the USGS met to identify the objectives of this project. The project is designed as a two-phase effort. The objectives of phase I are data acquisition, archiving, and distribution of the raw data. Phase II is product generation, and involves soliciting and applying internationally accepted processing algorithms to the data set to produce prototype and operational higher level products for global change research. Subsequent meetings held that year included members from IGBP and CEC and led to defining responsibilities and scientific endorsements for the Project.

In February 1992, NASA, NOAA, IGBP, and the USGS hosted the first Global Land 1-km AVHRR High Resolution Picture Transmission (HRPT) Ground Station Operators Meeting in Pasadena, California. Agreements were negotiated with the various organizations to ensure a smooth project implementation. On April 1, 1992, the project officially began with 23 stations worldwide plus the NOAA local area coverage (LAC) recorders capturing, copying, and transferring the 1-km data to the EDC.

2.0 Phase 1 - Data Acquisition, Archive, and Distribution

2.1 Global Land 1-km AVHRR Data Set Definition

The data set is composed of 5-channel, 10-bit, raw AVHRR data, at 1.1-km resolution (at nadir) for every daily afternoon pass over all land and coastal zones using data from NOAA's polar-orbiting TIROS. Initially the data were to be collected continuously for 18 consecutive months beginning April 1, 1992, and continuing through September 30, 1993, subsequently the period has been extended to September 30, 1996.

2.2 Data Acquisition Network

The six major participants (NOAA, ESA, USGS, NASA, CSIRO, and SMC) coordinate HRPT receiving stations to acquire and compile the data set. NOAA coordinates two HRPT receiving stations, Gilmore Creek, Alaska and Wallops Island, Virginia, and schedules the satellite's recorders to obtain local area coverage (LAC) AVHRR 1-km data over areas not covered by the other HRPT receiving stations.

Table 1 lists the participating HRPT receiving stations and their geographic locations. Figure 1 (212,779 bytes) illustrates the approximate ground coverage areas for the HRPT receiving stations. Those areas not covered by HRPT stations, or those areas where stations are not able to collect data throughout the 30-month period, are the responsibility of NOAA to utilize the onboard tape recorders to acquire LAC data. The one exception is Antarctica where complete land coverage is not always available. Since the major uses of the data sets are related to surface vegetation cover, this is currently not regarded as a major limitation.

Table 1.

Each of the major participants and their affiliates are committed to gather, manage, and share the data. Their responsibilities include routine collection of all NOAA-11 observations within their coverage region, quality control of the raw data, providing the data in a standard format, and delivery of the data on a timely basis. These responsibilities are intended to complement rather than interfere with normal station operations and commitments to other programs.

2.3 Data Acquisition and Archive Management

Data acquisition and archive management of the HRPT data received in real time at the EDC, the NOAA LAC data received via the DOMSAT receiving systems, and the data acquired by the ground station network are conducted on a central computer system at the EDC. The system is capable of ingesting, archiving, creating metadata, and creating digital quicklooks for 100 images per day. The system is equipped with a wide variety of magnetic tape devices, a communication network to facilitate data ingest, and a sufficient processing capability for large volumes of data.

Acquisition of HRPT and LAC data is a standardized and routine process. However, nearly all of the ground receiving stations use different data and media formats. The EDC receives the large volume of data on a wide variety of media including 4- and 8-mm, 3480 cartridges, and 9-track tapes. Therefore, the EDC wrote ingest procedures specific to each station's format from sample AVHRR HRPT data obtained from each station. The ingest routines create a Level 0 archive format image file along with a CEOS Inventory Exchange Format (IEF) header file that has a three-letter code specific to each receiving station. The IEF header file was based upon the Release Draft 0.4 from the CEOS Working Group on Data, Format Subgroup from November 2, 1992. The three-letter codes were adopted from the CEOS Working Group on Data, Catalog Subgroup which allows the data to be more universally accessible.

All data provided by the ground station network are logged in a data tracking system. Information such as the CEOS code of the contributor station, the number of images per media, and the media type are recorded. After the data are ingested the original contributor tape media are placed in an environmentally controlled archive.

The ingest routine creates a Level 0 archive copy of the data, metadata, quicklook microimage fiche, and digital quicklook images. The Level 0 archive copy is stored on 3480 tape cartridges for permanent archive. The metadata are checked for coordinate accuracy, satellite direction, day versus night determination, and acquisition date and time. Channel two (near-infrared) is used for digital quicklook images for daytime scenes while channel four (thermal) is utilized for night scenes. The digital quicklook image is scaled to 8-bits and subsampled every fourth line and fifth sample. Compression routines of the Joint Photographic Experts Group are implemented to realize between a 10 and 13 times data reduction using lossy compression (Nelson 1991). The size of a typical browse file is between 20,000 and 40,000 bytes after compression. The metadata and browse files are transferred to an Information Management System (IMS) where data base management routines are used to check for data quality, data gaps and inconsistencies in the metadata.

