Automated detection and segmentation of smallholder fish ponds in Nigeria

Abstract Fish and other aquatic foods are a vital source of nutrients and livelihoods around the world, and with the stagnation of capture fisheries, countries have increasingly turned to aquaculture. While the geographic expansion of aquaculture is evident from national statistics, much less is known about subnational patterns of growth. Notably, the spatial extent of smallholder aquaculture ponds is poorly understood, limiting our ability to ask basic questions about local risks and benefits stemming from aquaculture investments. To begin to address this, we develop an open-source automated segmentation pipeline to systematically delineate individual fish ponds. As Africa's most populous nation and third-largest fish producer, Nigeria plays a crucial role in regional food security. With its diversity of geographies and fish production systems, Nigeria offers the ideal case for developing automated approaches to fine-scale fish pond mapping. First, we use georeferenced point labels of fish pond locations to manually delineate fish pond boundaries and create a labeled dataset (N=14,021). Combining these labels with high-resolution satellite imagery, we then train a national segmentation model and perform predictions across the entire country (precision: 86.3%; recall: 87.9%). We then train a post-processing random forest classifier to distinguish fish ponds from false positives (precision: 99.1%; recall: 98.5%). Applying this approach, we identify a total of 250,936 fish ponds throughout Nigeria. Our scalable approach offers a flexible means for fine-scale tracking of smallholder aquaculture patterns potentially applicable across other smallholder-dominated geographies. Such data are central for food system monitoring and informing interventions to bolster smallholder productivity and market access.