Search

The availability of crop type reference datasets for satellite image classification is very limited for complex agricultural systems as observed in developing and emerging countries. Indeed, agricultural land use is very dynamic, agricultural censuses are often poorly georeferenced and crop types are difficult to interpret directly from satellite imagery. In this paper, we present a database made of 24 datasets collected in a standardized manner over nine sites within the framework of the international JECAM (Joint Experiment for Crop Assessment and Monitoring) initiative; the sites were spread over seven countries of the tropical belt, and the number of data collection years depended on the site (from 1 to 7 years between 2013 and 2020). These qualitycontrolled datasets are distinguished by in situ data collected at the field scale by local experts, with precise geographic coordinates, and following a common protocol. Altogether, the datasets completed 27 074 polygons (20 257 crops and 6817 noncrops, ranging from 748 plots in 2013 (one site visited) to 5515 in 2015 (six sites visited)) documented by detailed keywords. These datasets can be used to produce and validate agricultural land use maps in the tropics. They can also be used to assess the performances and robustness of classification methods of cropland and crop types/practices in a large range of tropical farming systems. The dataset is available at https://doi.org/10.18167/DVN1/P7OLAP (Jolivot et al., 2021).

With the unprecedented availability of satellite data and the rise of global binary maps, the collection of shared reference data sets should be fostered to allow systematic product benchmarking and validation. Authoritative global reference data are generally collected by experts with regional knowledge through photo-interpretation. During the last decade, crowdsourcing has emerged as an attractive alternative for rapid and relatively cheap data collection, beckoning the increasingly relevant question: can these two data sources be combined to validate thematic maps? In this article, we compared expert and crowd data and assessed their relative agreement for cropland identification, a land cover class often reported as difficult to map. Results indicate that observations from experts and volunteers could be partially conflated provided that several consistency checks are performed. We propose that conflation, i.e., replacement and augmentation of expert observations by crowdsourced observations, should be carried out both at the sampling and data analytics levels. The latter allows to evaluate the reliability of crowdsourced observations and to decide whether they should be conflated or discarded. We demonstrate that the standard deviation of crowdsourced contributions is a simple yet robust indicator of reliability which can effectively inform conflation. Following this criterion, we found that 70% of the expert observations could be crowdsourced with little to no effect on accuracy estimates, allowing a strategic reallocation of the spared expert effort to increase the reliability of the remaining 30% at no additional cost. Finally, we provide a collection of evidence-based recommendations for future hybrid reference data collection campaigns.