2.4 Orbital Pass Generation

Nearly 45,000 individual observations were acquired during the first 30 months of this project. The number of individual scenes and overall data volume, an estimated 3.2 terabytes, are large by contemporary standards. As a result a new process was developed to improve data management.

Figure 2 (247,035 bytes) illustrates how individual AVHRR observations from the ground receiving network and the NOAA tape recorded data are collected along a group of orbits in a single day (24 June 1992). Note the overlapping coverage along each orbit.

Orbital stitching is the process of combining a group of consecutive AVHRR observations to form an orbital pass. The global land 1-km AVHRR data set consists of only the afternoon (ascending) passes, and no descending (night time) data, from the NOAA polar orbiting satellites. As a result only half of the data from an individual orbit is needed. When the afternoon observations are stitched, a pole to pole half-orbital pass is generated. Combining the observations reduces the data volume by eliminating redundant data from the overlap areas between receiving stations, improves data quality by removing dropped and bad scan lines, and facilitates data distribution and product generation by reducing the overall number of units of data that must be handled.

Data quality of the individual observations was a major concern. Examination of the data received from the ground stations and the tape recorded LAC data identified several data quality problems. The most common problems were dropped lines, anomalous line and pixel noise and repeated lines. Dropped lines are generally replaced with the zero value automatically during the acquisition process. These were easy to locate and present very little problem in subsequent processing. The other problems were much more serious.

Anomalous lines and pixels were very difficult to detect because they contain non-zero values. The anomalous lines were generally common to all channels, whereas single pixel noise was not. The anomalous values in the lines and pixels cover the complete data range. However, the anomalous lines and pixels generally occurred at the beginning or end of an image and were assumed to be related to problems associated with acquisition and loss of signal due to low antenna angles at the horizon. Repeated lines were not very common and were also assumed to be associated with acquisition and loss of signal due to low antenna angles at the horizon.

Detection of the bad data was necessary in order to avoid serious problems in subsequent processing. The problems were very evident following the production of the first 10-day normalized difference vegetation index (NDVI) composite. The anomalous lines and pixels produced either very high or low NDVI values. The problems with the low NDVI values were apparently corrected by the maximum NDVI compositing process because pixels from other observations with higher NDVI values were selected for the composite. However, the problems that produced high NDVI values were carried throughout the process and remained in the composite. They were most evident in areas with low NDVI. The lines showed up as linear features and the pixels showed up as speckles in the data. The data in the composite became useless in areas where the bad data occur on a daily basis.

Once the bad data were present in the composite, there was no effective method for removing it. Therefore, it was necessary to detect them in the raw data. Two bad data detection algorithms were developed; one for lines and the other for pixels. The algorithms were used on the raw data prior to stitching. Bad lines and pixels were detected using a series of tests of different combinations of channels. If a bad line or pixel was detected it was set to the zero value. Zero values were handled in such a way that there were no detrimental affects on subsequent processing. The effect of the bad data detection algorithms was monitored and the algorithms were adjusted to optimize bad data detection and minimize false detections.

The stitching process began with the southernmost observation (the observation in the group with the earliest start time). Table 2 lists the start and stop times of four orbital segments to be stitched. The start time of the next observation was used to determine if an overlap area or gap existed between observations. When the end time of the previous observation in the half-orbit pass was later than the start time of the next observation, an overlap area existed. Instead of automatically copying each record from the half-orbit pass that overlaps with the next observation, each scan line was read and checked to determine if the line contained valid data.

Table 2. Acquisition start and stop times for observations in four AVHRR orbital segments.

The most frequent form of invalid data was a dropped line. A dropped line, a record containing all zeros, was identified by a gap between the consecutive time stamps. Zeros were added to the pass by the acquisition systems to maintain the proper along-track perspective.

When a zero value (a line or pixel that was detected as missing or bad data) was encountered in the first observation, the coincident record from the overlap area was read to determine if it was valid data. If this record was valid, it was used to replace the zero values in the half-orbit pass. If a replacement line was not found, the zero value was left and processing continues. After the overlap area was completed the remainder of the observation was copied to the half-orbit pass. This process was repeated until all the observations had been added to the half-orbital pass.

If a gap between observations existed when the start time of the next pass was later than the end time of the orbital pass, the gap was zero filled in the half-orbit pass to maintain the proper along-track perspective. To conserve data storage, missing data at the beginning or end of the orbital pass, such as missed passes or ocean data, were not zero filled. Thus each half-orbital pass did not stretch from pole to pole.

