Water Observations from Space (WOfS) is derived from Landsat-5 and Landsat-7 satellite imagery acquired over Australia between 1987 to November 2014. The Landsat data underpinning WOfS is ARG25 standard data located in the Australian Geoscience Data Cube (AGDC) at the National Computational Infrastructure (NCI) in the Australian National University (ANU), Canberra. The WOfS product is calculated from all acceptable Landsat scenes in the Geoscience Australia archive for the time period. The detection process is based on spectral analysis of each pixel in each Landsat scene.

The water detecton algorithm used to detect water from each observed pixel is based on a statistical regression tree analysis of a set of normalised difference indices and corrected band values. The regression is based on a set of water and non-water samples created by visual interpretation of 20 Landsat scenes from across Australia. The sample locations, ensure that the logistic regression is based on the full geographic range of conditions experienced in Australia.

The regression analysis determined a set of best indices and bands for the analysis and the associated thresholds in each component to derive a final classification tree, producing a water/non-water classification for every pixel in the Data Cube. The final water classification for each pixel is modified by Pixel Quality (see associated RG25 - PQ product information) and terrain.

Once the water algorithm has completed its process, the water detection for a pixel through time is combined to produce a total number of water observations for each pixel. This is compared to a total number of clear observations for the same pixel, derived from the PQ analysis. The ratio is expressed as a percentage water recurrence.

A separate analysis produces a confidence dataset, providing an assessment on whether a pixel depicted as having had water detected at some time is likely. The layer is computed by combining a set of confidence factors using a weighted sum approach, with the weightings derived by logistic regression. The confidence factors are:

1. MrVBF, a multi-resolution valley bottom flatness product (Gallant et al., 2012) derived from SRTM as part of the Terrestrial Ecosystems Research Network. Surface water pixels identified in valley bottoms were more likely to be positively detected.

2. Slope calculated from SRTM Digital Surface Models. Water pixels on a slope were considered less plausible than those on a flat surface.

3. MODIS Open Water Likelihood (OWL) (Ticehurst et al, 2010) provides a plausibility based an independent water detection algorithm employing the MODIS sensor. If both detection algorithms agree on the presence of a surface water pixel, there is a greater plausibility that the detection is correct.

4. Australian Hydrological Geospatial Fabric (Geofabric) is a GIS of hydrological features derived from manually interpreted topographic map grids. If known hydrologic features (pixels) from GeoFabric coincide with detected water pixels, the plausibility of detection is greater.

5. P, the number of observations of water as a fraction of the number of clear observations of the target pixel. P is high for more permanent water bodies.

6. Built-Up areas indicating areas of dense urban development. In such areas the water detection algorithm struggles to cope with the deep shadows cast by multi-story buildings and the generally noisy spectral response created by structures. The Built-Up layer is derived from the Ausralian Bureau of Statistics ASGS 2011 dataset, for urban centres of populations of 100 000 and over.

The product creation workflow is as follows:

1. Landsat raw data capture and storage

2. Data pre-processing (ARG25 and PQ products)

3. Water detection

4. Pixel Quality filtering

5. Data product storage and delivery

6. Time series data preparation

7. Summary and extent data preparation

8. Application of Confidence information

9. WMS/WCS service delivery