SBIR Phase II: Instance Segmentation in Support of Sustainable Development

The broader impact of this Small Business Innovation Research (SBIR) Phase II project will be to improve the efficiency of food production and supply chains for small-scale farming systems. This project advances high-resolution, in-season crop yield forecasts, focusing on maize yields in Sub-Saharan Africa with technologies that can be extended to global small-holder agriculture. The project will address three needs: 1) the design of financial products and services for small-holder farmers, including credit and crop insurance models; 2) the planning of harvest operations and efficient linkage of produce to markets; and 3) the detection of lower than average yields, and the mitigation of resulting threats to food security. This can help service providers, producer groups, traders and aggregators, and government policy-makers. In addition to commercial and societal impacts, this innovation will advance the state of the science in yield forecasting, by adapting methods used in large-scale commercial production for the smaller-scale, heterogeneous farm plots typical of the developing world. This Small Business Innovation Research (SBIR) Phase II project will develop a novel method for forecasting plot-level maize yields, using high resolution satellite imagery and other remotely sensed data as inputs. The method is calibrated and tested using field data from four countries in Sub-Saharan Africa. A first research objective is to implement and evaluate a variety of computationally efficient modeling approaches for in-season crop area classification, at the level of the small-holder plot (for which no method is currently established). A second objective is to design and calibrate a pixel-level yield forecasting model that generates estimates at multiple timepoints across the growing season. Various calibration approaches will be tested, using both public and proprietary data on historical yield anomalies. The project addresses several persistent challenges in yield forecasting, including the needs for: flexible fusion of remote sensing data that span multiple spatial resolutions, temporal frequencies, and sensing modalities; model architectures that can handle sparse data (given limited access to field-level ground-truth data for calibration and validation); and scalable approaches that can perform in different geographies and agro-ecologies. 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.