Figure 3. Example of two stitched orbital passes over Africa and western Europe. (252,349 bytes)

The orbital stitching process will reduce the number of archive elements from approximately 45,000 images to 11,700 orbits and reduces data volume from 3.2 terabytes to approximately 2.0 terabytes.

2.5 Acquisition and Archive Status

From 1 April, 1992, through 13 September, 1994, the EDC received, archived, and created metadata and browse for over 45,000 scenes. Orbital segments will be produced chronologically beginning with data from April 1, 1992.

2.6 Data Distribution

The main objectives of phase I are the acquisition, archiving, and distribution of a global land 1-km AVHRR data set. To facilitate the data distribution, a complete copy of the data set will be held by NOAA, ESA and the EDC. The EDC provides ESA a copy of all the data provided by NOAA and the EDC HRPT station and will provide NOAA with a copy of all data received by ESA and the EDC HRPT station.

Each agency is committed to facilitate access to the data by traditional methods and networking. Information Management Systems (IMS) have or are being established to provide access to the data. The IMS at the EDC for these metadata records and digital quicklooks is called the Global Land Information System (GLIS) (Oleson et al., 1991) that allows users easy and flexible access to the AVHRR data base. Users can conduct queries, graphically view the coordinate metadata, observe the digital quicklooks and order data. Information about the IMS's can be obtained from the respective agencies.

3.0 Prototype Product Generation

The science requirements for the global land 1-km AVHRR data set are documented by IGBP, CEOS, NASA, CEC and other international research organizations. Among the requirements is the need for higher level products that are derived from the raw 1-km AVHRR data, such as vegetation indices and periodic temporal composites.

The raw data are an important product simply because many AVHRR data at 1-km were either unavailable, or at least very difficult to acquire from foreign receiving stations; these are now readily available from the archive of the EDC, ESA, and NOAA. The raw data are also most desirable for development of calibration, atmospheric correction, and other algorithms requiring the basic raw data. The raw data will initially be available on a scene basis as it is acquired by the ground receiving stations and eventually as part of an orbital segment. The use of the orbital segments will improve along track scene- framing so the user will be able to use a single continuous scene instead of having to mosaic multiple scenes to obtain coverage of a large study area.

The periodic temporal composites must include the minimum of the ten bands listed in Table 3. The composites must be generated from radiometrically calibrated, atmospherically corrected, and geometrically registered data. As a result, the typical user, who is mainly interested in the data for assessment of vegetation condition, derivation of biophysical parameters, mapping of land cover or monitoring land surface characteristics, is freed of the burden of preprocessing the raw data. The assumption here is that raw data have been processed using widely accepted, well defined and documented processing standards.

Table 3. Band description of composite images.

The definition of the processing standards for the products has been reached by consensus as a result of the efforts of the IGBP-DIS land cover working group (Townshend et al., 1994), CEOS, AVHRR Pathfinder land science working group (James and Kalluri, 1994). The standards for the primary processing steps of radiometric calibration, atmospheric correction, geometric registration, and compositing have been documented and guidelines for other functions such as map projection, compositing period, and product format were provided. Overall, the recommendations of the IGBP were the most comprehensive and are being used to create the prototype products.

In a project of this magnitude, data processing efficiency is very important. Each day, the amount of data to be processed increases incrementally. The goal for any processing scenario is the ability to process a days worth of data in at least a day. Existing operational programs, such as the greenness mapping program at the EDC meet the requirement for regional and subcontinental data sets of the conterminous U.S., Eurasia, and Alaska. The AVHRR Pathfinder Program has developed the capability to process daily 4-km Global Area Coverage (GAC) data (James and Kalluri, 1994). Therefore, new processing technology was developed and existing technologies were refined to meet the requirement of processing the data in a reasonable time period.

3.1 Processing Methods

The implementation of the data processing flow is generally a stepwise process that incrementally applies higher order processing in a logical and efficient manner. The steps identified are:

1. radiometric calibration
2. atmospheric correction
3. computation of NDVI
4. geometric registration
5. compositing

The following is a discussion of the methods and parameters that have been recommended for each step and a review of the current status and plans for implemenation of the processing algorithms.

