Item talk:Q149404

Mapping 30m irrigation extent across the United States by using a semi-automatic training approach on Google Earth Engine

Accurate and timely information on the distribution of irrigated croplands is crucial to research on agriculture, water availability, land use, and climate change. While agricultural land use has been well characterized, less attention has been paid specifically to croplands that are irrigated, in part due to the difficulty in mapping and distinguishing irrigation in satellite imagery. In this study, we developed a semi-automatic training approach to rapidly map irrigated croplands across the conterminous United States (CONUS) at 30 m resolution using Google Earth Engine. To resolve the issue of lacking nationwide training data, we generated two intermediate irrigation maps by segmenting Landsat-derived annual maximum greenness and enhanced vegetation index using county-level thresholds calibrated from an existing coarse resolution irrigation map. The resulting intermediate maps were then spatially filtered to provide a training data pool for most areas except for the upper midwestern states where we visually collected samples. We then used random samples extracted from the training pool along with remote sensing-derived features and climate variables to train ecoregion-stratified random forest classifiers for pixel-level classification. For ecoregions with a large training pool, the procedure of sample extraction, classifier training, and classification was conducted 10 times to obtain stable classification results. The resulting 2012 Landsat-based irrigation dataset (LANID) identified 23.3 million hectares of irrigated croplands in CONUS. A quantitative assessment of LANID showed superior accuracy to currently available maps, with a mean Kappa value of 0.88 (0.75–0.99), overall accuracy of 94% (87.5–99%), and producer’s and user’s accuracy of the irrigation class of 97.3% and 90.5%, respectively, at the aquifer level. Evaluation of feature importance indicated that Landsat-derived features played the primary role in classification in relatively arid regions while climate variables were important in the more humid eastern states. This methodology has the potential to produce annual irrigation maps for CONUS and provide insights into the field-level spatial and temporal aspects of irrigation.