Tracking four years of smallholder cropland dynamics at high resolution and national extent

Keywords: smallholder agriculture, machine learning, CubeSats, land cover change, field boundaries, Ghana
Session Type: Virtual Paper Abstract
Day: Friday
Session Start / End Time: 2/25/2022 11:20 AM (Eastern Time (US & Canada)) - 2/25/2022 12:40 PM (Eastern Time (US & Canada))
Room: Virtual 38

Abstract

Tracking the changing characteristics of small-scale croplands is crucial for understanding a host of socioeconomic and environmental issues, including questions about food security, the trajectory of agricultural development, and the environmental consequences of cropland change. Until recently, mapping smallholder-dominated croplands for even a single point in time, let alone across multiple time points, has proven to be major challenge, given the reliance on satellite remote sensing for the task. However, new generations of satellite fleets together with advances in cloud computing and machine learning have made it increasingly possible to create accurate annual maps at national to regional scales. In this presentation, we develop a four-time series of field boundary maps for Ghana for the years 2018-2021. Our approach builds on a country-wide proof-of-concept field boundary map for 2018, from which we estimated the average size and total number of fields to be 1.73 ha 1.7 million fields, respectively. The mapping approach integrates high resolution PlanetScope imagery collected from two seasons in each agricultural year, and a convolutional neural network (U-Net) trained to distinguish between field interiors and field edges, using labels collected with a custom platform designed to improve label accuracy. The resulting classifications of field interiors are combined with an object-based image analysis to delineate the boundaries of individual fields. We use the resulting maps to identify recent changes in Ghana’s agricultural systems, providing insight into the dynamics of swidden systems as well as trends in agricultural intensification and expansion.