3.2 Radiometric Calibration

Radiometric calibration of the AVHRR visible and near-infrared channels (channels 1 and 2) is difficult because there is not always reliable preflight calibration, no onboard calibration, and difficulty with inflight calibration. Preflight calibration coefficients can change while the instrument is in storage or after launch because of the space environment. Degradation of AVHRR sensors after launch is well documented (e.g., Rao, 1987; Price, 1987; Holben et al., 1990). These studies have used a variety of approaches such as ground-based measurements from stable sites such as homogeneous desert targets to monitor the degradation of the sensors. Although the results from the different approaches often are not in close agreement, there is an increasing consensus on suitable coefficients. The calibration method that is recommended for this data set accounts for sensor degradation by using coefficients developed by Teillet and Holben (1994). Their calculation uses a desert calibration to develop time-dependent calibration coefficients for the AVHRR sensor on NOAA-11. Use of calibration coefficients involves extrapolation of the most recent calibration results for processing data on a near real-time basis. Therefore, the time-dependent coefficients are based on a piecewise linear fit of the desert results. A piecewise linear fit is recommended for operational use because, unlike polynomial fits, they will not change retroactively when new data are added to the end of the time series.

The equations for radiometric calibration to radiance and reflectance for the visible and near infrared channels are described by Teillet and Holben (1994). The calibration coefficients for AVHRR thermal channels 3, 4, and 5 are derived onboard the satellite using a view of a stable blackbody and deep space as a reference. The calibration process converts raw digital counts to radiance as described by Kidwell (1991). The radiance values for all channels are stored with 10-bit precision.

3.3 Atmospheric Correction

The impact of atmospheric effects on the AVHRR channel 1 and 2 data and NDVI can be significant. Four principle atmospheric factors, water vapor, aerosols, ozone, and Rayleigh scattering, are considered to have the most impact.

The corrections for ozone and Rayleigh scattering are straightforward (Teillet, 1991). Appropriate Rayleigh scattering correction must include an adjustment for topography. Recommended reference values for Rayleigh optical depths for standard pressure and temperature conditions are available (Teillet, 1990). The local elevation adjustment can be derived using ETOPO5 which is the best available global digital terrain data at this time and for this task has sufficient accuracy.

The correction for ozone should be based on actual measurements derived from the Total Ozone Mapping Spectrometer (TOMS) or other appropriate sensors. However, access and utilization of these data can be difficult. Using the concentration values from standard climatic tables with latitudinal and seasonal dependence is acceptable and is the approach that has been implemented for this data set.

Proper use of the radiative transfer code is important. Bandpass calculations of 0.005-micrometer spacing or better are recommended. No specific radiative transfer code is recommended but it should be noted that results from different code using large optical depths for aerosols and off-nadir viewing angles greater than 60 degrees can vary significantly.

Several approaches for correction of water vapor exist but there is no community agreement on a feasible method. The basic problem is determination of the spatial and temporal variability of water vapor concentrations. The same circumstances affect aerosol corrections. Therefore no water vapor or aerosol correction will be applied.

The input to the atmospheric correction process is radiance values from the calibrated visible and near-infrared channels. The output of the atmospheric correction process is surface reflectance (in percent) of the visible and near-infrared, albeit without corrections for water vapor and aerosols.

3.4 Computation of NDVI

The NDVI is the difference of near-infrared (NIR, channel 2) and visible (VIS, channel 1) reflectance values normalized over the sum of channels 1 and 2 ((NIR-VIS)/(NIR+VIS)). The NDVI equation produces values in the range of -1.0 to 1.0, where increasing positive values indicate increasing green vegetation and negative values indicate nonvegetated surface features such as water, barren, ice, and snow or clouds. To obtain the most precision, the NDVI is derived from calibrated channels 1 and 2 data in 16-bit precision, prior to geometric registration and resampling.

It is recommended to scale the computed NDVI results to 8- bit range to minimize the volume and optimize analysis and display. This is acceptable because there is very little quantitative understanding of the relationship between excessive mathematical precision in the NDVI and physical measurements of the vegetation condition. If greater precision is required, the NDVI can be computed from the calibrated channel 1 and 2 values that are stored with 10-bit precision.

The scaling method is chosen to emphasize different types of vegetation or vegetation condition. The EDC routinely uses a formula that scales the NDVI from -1.0 to 1.0 as 0 to 200, where each value represents 1.0 percent of the total possible range.

3.5 Geometric Registration

Geometric registration involves precise transformation of the image from the sensor-based projection to an earth surface- based projection. This process includes calculating a satellite model, matching ground and image-based control points, and transformation and resampling the data to a map projection coordinate system. The satellite model is also used to compute satellite zenith, solar zenith, and relative azimuth viewing angles for each pixel.

Based on a comprehensive evaluation of map projections for global data sets (Steinwand et al., in press), the Interrupted Goode Homolosine is recommended for the data set. The Interrupted Goode Homolosine (Goode, 1925; Steinwand, 1994 ) has two important features. First, it is an equal area projection that facilitates spatial analysis. Second, it essentially divides the world into 12 regions that can be mosaicked into a global map. The regionalization of the global map has advantages for data handling.

