Introduction

This document sets out the first of a series of evaluations of algorithms, which comprise the core processing steps in the creation of the CSIRO AVHRR Time Series (CATS).

CATS is the principal demonstration product of EOC Task 4.3: Product Presentation, Quality & Standards. The overarching objective of Task 4.3 is to demonstrate the quality and consistency of a widely applicable dataset (CATS) that results from coordinated and systematic processing enjoying the cooperation of CSIRO best teams.

The CATS demonstration will include quantifying the benefits of each of the CSIRO Best Practice algorithms applied in the processing. The scientific research component underpinning this first demonstration objective is summarised as the systematic, progressive and quantifiable removal of non-surface contributions to the AVHRR data archive.

CSIRO Best Practice Algorithms

Task 4.3 has identified six processing algorithms, which are critical to the overall outcome. These are, in order of processing sequence:

1. Noise removal
2. Calibration
3. BRDF
4. Atmospheric correction
5. Cloud flagging
6. Navigation and remapping.

The objective of the evaluation exercise is to choose amongst candidate algorithms for each of the above processing steps, and then incorporate the chosen algorithm into the CATS processing stream.

Because CATS carries the Organisation name and will be used by scientists in differing research areas, the Best Practice selection procedure must represent CSIRO wide opinion and be transparent to any interested party. This document is one component of the effort of Task 4.3 to address these important requirements. One prior component has been the open discussion of these topics at the last two EOC Annual Science Meetings – particularly that in Perth, 1999. In addition, most of the selections are not or cannot be contentious because we are dealing with mature areas of AVHRR processing experience.

Outcomes of the Selection Process

The principal and immediate outcome of the selection of a processing algorithms as CSIRO Best Practice is that it will be applied in the production of (the latest edition of) CATS. As set out in the Task 4.3 documentation, it is anticipated that CATS will pass through several editions, with each successive one being a demonstrable improvement over the preceding version. It follows that the Best Practice algorithms will be subject to continuing testing as research and validation continues.

A secondary outcome is the implementation of a processing algorithm within software available to CSIRO staff. Task 4.1 (CAPS) represents a formal effort in the production of such software and the CAPS Working Group is to ensure its evolution. However, other AVHRR software packages or implementations exist within and without CSIRO, having been developed independently to serve specific needs. They exist in the individual, public and commercial domains. As part of the Best Practice evaluations, all three sources of implementation will be each evaluated in terms of utility and efficiency.

The secondary outcome is acknowledged by the Best Practice evaluations set out below but it remains a secondary outcome to be undertaken by others. Task 4.3 is not the CAPS Working Group. The future usage within CSIRO of AVHRR processing using CAPS, or LAS or ENVI/IDL or any other package, will be determined by the level of support it generates by satisfaction of core user needs, namely accuracy, utility and efficiency. The Best Practice evaluations all include measures of utility (the ease of scripting for batch processing) and efficiency (CPU time).

Evaluations in progress

1. Noise Removal (Edward King) [ACTIVE]

The objective of this algorithm, or processing module is to convert the fragmented ‘raw data’ archived from four receiving stations (Darwin, Townsville, Hobart and Perth) into minimal noise, continent spanning passes. Data from each of these stations have different noise characteristics and their full horizon-to-horizon passes do not alone span the continent. The ‘stitching’ of individual station passes into one is prerequisite to down-stream processing, while the noise minimisation provides the best foundation data possible. Furthermore, the creation of continent-spanning passes reduces redundancy in the archive.

There was no satisfactory algorithm and implementation available to achieve the overall objective. Therefore, guided by the considerable experience of colleagues in CMR and CAR and an existing implementation within the LAS package created for the ‘Global 1 km Land Data Project’, a new set of algorithms was written (based on LAS) and implemented at the EOC. It is currently stitching the HRPT archive, 1992-present.

Based on visual assessment of the quality of the stitched output, particularly of the very noisy N15 data stream, the results are excellent.

A quantitative demonstration of the performance of the stitching module will be made by comparing the noise characteristics of stitched images with equivalents from just one (high quality) station (ACRES, Alice Springs) which is not a contributor to the archive. These comparisons of mean residual noise per scene will be framed as an ANOVA to include the factors of spacecraft (N11, N14) and time post-launch. For this demonstration, there are no requirements for ancillary data, and no requirements for independent validation. The Noise Removal Best Practice evaluation is in the process of publication as an EOC Report to be completed before the next EOC Annual Meeting.

