UNSUPERVISED SEGMENTATION OF SMALLHOLDER FIELDS IN MOZAMBIQUE USING PLANETSCOPE IMAGERY

Abstract

Smallholders produce about a third of the global crop production. Supporting these smallholder farms is an important lever for poverty alleviation. Farm and field sizes are key indicators of many smallholder dynamics, including fragmentation, farm consolidation, and interactions between smallholders, medium-scale commercial farming, and large enterprises. Despite the socio-economic, environmental, and political importance of these dynamics, spatially explicit data on farms and field sizes are still lacking. Identifying small-scale agriculture using satellite imagery is challenging due to the heterogeneity in the crop types and management practices. This study compared three unsupervised segmentation approaches that have not been widely explored for delineating smallholder fields: mean shift, multiresolution segmentation, and simple non-iterative clustering (SNIC), using PlanetScope imagery. The study area is located in northern Mozambique, where 71% of the farms cover less than 2 ha. The results were evaluated using four segmentation accuracy metrics based on object geometries: Area Fit Index (AFI), Quality Rate (QR), Oversegmentation (OS), and Undersegmentation (US). The results showed that the multiresolution segmentation algorithm outperformed the other methods to delineate smallholder fields. This work will support future regional-scale mapping efforts.