The most important factor to consider in geometric registration is the positional accuracy of each pixel as it is moved from the sensor-based projection to the surface-based projection. The recommendation for this project is a positional accuracy of 1000 meters or less. A systematic correction using the satellite model alone cannot achieve this positional accuracy. Thus, it is necessary to perform control point matching and correction for terrain elevation. If terrain is not accounted for in the satellite's geometric model, registration errors of up to 12 kilometers can occur for extremely off nadir pixels in areas of high relief. Incorporating heights above the reference ellipsoid derived from digital elevation models into the satellite's geometric model can greatly reduce these registration errors. The digital elevation data used for this process is ETOPO5.

Most methods for control point matching use automatic correlation of image segments with ground control points and then integrate the adjustments derived from the correlation with the orbital model (Cracknell and Paithoonwattanakij 1989; Kelly and Hood 1991; Brunel and Marsouin 1987). The problem with this approach is the identification of enough control points on a global basis to ensure the required registration accuracy. The problem is complicated by the routine presence of clouds that prevents a successful correlation process. The problem is minimized by having an extremely dense and evenly distributed set of control points.

Control points for AVHRR data are commonly identifiable features along coast lines, lakes, and rivers. Some prominent physiographic features are used in arid regions where hydrologic features are sparse. One approach is to use satellite imagery to develop the control points. Control points are chosen from a near-nadir, cloud-free base image of 1-km AVHRR data. Another possible approach is to select the control points from Landsat Multispectral Scanner (MSS) data and then degrade the image to AVHRR resolution. Although the use of MSS data improved the precision of the control point marking process it is time consuming and expensive to develop.

Yet another approach is to use hydrologic features from vector data sets such as the Digital Chart of the World (DCW) and the World Vector Shoreline (WVS). The vector data representing coastlines, shorelines, and other major hydrologic features are rasterized to AVHRR resolution and warped into satellite projection. A comparable edge feature is derived from the AVHRR data at approximately the same location using edge extraction filtering methods. The correlation is performed on several control points and adjustments are made. The advantage of this approach is that DCW and WVS are available worldwide. The accuracy of the DCW, however, is variable, depending on the geographic location and original map source. The WVS is considered more accurate for shorelines than DCW, but DCW is the best source of inland water bodies and rivers.

The prototype procedure that was developed using the DCW had an accuracy of 0.8 rms to 1.3 rms pixels based on image to map verification and, therefore, meets the recommended standards in most cases. The accuracy of the registration was improved slightly by constraining the correlation process. However, the accuracy was still slightly greater than 1.0 rms pixels. The most limiting factor is still the overall accuracy of the DCW and WVS. However, the important point was to achieve multi-temporal (image to image) registration rather than absolute positional accuracy in order to prevent "blurring" in the compositing process. Verification of image to image registration had an accuracy of less than 1 pixel rms.

3.6 Compositing

The first factor to consider in the compositing process is the length of the compositing period. Compositing periods of 7, 10, and 14 days have been used most commonly. The choice of the period is usually based on the length of time necessary to obtain a composite with minimal cloud contamination and or the amount of time necessary to observe meaningful changes in surface characteristics. The compositing period that is recommended for the prototype products is approximately 10 days created by month. Thus, January has three composites of 10, 10, and 11 days; February has 10, 10, and 9 or 8 depending on whether it is a leap year, and so on. This procedure has the advantage of creating calendar month composites, which is a common reporting period for agronomic and biophysical characteristics.

The recommended method is maximum NDVI compositing. The NDVI is examined pixel by pixel for each observation during the compositing period to determine the maximum value. The retention of the highest NDVI value reduces the number of cloud-contaminated pixels and selects the pixels nearest to nadir (Holben, 1986).

Other compositing techniques, including those using multiple criteria such as maximum NDVI and maximum apparent temperature, are currently being investigated. The results of these studies may provide a basis for a different compositing technique.

3.7 Processing Flow

The order of the processing steps can vary, depending on the desired types of products. The processing flow for the global land 1-km AVHRR 10-day composites was designed to take advantage of computer resources and provide optimum product flexibility.

The first step was to read in the portion of an orbital segment corresponding to the region of the Interrupted Goode Homolosine projection that was being processed.

The calibration of the five AVHRR channels to radiance was completed next. The date of the observation was determined and the appropriate time adjusted coefficients for channel 1 and 2.

The next step was to perform the control point matching and terrain correction process and develop the adjustments to the orbital model. The control points and the terrain data were warped into the satellite projection for the correlation process. A transformation grid was computed but not applied in this step.

The satellite model was used to compute the three solar/satellite viewing angles (Documentation updated: 11/98). The angles are eventually used in the atmospheric correction process.

