Landsat-based maps of cropping practices in the irrigated drylands of the Aral Sea Basin (1987-2019)

Overview

A set of maps revealing agricultural land use patterns in the irrigated drylands of the Amu Darya and Syr Darya basin between 1987 and 2019. The maps were produced using time series of 30m Landsat TM, ETM+ and OLI Collection 1 surface reflectance products and a Random Forest classification model trained with ~30,000 samples from the years 1987, 1998, 2008, and 2018. All processing steps were conducted in Google Earth Engine. The target classes of this product are "wet season cropping", "dry season cropping", "double cropping including fodder crops", and "non-cropland". Mapping was conducted across nine provinces in Uzbekistan, two in Turkmenistan, and three in Tajikistan, and post-processing was used to constrain the study area to regions which were irrigated in at least two years in the study period, areas below 2,000 m above sea level, and regions/years with a sufficient number of cloud-free Landsat images (n>6). Annual maps were aggregated temporally and spatially, resulting in different datasets described below.

We advise map users to read the open access paper and the associated supplementary materials for detailed insights. In case of further questions please contact the lead author of the work.

Data

This download contains three folders:

• landuse_30m: Land use layers with 30m spatial resolution, representing the percentage (0-100%) of the three key land use types ("wet season cropping", "dry season cropping", "double cropping including fodder crops") within two time frames (1987-2000 and 2001-2019). Years with insufficient cloud-free Landsat observations were excluded from the calculation of percentages.
• landuse_3km: Land use metrics with 3km spatial resolution, representing - for each grid cell and year - the "percentage of cropland", "percentage of dry-season cultivation", and "cropping frequency". The layers in these datasets are ordered chronologically, with the layer 1 representing the year 1987, and layer 33 the year 2019. Gaps in the time series were filled using linear interpolation for subsequent trend calculation using remotePARTS.
• mask: A mask with 30m spatial resolution constraining the study area to regions below 2,000 m above sea level and regions which were irrigated at least twice in the study period.

Map accuracy

We conducted an area-adjusted accuracy assessment based on a stratified random sample (n = 2,784 per year), which yielded important insights regarding accuracies and error types. The median area-adjusted overall accuracy of the maps across the study period is 91.4%, but class-specific user´s and producer´s accuracies vary substantially. Users should consult the supplementary materials of the article for details on accuracies, confusion matrices, and the most important error types.

Further resources
The production of this map was made possible through the Landsat Program of the United States Geological Survey (USGS) and the Google Earth Engine cloud computing platform for preprocessing of the satellite data and classification. The code for preprocessing the Landsat time series is based on the Google Earth Engine Python API and made available at https://github.com/philipperufin/eepypr/.