Coupling Earth Observations, Machine Learning, and Optimization to Uncover Productivity Drivers in Insect Pollination-Dependent Crops

The global food system is highly dependent on a range of ecosystem services. Understanding the interdependence between crop productivity, climate, and ecosystem services is crucial for resilient and plentiful supply of food. A key challenge in food security is reliable prediction of crop yields because they are dependent on several biotic and abiotic factors with complex and often non-linear interactions. Recent research has made important progress to predict crop yields based on earth observation and weather datasets. However, there is a gap in the literature regarding the inclusion of ecosystem services in yield models. One such critical ecosystem service is pollination provided by wild and managed bees. Studies suggest that the abundance and diversity of bees are threatened by insecticide use, climate change, and habitat loss. However, the role of crop management practices and environmental conditions on pollinators and consequently on crop yields is poorly understood. There is also a need for systems-level studies to identify production areas that are efficient at resource utilization and areas suitable for interventions and restoration efforts.

We develop machine learning models leveraging the USDA-NASS Cropland Data Layer, satellite observations from the MODIS sensor, and climate data from PRISM to predict crop yield at both county-level and 250m scale. The downscaled yield is then coupled with Data Envelopment Analysis (DEA) to understand how resources are used and what drives productivity in crop fields. We specifically derive relative efficiencies as the ratio between outputs (e.g. yield) and inputs which include management practices such as surrounding landscape, irrigation, and pesticide use. The model is applied to soybeans due to its importance to US agriculture and the need to include pollination in soybean studies. The machine learning models can predict yield with high coefficient of determination (0.85). Preliminary results indicate that a very small portion of fields are operating at maximum relative efficiency, with toxic load from pesticides being a key element driving inefficiency in most cases. Current work is in progress to evaluate almonds, a crop heavily dependent on bee pollination. The implications of our findings for pollinator management and policymaking will be discussed.