2. Calibration (Dean Graetz, Susan Campbell, Jenny Lovell) [PENDING]

Within the AVHRR data archive, calibration of the optical channels represents the largest uncertainty in interpretation. Uncertainty in calibration exceeds that contributed by the atmosphere or by BRDF.

The AVHRR calibration uncertainty issue is very significant because it precludes any credible demonstration of multi-year trends in the behaviour of the land surface. With uncertain calibration, there can be no monitoring - which is the principal application driving Task 4.3.

Current CSIRO Best Practice in AVHRR Calibration is CALWATCH; see CSIRO Atmospheric Research Technical Paper No. 42. This very valuable summary of published and unpublished calibration data supports the CAPS calibration module. CALWATCH is essentially expert opinion because it is not underpinned by any measurements made by CSIRO. Effectively, and perhaps unwisely, CSIRO is a passive user of critical parameters provided by others.

The current passive situation is unsatisfactory. Calibration is the key to the quality and thus usefulness of the CATS output. It is critical that Task 4.3 has complete confidence in the sensor calibration used. CALWATCH needs to be supplemented by an EOC capacity to choose between alternative calibration datasets provided by NOAA, and others.

To illustrate: Ross Mitchell (CSIRO Atmospheric Research Technical Paper No. 42, Figure 1) offers two very different calibration datasets for N14 provided by the operators of the spacecraft, NOAA (Rao & Chen 1999), and by an independent research group (Vermote & El Saleous, unpublished). The differences in the temporal behaviour of gain are quite significant for long term trend detection. Is either of these datasets correct?

A proposal for a measurement program to underpin Best Practice in Calibration is under development for presentation to the next EOC Annual Meeting.

3. BRDF (Jenny Lovell, Dean Graetz) [ACTIVE]

The objective of this module is to quantify the influence of varying sun-target-view geometry (BRDF) on the recorded signal. Having quantified this contribution, it can be normalised so that it has little or no influence on the interpretation of time series observations.

BRDF is assessed as contributing the third largest source of ‘noise’ to the surface signal, after calibration and atmospheric correction. However, BRDF is not noise but signal because it contains information about the surface. Overall, the absolute size of its influence on interpreting an AVHRR time series is assessed at approximately 10%.

Incorporating BRDF modelling and BRDF correction into the processing of a multi-year AVHRR archive remains a pioneering effort. The CCRS ABC3 project represents the most advanced attempt at generating a BRDF-corrected AVHRR time series. Their approach was to first stratify the landcover of Canada into nine types and then fit the model of Roujean et al (1992) to obtain coefficients for each of the landcover types. It was a ‘stratify and designate’ strategy. The CCRS effort found that applying the strategy nation-wide resulted in a relative reduction of BRDF-dependence in reflectance of 50%(Channel 1) and 60%(channel 2).

The Best Practice strategy has evolved from the contributions of others, such as Ian Grant and Task 3.3. Based on a meticulous processing of the (noisy) POLDER data, a (Roujean model) description of the BRDF of the continent has been derived. This dataset, which was acquired over a period of 8 months in 1996/7, is being interpreted in terms of static and dynamic landcover attributes. Given a satisfactory and robust mechanistic interpretation, the (Roujean model) BRDF will be used to normalise the observed AVHRR reflectances to albedos.

Best Practice evaluation will involve comparison between the BRDF derived from POLDER with that derived from AVHRR itself, algorithms and code that is a product of Task 4.2 and to be incorporated into CAPS (Task 4.1)? However, that evaluation cannot be made just at present but will occur later in 2000.

4. Atmospheric correction (Dean Graetz, Jenny Lovell, Susan Campbell) [PENDING]

After calibration, atmospheric correction (ATMOCO) is the next largest source of uncertainty in interpreting surface behaviour using the AVHRR archive. ATMOCO of continental data is not yet routinely attempted but will be done in stages within CATS – see Task 4.3 documentation.

The question is - which is the best ATMOCO model given the chronic absence of input variables throughout the AVHRR archive? It can be confidently predicted that the chronic absence of the required input measurements (aerosols and water vapour) will limit adequate ATMOCO of the AVHRR archive for some time yet.

Nevertheless, it is currently planned that Best Practice will be selected from three models: NONE, ATCO (Task 4.2) and SMAC (Rahman & Dedieu, 1994).

A proposal outlining strategy and a measurement program to underpin Best Practice in ATMOCO is under development for presentation to the next EOC Annual Meeting.