Next, the geometric registration was completed for the five AVHRR spectral channels and the three viewing angles. The transformation grid developed in the control point correlation process was applied. The grid postings were at 10-km intervals to be consistent with the ETOPO5 digital elevation data. Intervening pixel locations were interpolated.

The next step was to compute the NDVI and perform the maximum NDVI compositing. The output of the compositing process was defined as a 10-band image that includes the NDVI value for each pixel selected by the maximum value compositing for the 10- day period, the radiometrically calibrated channel 1-5 values, the satellite viewing geometry data, and a date index value. Table 3 lists the data included in each of the 10 bands. The date index band is provided to allow a user to identify the specific date and scene identification of the observation for each pixel in the composite period. Since there are multiple observations of the same ground location each day, especially in higher latitude regions, an ancillary table of the specific date and scene information was provided. The viewing geometry information was provided for several reasons including the identification of pixels that would be considered off nadir or for use in user defined atmospheric correction algorithms. When solar zenith angles are greater than 80 degrees, it is essentially near dark and those data are not included in the compositing process. Therefore, in the winter months there is likely to be no data in high latitude regions of a 10-day composite. The steps described so far were repeated on all orbital segments for the 10-day period.

The final step was atmospheric correction. Typically, atmospheric correction is performed in conjunction with calibration. However, in a recent study (Cihlar and Huang, personal communication) it was shown that using atmospherically corrected data in the maximum NDVI compositing process increased the probability of selecting pixels with higher satellite zenith angles, and preferably in the backscatter direction. Based on these findings and one other important factor, the atmospheric correction will be applied following the compositing process.

The rationale for applying atmospheric correction after compositing is feasible because the maximum NDVI value selection process yields the same results whether the NDVI is computed from raw digital numbers or calibrated radiance. The same pixel is selected from the array of choices. In this process, the NDVI used in the compositing process were computed from the calibrated radiance values that were subsequently be used in the atmospheric correction process. After the compositing process was completed, the channel 1 and 2 radiance values used in the computation of the NDVI that were selected as the maximum value were atmospherically corrected and converted to surface reflectance. The reflectance values were used to recompute the NDVI for the composite.

Perhaps the most significant advantage of this approach is that it provides an opportunity to apply improved atmospheric corrections to the data in the future. For example, there is currently no consensus method for water vapor or aerosol corrections on a global basis. However, ongoing research and validation is likely to provide a suitable method at some point in the future. When this occurs, it will be possible to reprocess the archived radiance values and recompute a better atmospherically corrected NDVI. The computation will be completed without having to use extensive resources to calibrate, geometrically register, and composite the raw data.

The basic assumption is that the original maximum value compositing based on the uncorrected NDVI is acceptable. If the improved atmospheric correction procedures would eliminate the initial concern regarding selection of off nadir views or a new compositing process is developed, the argument for saving processing resources is mute. However, this process still offers the most flexibility given the current state of the art.

3.8 Data Scaling Characteristics

The Global Land 1 km AVHRR Data Set products are processed to maintain the maximum precision of the data. The AVHRR channels 1-5 are scaled to 10-bit precision within a 16-bit (signed) integer data type. The NDVI, viewing geometry, and data pointer are stored as byte data. The following describes the data range, scale, offset, and methods for unscaling the data.

Click here to obtain the data used in production flow as of 11/98 . Note that these data supercede the information below, which remain for legacy purposes only.

1. Equations for scaling and unscaling values.  

   scaled = (actual + shift) * scale + offset

   actual = (scaled - offset) / scale - shift

2. Values to be scaled are stored in signed data types, except unsigned byte.
3. Each value to be scaled must be greater than or equal to 0. This allows the values 0 to 9 to be reserved for special masks (water, interrupted area, no data, etc). To accomplish this, a shift value is added to each pixel prior to scaling.
4. Care must be taken when unscaling the data. The mask values must be reset to some value not present in the data range or they must be reset to a fixed value. For example, if all mask values are set to zero, they will be indistinguishable from a data value of zero.
5. Satellite zenith angles are truncated to the range -90 to 90.

3.9 The First Prototype Products

The challenge of processing a global land 1-km AVHRR data set is great. One of the areas of concentration is the development of the 10-day composite product. A benchmark 10-day composite period was identified to test development of the processing system. The benchmark period chosen was June 21-30, 1992. This is the summer/winter solstice and represents a period with extreme sun angles and spatial variation in global vegetation condition.

The first test was to create a single day 1-km global image. The date, June 24, 1992, was selected from the benchmark period. The objective of the test was to determine the feasibility of registering the half-orbit passes into the Interrupted Goode Homolosine projection using a systematic satellite model correction (i.e., without ground control points).

