SBIR Phase I: Instance Segmentation in Support of Sustainable Development

The broader impact/commercial potential of this Small Business Innovation Research (SBIR) Phase I project is to enable production of high resolution, timely, affordable data on agricultural productivity and economic outcomes across Sub-Saharan Africa (SSA), where access to demographic and market data is severely limited. The project proposes a set of machine learning methods for extracting objects, like farm fields and housing, from earth observation datasets. There are three key impacts enabled by the proposed technology. First, there are open scientific questions related to the efficient detection of objects in satellite imagery, particularly in rural developing country contexts, where training data are sparse. Second, the project is expected to significantly improve the accuracy of existing models for the prediction of socio-economic outcomes, by generating valuable new model inputs and improving the handling of missing data. Third, it seeks to expand the universe of socio-economic information that can be extracted from satellite imagery, by enabling the prediction of outcomes that previously have not been feasible. The resulting market intelligence products will revolutionize the cost effectiveness of agribusinesses, non-governmental organizations, and governments operating in SSA, enabling them to enter and serve fast-growing consumer markets across the continent. This Small Business Innovation Research (SBIR) Phase I project focuses on developing computationally efficient methods for segmenting images with rich geospatial content. It evaluates a series of deep learning architectures that extract relevant objects from composites of different earth observation datasets. A key technical challenge is designing models that make effective use of satellite imagery that spans multiple spatial resolutions, temporal frequencies, and sensing modalities. This presents unique opportunities and challenges compared to mainstream research on weakly supervised learning in traditional computer vision applications. The project's objectives are to 1) develop and test instance segmentation techniques that leverage the unique spatiotemporal information available in different types of earth observation data; and 2) innovate new learning approaches for handling missing inputs and dealing with sparse training data. Algorithms developed by this project will be evaluated using standard metrics for model performance, leveraging existing household survey data for testing and validation. Downstream benefits of the project include lower-cost access to market intelligence that can unlock underserved markets, expansion of the set of economic indicators that can be derived from earth observation data, and improved targeting of development assistance to people living in emerging economies, enabling cost-efficiencies for donors like USAID. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.