Traditional smallholder farming systems dominate the savanna range countries of sub-Saharan Africa and provide the foundation for the region’s food security. Despite continued expansion of smallholder farming into the surrounding savanna landscapes, food insecurity in the region persists. Central to the monitoring of food security in these countries, and to understanding the processes behind it, are reliable, high-quality datasets of cultivated land. Remote sensing has been frequently used for this purpose but distinguishing crops under certain stages of growth from savanna woodlands has remained a major challenge. Yet, crop production in dryland ecosystems is most vulnerable to seasonal climate variability, amplifying the need for high quality products showing the distribution and extent of cropland. The key objective in this analysis is the development of a classification protocol for African savanna landscapes, emphasizing the delineation of cropland. We integrate remote sensing techniques with probabilistic modeling into an innovative workflow. We present summary results for this methodology applied to a land cover classification of Zambia’s Southern Province. Five primary land cover categories are classified for the study area, producing an overall map accuracy of 88.18%. Omission error within the cropland class is 12.11% and commission error 9.76%.

The objective of this analysis is to develop a method of classification error assessment that can contribute to techniques designed to evaluate change products with an emphasis on edge locations in landscapes. Reference land-cover data were developed within 10 randomly sampled regions across the study site, approximately 9 km² each. Sample pixels were partitioned and allocated among two domain categories, interior and edge. The assignment of pixels to domain categories was based on neighbourhood heterogeneity. Classification error was assessed within domain categories on a pixel-by-pixel basis, evaluating agreement between map values and majority land cover. To demonstrate the methodology, we present summary results for this technique applied to a land-cover classification of south-central Indiana. Omission error was less than 8.3% within interior domains and from 25.8% to 36.4% in the edge. Commission error was below 26.7% within the interior domains and ranged from 13.9 to 74.6% in the edge domains.