Using the Goode's Interrupted Homolosine at 1-km resolution, the global image is 17,347 lines by 40,031 samples (694 megabytes per band) for an 8-bit image. The uncalibrated VIS and NIR data of the 14 orbits were processed. The individual orbits were overlapped left to right to create the final global image. The band combination 2, 1, 1 as red, green, blue was used in creating a false color image shown in Figure 4 (71,667 bytes).

The second test was to create a global 1-km 10-day composite for the benchmark period of June 21-30, 1992. The processing for the composite includes calibration, geometric registration, computation of NDVI, and compositing. No atmospheric correction was performed on the data. The calibration of channels 1 and 2 adheres to the recommended procedure that accounts for sensor degradation over time. The thermal channels were corrected using the standard procedure as described by NOAA (Kidwell, 1991). The NDVI values were computed from calibrated reflectance maintained in 16-bit precision. Afterwards, all five channels and the NDVI were converted to byte data to reduce the data volume using standard EDC procedures (Eidenshink, 1992). The geometric registration was performed using DCW for ground control points, and the data were transformed to the Interrupted Goode Homolosine projection. The data were composited using maximum NDVI compositing and a 10-band image, as described earlier, was created.

The input to the processing was the stitched orbital segments. There were 386 scenes stitched into 130 orbital segments for the 10-day period. The raw data volume for the 386 scenes was approximately 21 gigabytes. The problems associated with processing extremely large data volumes were minimized by splitting the world into the 12 regions defined by the Interrupted Goode Homolosine projection. The portion of each orbital segment that intersects a region was extracted and processed. As multiple orbital segments were processed for a region, they were composited. After the processing for all the regions was complete they were mosaicked to form the global image, and a land/sea mask was applied to the data. The 10-day composite is represented as a greenness image ( Figure 5 19,755 bytes).

4.0 Conclusion

The existing 1-km AVHRR data collection capabilities of NOAA, the USGS, ESA, CSIRO, SMC, and a number of other foreign AVHRR receiving stations have been organized to collect a global land 1-km AVHRR data set over all land surfaces. These agencies were committed to gather, manage and share these data with the global change research community and to share the responsibility for the stewardship and preservation of the data. This unprecedented accomplishment of already acquiring 30 months of global land 1-km AVHRR data is testimony to the commitment of these agencies.

The project is committed to using the appropriate international standards to facilitate development of consistent data and information products. The active involvement of the international research and user community is crucial to the advancement of the processing and application of these data.

The data set is composed of 5-channel, 10-bit, raw AVHRR data, at 1.1-km resolution (at nadir) for each daily orbital pass over all land and coastal zones using NOAA's TIROS afternoon polar-orbiting satellites. The data are to be collected continuously beginning 1 April, 1992, and continuing through 30 September, 1996. The raw data are permanently archived, catalogued, and available to the science community.

The science requirements for the higher level products that are derived from the raw 1-km AVHRR data, such as vegetation indices and periodic temporal composites, have been well defined by IGBP, CEOS, NASA, CEC, and other international research organizations. The raw data will be processed to standards reached by consensus of several research scientists and science working groups.

The project will continue to solicit active involvement of the national and international data acquisition, research and user communities to participate in gathering, managing, and using the global land 1-km AVHRR data set.

5.0 Acknowledgements

The EROS Data Center wishes to acknowledge the support of Martha Maiden (NASA), Arthur Booth and Chuck Liddick (NOAA), Giancarlo Pittella, Luigi Fusco, and Alessandra Buongiorno (ESA), Jeff Kingwell and Murray Wilson (CSIRO), and the other international participants in the acquisition of the data set.

We also wish to acknowledge John Townshend and the contribution of the International Geosphere Biosphere Programme and the Committee on Earth Observations Satellites in the definition of the science requirements and processing standards for the data.