5. Cloud flagging (Susan, Jenny, and Dean) [ACTIVE]

Operational (machine-only) cloud flagging for AVHRR data over land during day and night is a stale rather than mature area of research. Many methods have been published but demonstrations of their effectiveness are usually reduced to a few images of landscapes that are mostly easy to work with, ie no bright deserts or salt lakes. While several cloud detection strategies are in use, there appears no clear winner and the probability of undetected residual cloud is real, persistent but small. This unflagged cloud must always be expected and must be removed by other processing, such as during compositing.

Cloud shadow is a real and significant problem, especially for vegetation indices based on channel 2. Cloud shadow is usually ignored on the assumption that it will be removed in a composited product. One algorithm for explicitly locating and flagging shadows has recently been published (Simpson and Stitt, 1998). The computational load appears significant and a far more efficient technique appears to be flag cloud first and then apply a dilation filter of 1-2 pixels wide (depending on sun angle) to produce a second more extensive area of flagged cloud + shadow.

The desired result of a CATS cloud flagging module is the generation of an appended data layer of cloud (and cloud shadow) flags for day and night images of the land surface. No actual cloud clearing would be done because the CATS processing strategy do not foreclose how the user subsequently interprets these flags. Cloud clearing in a CATS product is only forced at the aggregation stage.

During the EOC Annual Meeting in Perth (1999), the above strategy was advocated by the experienced. As Best Practice, the CLAVR algorithm of Stowe et al (1999) was recommended as the most productive option by Peter Turner who subsequently implemented it into CAPS. CLAVR produces a set of flags depending upon the results of a sequence of tests but the interpretation of these flags for the Australian continent is not yet determined.

Because the users of cloud flagging have differing objectives, the Best Practice evaluation has been enlarged to a comparison of three techniques/implementations:

1. The current CAPS implementation of CLAVR (spectral-spatial)
2. An intelligent filtering of a time series (spectral-temporal)
3. GEMI (custom spectral)

The factors expected to influence the performance of the three implementations are:

1. Cloud type and patterning (3 levels: cumulus, stratus, cirrus)
2. Land/cloud spectral contrast (2 levels: high, low)
3. Scan angle (2 levels: _ 27.5 °, _ 27.5 °)

The data used for the test is a geocoded subset of a navigated, remapped and registered time series of N14 afternoon passes for the period October 1-30, 1995. The target area will be quite large, approx. 30° x 30 °.

The absolute standard will be that which human interpreters manually produce. The comparison between machine and human interpretation will be made for replicated 1° x 1° subsets chosen to cover the twelve factors by level combinations set out above, ie. 2 * 3 * 2 * 2 or a total of 24. The three implementations will be compared for accuracy of outputs at each of these 24 subsets, and overall in terms of efficiency (CPU time) and utility (ease of scripting for batch processing).

The processing will begin when access to the CSIRO silo becomes available (May 1). The findings will be presented at the next EOC Annual Meeting and published as an EOC Report.

6. Navigation and remapping (Susan, Jenny, Dean) [ACTIVE]

Though listed last, navigation and remapping is critically important. Any use of AVHRR data is primarily constrained by the quality of its navigation and remapping. Navigation and remapping were left until last because in the CATS processing stream, increasing data storage space imposes a greater penalty than increasing processing load. Therefore, CATS processing is planned around (temporarily) remapping any ancillary datasets into the projection of the non-navigated image while leaving the remapping of the much larger AVHRR image until last. However, the 1999 Perth EOC meeting recommended that Best Practice in navigation and remapping be evaluated this FY.

This evaluation is by far the largest to be undertaken as part of Task 4.3. It will be attempted in stages but we start by setting out the full experimental design.

Implementations:

1. CAPS (current version)
2. LAS 1 (coastline vector method)
3. LAS 2 (continental image method)
4. ENVI (version 3.2)

Factors

1. Sensor (5 levels: N9, N11, N12, N14, N15)
2. Overpass time (2 levels: day reflective, night-thermal)

Data

Metric

1. The overall r2 of the polynomial warp required to match the output image to the reference image.
2. Sum (delta line**2 + delta pixel**2)
3. The ‘Predictive Error Map’ function, to be ported from MicroBrian, if possible.

The first (and learning) stage of the evaluation will be for 4 implementations, 1 sensor (N14), 1 overpass time (afternoon), and replicate images from the current stitching activity, such as the month of October 1995, the same dataset being processed for the Cloud Flagging Best Practice evaluation

The second stage of the evaluation will be a repeat except that less replicates will be used and nighttime images will be included. The third and final stage will include the five sensors for day and night images.

The processing will begin when access to the CSIRO silo becomes available (May). The findings will be presented at the next EOC Annual Meeting and published as an EOC Report.