Riparian vegetation dynamics in the Beas and Sutlej catchments, India, 1989-2020

Winter season Landsat 5, 7 and 8 top of atmosphere (TOA) images between 1989 and 2020 were processed in Google Earth Engine (GEE) using images from the winter season (January to March). Cloudy pixels were masked by the algorithm provided in GEE for analysis of Landsat data. NDVI and mNDWI were calculated at each pixel, then averaged across all images to produce an image of winter season mean values. Those pixels with mean mNDWI > 0.15 (water surface) were excluded from the NDVI analysis. For each combination of section and zone, a spatially averaged, mean winter season NDVI was calculated for each year using GEE zonal statistics functions.

Landcover classification maps were created in GEE, using the same Landsat images as the NDVI analysis, by Gradient Tree Boost classifier with 10 decision trees. The raw data comprised the blue, green, red, NIR, SWIR1 and SWIR2 channels (bands 1-5 and 7 of Landsat 5/ 7 and bands 2-7 of Landsat 8). The classification was repeated for all seasons; winter (January to March), pre-monsoon (April to June), monsoon (July to September) and post-monsoon (October to December). The bands were averaged within each season using all the available cloud-free images.

The training datasets were created by manual inspection of the visible band colour images for each year of the 32-year record, identifying locations that were invariant in landcover. The classes and training locations are summarised in the supplementary information. Separate training sets (locations provided in .csv files) were used for map creation below the dams (Sutlej and Beas) and above the Pong Dam (Beas only) to mitigate effects due to differing spectral characteristics of soil and vegetation due to topography.

Taking each image in turn, as the output of the supervised classification process, the number of pixels within each section and zone were counted by landcover class. These counts were divided by the total pixel count in each area to obtain a fractional landcover for each class, by section and zone. This process was also repeated for the three test sites without differentiation into zone or section.

Quality control: Landsat satellite data was used to collect provide information on spectral surface reflectance, respectively. These data are accompanied by quality assessment bands which were used to select det afro processing, particularly with respect to obtaining cloud-free images. Processing was conducted by the project team, with using statistical analysis (such as consumer’s accuracy) and spot validation against aerial satellite photography. Data were checked for errors at all stages by members of the project team at Cranfield University.

Related datasets

River planform dynamics in the Beas and Sutlej catchments, India, 1847 and 1989-2018