A Deep Learning Approach to Agricultural Field Delineation in Nigeria

Nigeria plays an essential role in feeding Africa, providing a quarter of the continents crop production. With rapid growth in population and food demand projected for Africa in the coming decades, maximizing yields while minimizing environmental impacts will be essential for Nigeria to sustainably contribute to increased food supply for the continent. However, informing food security and sustainability efforts requires reliable agricultural statistics, which are not regularly collected in Nigeria, and advances in satellite sensors and remote sensing techniques provide a valuable alternative for gathering this much needed information. To this end we apply deep learning models to high-resolution (4 m) Planet NICFI basemaps to delineate individual agricultural fields in Nigeria. June 2017 cloud-free imagery is acquired from Planet user interface for the country. After dividing Nigeria into grid cells, random sampling methods generated over 700 sample points covering the various agroecological zones of Nigeria taking into consideration the f-weights of the samples. The sample points are the base for manual field delineation to train a U-Net model with deep learning algorithm to identify the fields within the study area. Using the training data, the model labels the fields and assesses for accuracy to validate the results. The generated field boundary map will be used to identify and understand spatial relationships between field sizes and other factors such as climate, soils, and proximity to cities and water bodies. Outputs from this work will provide a foundation for mapping staple crops and intercropping patterns and will provide valuable information for agricultural investment and policy in a country central to achieving food security in Africa.