6.0 References

• Brunel, P., and Marsouin, A., 1987, An operational method using ARGOS orbital elements for navigation of AVHRR imagery. The International Journal of Remote Sensing, 8, 569-578.
• Cihlar, J. and Huang, F., 1994, Effect of atmospheric correction and viewing angle restriction on AVHRR data composites. Canadian Journal of Remote Sensing,20(2), 132-137.
• Cracknell, A. P. and Paithoonwattanakij, K., Pixel and subpixel accuracy in geometrical correction of AVHRR imagery. International Journal of Remote Sensing, 10, 661-667.
• Eidenshink, J. C., 1992, The 1990 conterminous U.S. AVHRR data set. Photogrammetric Engineering and Remote Sensing, 58(6), 809-813.
• Faizoun, A. C. (editor), Final report of the OSS working group on NOAA-AVHRR standard processing definition. (CNES, Paris: l'Observatoire Sahara Sahel) September, 1992.
• FOOD AND AGRICULTURE ORGANISATION OF THE UNITED NATIONS, 1990, Expert Consultation on Forest Resources Assessment 1990 Project, FAO Headquarters (Rome: Forest Resources Assessment).
• Gallo, K. P., McNab, A. L., Karl, T. R., Brown, J. F., Hood, J. J., and Tarpley, J. D., 1993, The use of vegetation index for assessment of the urban heat island effect. International Journal of Remote Sensing, 14, 2223-2230.
• Goode, J. P., 1925, The Homolosine projection: a new device for portraying the Earth's surface entire. Association of American Geographers, Annals, 15, 119-125.
• Holben, B. N., 1986, Characteristics of maximum-value composite images from temporal AVHRR data: The International Journal of Remote Sensing, 7, 1417-1434.
• Holben, B. N., Kaufman, Y. J., and Kendall, J. D., 1990, NOAA-11 AVHRR visible and near-IR inflight calibration: The International Journal of Remote Sensing, 11, 1511-1519.
• James, M. E., and Kalluri, S., 1993, The Pathfinder AVHRR land data set: An improved coarse resolution data set for terrestrial monitoring: The International Journal of Remote Sensing, 15, 3347-3364.
• Kelly, G. G., and Hood, J. J., 1991, AVHRR Conterminous United States Reference Data Set, Technical Papers, 1991 ACSM-ASPRS Annual Convention, 3, 232-239.
• Kidwell, K. B., 1991, NOAA Polar Orbiter Data Users' Guide, National Oceanic and Atmospheric Administration, World Weather Building, Room 100, Washington, D.C.
• Malingreau, J. -P., and Belward, A. S., 1994, Recent activities in the European Community for the creation and analysis of global AVHRR data sets. International Journal of Remote Sensing, 15, 3397-3416.
• Nelson, M., 1991, The Data Compression Book (Redwood City, California: M&T Books).
• Oleson, L. R., Doescher, S. W., and Holm, T. M., 1991, Image browse in the global land information system. Technical Papers 1991 ACSM-ASPRS Annual Convention, 3, 294-301.
• Price, John C., 1987, Calibration of satellite radiometers and the comparison of vegetation indices: Remote Sensing of the Environment, 21, 15-27.
• Rao, Nagaraja C. R., 1987, Pre-launch calibration of channels 1 and 2 of Advanced Very High Resolution Radiometer: NOAA Technical Report NESDIS 36, Satellite Research Laboratory, National Environmental Satellite, Data, and Information Service, Washington, D.C.
• Running, S. W., Justice, C. O., Salomonson, V. V., Hall, D., Barker, J., Kaufmann, Y, J., Strahler, A. H., Huete, A. R., Muller, J.-P., Vanderbilt, V., Wan, Z. M., Teillet, P. M., and Carneggie, D., Terrestrial remote sensing science and algorithms planned for the Moderate Resolution Imaging Spectrometer (MODIS) of the Earth Observing System (EOS): The International Journal of Remote Sensing, 15, 3587-3620.
• Steinwand, D. R., Mapping raster imagery to the interrupted Goode Homolosine Projection: The International Journal of Remote Sensing, 15, 3463-3472.
• Steinwand, D. R., Hutchinson, J. A., and Snyder, J. P., 1993, Map projections for global and continental data sets and an analysis of pixel distortion caused by projection: in press.
• Teillet, P. M., 1990, Rayleigh optical depth comparisons from various sources: Applied Optics, 28, 1897-1900.
• Teillet, P. M., 1991, Radiometric and atmospheric correction procedures for AVHRR preprocessing in the solar reflective channels. Proceedings of the Fifth International Colloquium on Spectral Signatures of Objects in Remote Sensing, Courchevel, France, 1991, 511-516.
• Teillet, P. M, and Holben, B. N., 1993, Towards operational radiometric calibration of NOAA AVHRR imagery in the visible and infrared channels: Canadian Journal of Remote Sensing, 20 (1) 1-10.
• Townshend, J. T. (editor), 1992, Improved Global Data for Land Applications. IGBP Global Change Report No. 20. The International Geosphere Biosphere Programme: A Study of Global Change of the Council of Scientific Unions, Stockholm.
• Townshend, J. R. G., Justice, C. O., Skole, D., Malingreau, J.- P., Cihlar, J., Teillet, P. M., Sadowski, F., and Ruttenberg, S., The 1-km AVHRR global data set: needs of the International Geosphere Biosphere Program: The International Journal of Remote Sensing, 15, 3319-3332.
• Tropical Ecosystem Environment Observations by Satellite (TREES), 1991, A strategy proposal 1991-1993: Joint Research Center, TREES Series A, Ispra, Italy.