Mapping cropland management in Niger using remote sensing time series

Cropland plays an essential role in guarding food security and human well-being. To meet the increasing demand for food and energy, monitoring of cropland is much needed for supporting evidence-based decision-making. In West Africa (WA), one of the most vulnerable areas to climate change worldwide, given the unprecedented population growth and recurrent extreme events (heat weaves, drought and flooding), the need is particularly urgent. Remote sensing provides a great tool for cropland monitoring. However, the heterogeneous landscapes, cropland dynamics and diverse land use management (e.g. water-harvesting practices, Zaï technique) pose substantial challenges. Mapping such heterogeneity is difficult due to the overlap of spectral signatures and the requirement of dense observations. Here we present the usage of open-access Sentinel-1, Sentinel-2 and Landsat imagery time series to map cropland extent, crop types, and management (i.e. irrigated, rain-fed cropland, and Zaï). Specifically, (i). we generated different features from Sentinel-1, Sentinel-2 and Landsat imagery for classification. We performed feature selection and compared mapping accuracies with and without feature selection; (ii). we tested three commonly used classifiers, Maximum Likelihood (ML), Support Vector Machine (SVM), Random Forest (RF) and an ensemble approach for cropland monitoring. Compared to existing land cover products or cropland maps, our cropland maps not only showed a higher mapping accuracy, but also depicted more detailed land use management. Our study demonstrated the challenges and opportunities of remote sensing time series for cropland monitoring in a heterogeneous environment. The generated maps provide a reference for assessing distribution and resilience of food production systems and thereby for supporting spatial planning towards the implementation of the Sustainable Development Goals in WA.