Annual Image Composites

Our preliminary analysis showed that much of the rasters over India have no data until about 1990. Further, these data being originally derived from the historical data archive of Landsat satellites which are optical satellites prone to cloud occlusion, much of the rasters over India for the monsoon months (June to September) showed no data. 

We took India's dry (summer) season to be February-March-April and wet (post-monsoon) season to be October-November-December. Combining data from these months for each year, we calculated water occurrence maps, called seasonal composites, for every year* from 1991 to 2020, for dry and wet seasons separately. Compositing rule shown in the table below was used for arriving at the label value at every pixel based on its value in the three months of the season.

*Note: Here, we considered year to mean a hydrologic year going from June to May, and not a calendar year from January to December. Hence here, the year 1995, for example, is June 1995 to May 1996. Wet season of 1995 is October-November-December of 1995 and its dry season is February-March-April of 1996.

Using dry and wet season composites thus calculated, we further combined them to calculate water in dry season, water in wet season and permanent water (taken to be pixels that were water in both dry and wet seasons) for each year. We also calculated corresponding no data pixel values. The table below shows the rule applied to calculate these.

Conceptual motivation (and, hence, caveats) of our compositing approach

The compositing rules we used, shown above, are motivated by the goal of our analysis: to understand how surface water occurrence in India has changed at fine spatial scales over a period of 30 years, by looking at these changes separately in the dry season, wet season and permanent water scenarios. These three water regimes, individually as well as through dependencies and interactions among them, are vital for the ecology of our terrestrial freshwater ecosystems as well as our crucial socio-economic needs like agriculture and drinking water across India. Hence, quantifying trends in changes in these regimes can yield a granular understanding of our water systems.

We needed to combine, at the pixel-level, a year's 12 month data record of water occurrence—where a pixel can have one of three labels "water", "not water" and "no data"—to its seasonal (dry and wet) and permanent water. While combining a season's 3 months data to arrive at its composite, our rule is generous towards "water" pixels: water occurring at a pixel in any of the 3 months, regardless of the other 2 months, is taken to be sufficient to label it "water" for that season. Our rule towards "no data" pixels, on the other hand, is conservative: a pixel for a season is labelled as "no data" only when it is "no data" in all 3 months of the season.

The rules for calculating annual dry and wet season water is simple: a year's pixel is taken to be dry season water if it is a water in that year's dry season composite, regardless of the corresponding wet season. Likewise for wet season water. For permanent water, our rule is conservative: a year's pixel is taken to be permanent water only if it was "water" in both dry and wet season composites of that year. And finally, a year's pixel is taken to be dry season "no data" if it is "no data" in the year's dry season composite, regardless of the corresponding wet season composite. Likewise for wet season "no data". A year's pixel is permanent "no data" if it is "no data" in either or both of dry and wet season composites for that year.

Overall, our chosen compositing rules, and consequently our area estimates and trend analysis that follow, tend to be generous towards counting water area in dry and wet seasons and conservative in counting area of permanent water.