Annual trends

Annual rate of change in surface water area at the basin- and reach-scale

With the 30-years timeseries of annual surface water areas for every river basin and reach (transect), we estimated annual rate of change in surface water area using the Sen's slope linear trend estimator. This estimation was performed on dry season water, wet season water and permanent water separately for each of the basins and transects.

Part of our analysis flow chart, showing how annual rate of change in surface water area, by season, is estimated for river basins and reaches.

Sen's slope linear trend estimator is a popular non-parametric estimator which, unlike linear regression, does not require nor depend on Gaussian Normal distribution of regression errors. This makes it better suited for, among other scenarios, dealing with remotely sensed datasets and applications like as this one where such statistical structure does not apply.

A representative image of a surface water area timeseries chart along with straight line fitted to them.

A chart showing the timeseries surface water areas and the trend line estimated using Sen's slope estimator. Such a chart can be obtained from our Map for any river basin or reach in our analysis.

Handling data inadequacy during trend analysis

Some 30m x 30m pixels in the seasonal composites of water occurrence for each year can be "No data" pixels, indicating no valid data at these pixels. If a river basin or reach has too many such pixels, then the surface water area calculated for it—which comes from the set of remaining valid pixels, a count of "Water" pixels among them—would be an unreliable measure of its water area, and likely an underestimation. As a result, such data points in the timeseries could affect the accuracy of trend estimation.

To address this, we first the filter data points in each area timeseries to retain only those associated with almost no "No data" pixels. Water areas coming from time-points with area of "No data" pixels exceeding 5% of the total basin (or transect) area are deemed to be unreliable and hence dropped from regression analysis. The remaining points, deemed reliable, are used for estimating annual trends.

A representative image containing two time series charts. The first is of a surface water areas, with some data points missing in the series. The second is the corresponding timeseries of area of missing data ("no data" pixels area). Missing points in the first (water area) chart correspond to points in the second ("no data" area) chart with a high value.

Charts showing a timeseries of valid surface water areas along with the estimated trend line, and the corresponding timeseries of "No data" areas. Only those time points with fractional area of "No data" (bottom chart) less than 0.05, i.e., 5% of total area, are included in the water area time series trend estimation (top chart).

Reasons for data inadequacies

The JRC historical monthly surface water occurrence dataset contains pixels labelled as "No data" indicating that those pixels could not be confidently determined to contain valid data of either "Water" or "Not water". This could happen in snow-clad areas or because of reasons like pixel occlusion due to cloud cover (Landsat, an optical imagery satellite, cannot "see" through clouds), pixel not being imaged (which tends to happen in early parts of the timeseries, when Landsat data gathering did not have global coverage), etc. See Pekel et al. (2016) for more details.

The compositing rules we employ while calculating seasonal composites from this monthly history also yield "No data" pixels, based on particular combinations of pixel values in the constituent months. These pixels determine how reliable surface water area values in the timeseries are, and hence are used to filter out unreliable points before trend estimation.

Page updated

Google Sites

Report abuse