Understanding the Requirements for Surveys to Support Satellite-Based Crop Type Mapping

With the surge in publicly available high-resolution satelliteimagery, satellite-based monitoring of smallholder agriculturaloutcomes is gaining momentum. This paper providesrecommendations on how large-scale household surveysshould be conducted to generate the data needed to trainmodels for satellite-based crop type mapping in smallholderfarming systems. The analysis focuses on maize cultivationin Malawi and Ethiopia, and leverages rich, georeferencedplot-level data from national household surveys that wereconducted in 2018–20 and that are integrated with Sentinel-2satellite imagery and complementary geospatialdata. To identify the approach to survey data collectionthat yields optimal data for training remote sensing models,26,250 in silico experiments are simulated within a machinelearning framework. The best model is then applied to mapseasonal maize cultivation from 2016 to 2019 at 10-meterresolution in both countries. The analysis reveals that smallholderplots with maize cultivation can be identified withup to 75 percent accuracy. However, the predictive accuracyvaries with the approach to georeferencing plot locationsand the number of observations in the training data. Collectingfull plot boundaries or complete plot corner pointsprovides the best quality of information for model training.Classification performance peaks with slightly less than 60percent of the training data. Seemingly small erosion inaccuracy under less preferable approaches to georeferencingplots results in total area under maize cultivation beingoverestimated by 0.16 to 0.47 million hectares (8 to 24percent) in Malawi.