A 30-m landsat-derived cropland extent product of Australia and China using random forest machine learning algorithm on Google Earth Engine cloud computing platform

Abstract

Mapping high resolution (30-m or better) cropland extent over very large areas such as continents or large countries or regions accurately, precisely, repeatedly, and rapidly is of great importance for addressing the global food and water security challenges. Such cropland extent products capture individual farm fields, small or large, and are crucial for developing accurate higher-level cropland products such as cropping intensities, crop types, crop watering methods (irrigated or rainfed), crop productivity, and crop water productivity. It also brings many challenges that include handling massively large data volumes, computing power, and collecting resource intensive reference training and validation data over complex geographic and political boundaries. Thereby, this study developed a precise and accurate Landsat 30-m derived cropland extent product for two very important, distinct, diverse, and large countries: Australia and China. The study used of eight bands (blue, green, red, NIR, SWIR1, SWIR2, TIR1, and NDVI) of Landsat-8 every 16-day Operational Land Imager (OLI) data for the years 2013–2015. The classification was performed by using a pixel-based supervised random forest (RF) machine learning algorithm (MLA) executed on the Google Earth Engine (GEE) cloud computing platform. Each band was time-composited over 4–6 time-periods over a year using median value for various agro-ecological zones (AEZs) of Australia and China. This resulted in a 32–48-layer mega-file data-cube (MFDC) for each of the AEZs. Reference training and validation data were gathered from: (a) field visits, (b) sub-meter to 5-m very high spatial resolution imagery (VHRI) data, and (c) ancillary sources such as from the National agriculture bureaus. Croplands versus non-croplands knowledge base for training the RF algorithm were derived from MFDC using 958 reference-training samples for Australia and 2130 reference-training samples for China. The resulting 30-m cropland extent product was assessed for accuracies using independent validation samples: 900 for Australia and 1972 for China. The 30-m cropland extent product of Australia showed an overall accuracy of 97.6% with a producer’s accuracy of 98.8% (errors of omissions = 1.2%), and user’s accuracy of 79% (errors of commissions = 21%) for the cropland class. For China, overall accuracies were 94% with a producer’s accuracy of 80% (errors of omissions = 20%), and user’s accuracy of 84.2% (errors of commissions = 15.8%) for cropland class. Total cropland areas of Australia were estimated as 35.1 million hectares and 165.2 million hectares for China. These estimates were higher by 8.6% for Australia and 3.9% for China when compared with the traditionally derived national statistics. The cropland extent product further demonstrated the ability to estimate sub-national cropland areas accurately by providing an R2 value of 0.85 when compared with province-wise cropland areas of China. The study provides a paradigm-shift on how cropland maps are produced using multi-date remote sensing. These products can be browsed at www.croplands.org and made available for download at NASA’s Land Processes Distributed Active Archive Center (LP DAAC) https://www.lpdaac.usgs.gov/node/1282.