Improving smallholder representation in crop maps

Abstract

Abstract Farms smaller than five hectares produce more than half of the world's crops. Accurate maps of smallholder croplands thus hold great potential to inform a wide variety of policies including those that aim to address food insecurity, provide agricultural support, or encourage adaptation to climate change. However, many remotely sensed maps of croplands have been developed in ways that preferentially identify large, intensively managed fields. Here, we illustrate how remotely-sensed land cover maps could be improved to yield better information about smallholder farming, using Paraguay as a study system. We show that prior land cover maps dramatically underestimate smallholder croplands in Paraguay due to misclassification of these systems as a mixture of pastures, shrublands and forests. These maps fail to classify 55-99\% of smallholder fields ($<$ 5 ha) as croplands, overlooking more than 50 to 120 thousand farmers (or 16-43\% of the farming population) in the country. We develop national-scale land cover maps that more accurately identify smallholder croplands, conducting a series of experiments to identify specific methodological practices that enable the greatest improvements in smallholder cropland mapping. We show that modest adjustments to the training sample can improve smallholder crop recall in Paraguay by more than 27 percentage points compared to the best performing existing crop map, while other modeling adjustments, such as field segmentation and 10 m pixel resolution, improve smallholder crop recall by an additional 8 percentage points. Our analysis underscores the need for remote sensing scientists to more carefully consider crop system heterogeneity and to ensure that all farming systems are represented in training data for crop mapping.