Mapping croplands of Europe, Middle East, Russia, and Central Asia using Landsat, Random Forest, and Google Earth Engine

Accurate and timely information on croplands is important for environmental, food security, and policy studies. Spatially explicit cropland datasets are also required to derive information on crop type, crop yield, cropping intensity, as well as irrigated areas. Large area – defined as continental to global – cropland mapping is challenging due to differential manifestation of croplands, wide range of cultivation practices and limited reference data availability. This study presents the results of a cropland extent mapping of 64 countries covering large parts of Europe, Middle East, Russia and Central Asia. To cover such a vast area, roughly 160,000 Landsat scenes from 3351 footprints between 2014 and 2016 were processed within the Google Earth Engine (GEE) platform. We used a pixel-based Random Forest (RF) machine learning algorithm with a set of satellite data inputs capturing diverse spectral, temporal and topographical characteristics across twelve agroecological zones (AEZs). The reference data to train the classification model were collected from very high spatial resolution imagery (VHRI) and ancillary datasets. The result is a binary map showing cultivated/non-cultivated areas ca. 2015. The map produced an overall accuracy of 93.8% with roughly 14% omission and commission errors for the cropland class based on a large set of independent validation samples. The map suggests the entire study area has a total 546 million hectares (Mha) of net croplands (nearly 30% of global net cropland areas) occupying 18% of the study land area. Comparison between national cropland area estimates from United Nations Food and Agricultural Organizations (FAO) and those derived from this work also showed an R-square value of 0.95. This Landsat-derived 30-m cropland product (GFSAD30) provided 10–30% greater cropland areas compared to UN FAO in the 64 Countries. Finally, the map-to-map comparison between GFSAD30 with several other cropland products revealed that the best similarity matrix was with the 30 m global land cover (GLC30) product providing an overall similarity of 88.8% (Kappa 0.7) with producer’s cropland similarity of 89.2% (errors of omissions = 10.8%) and user’s cropland similarity of 81.8% (errors of commissions = 8.1%). GFSAD30 captured the missing croplands in GLC30 product around significantly irrigated agricultural areas in Germany and Belgium and rainfed agriculture in Italy. This study also established that the real strengths of GFSAD30 product, compared to other products, were: 1. identifying precise location of croplands, and 2. capturing fragmented croplands. The cropland extent map dataset is available through NASA’s Land Processes Distributed Active Archive Center (LP DAAC) at https://doi.org/10.5067/MEaSUREs/GFSAD/GFSAD30EUCEARUMECE.001 , while the training and reference data as well as visualization are available at the Global Croplands https://croplands.org > website, GEE code is accessible at: https://code.earthengine.google.com/1666e8bed34e0ce2b2aaf1235ad8c6bd .