For estimating annual trends in surface water extents, we use the Sen's slope trend estimator for the 30-year period 1991-2020 which yields an estimate of a monotonic (i.e., straight line) trend. Non-monotonic changes have not been analysed in this work.
Hence, while surface water flows and accumulations in rivers, lakes and water bodies are affected in complex ways by ground water, seasons, land cover, anthropogenic alterations, etc., the non-monotonic changes these can bring about to surface water have not been captured in this work.
The original JRC monthly surface water occurrence raster dataset has pixels labeled `No data`, indicating invalid data (Pekel et al. 2018). Hence, some pixel values in our annual image composites result in invalid data as well, based on criteria in our image compositing rules. These invalid data pixels are primarily due to snow cover, occlusion due to cloud cover or unavailability of satellite image in the Landsat archive. Refer to the JRC dataset description for more details.
The Landsat family of satellites are optical satellites, so they cannot "see" through the clouds. Hence the earth's surface (likewise, surface water) can be occluded by clouds in Landsat images. Such occlusions and atmospheric conditions can lead to invalid data values in surface water occurrence dataset. These factors affect the data particularly in the Indian monsoon months (typically, June to September) when the skies are thick in cloud cover,
Invalid data can also be a result of technical issues in the satellite sensor, or due to snow cover on the ground.
Satellites in the Landsat mission, while having orbits spanning the globe, did not acquire imagery over some regions or the acquired imagery do not meet certain quality thresholds. From preliminary analysis of the JRC dataset, we found that for most of India, regularly collected imagery starts after 1990.
From preliminary analysis of the JRC dataset, we found that for most of India, sufficient extents of valid data begins after 1990. The lack of valid data before that is in part because satellites in the Landsat mission, while having orbits spanning the globe, did not acquire imagery over some regions and this appears to affect the JRC dataset till around 1990. Missing data could also result from acquired images failing to meet acceptable quality thresholds.
The spatial resolution of the JRC dataset, like that of the Landsat image dataset it is derived from, is 30m/pixel. Hence water bodies at roughly this size or smaller can be missed in the process of surface water detection.
When water bodies have emerging vegetation in them or overhanging vegetation around their edges the water surface may not be openly visible in Landsat satellites. Such patches of water bodies, despite having surface water, may be missed because it is not open surface water visible to the satellites.
Refer to the published here for more details.
Landsat satellites have a typical revisit cycle of 16 days. These are used to build monthly snapshots of surface water occurrence. Since phenomena such as river floods, dam releases, spurts of snow-melt, etc. can occur at time scales shorter than these, they may not be well represented in this dataset. Refer to the published here for more details.
In the methods we follow, a pixel in a year's composite image for dry (or wet) season is labeled as a surface water pixel when that pixel had surface water in any one of the season's three months of that year. The same pixel in that year's composite image for permanent water, on the other hand, is labeled as a surface water pixel when that pixel had surface water in both dry and wet season composite images of that year. See our methods for more details.