Modeling N2O emissions with remotely sensed variables using machine learning

Nitrous oxide is the largest source of greenhouse gas emissions from crop production. There is significant interest in targeting marginal lands for growing biomass crops, however little information is available on how this will affect N2O emissions from these crops. Furthermore, to characterize N2O emission at the farm level to quantify mitigation using measurements is time intensive, costly, and impractical. We selected a highly diverse watershed varying in soil texture and topography to compare two approaches for modeling soil N2O emissions using machine learning, intensive measurements of soil environment and climate variables, with the other only using remotely sensed variables. We confirmed that soil nitrogen was the most important variable followed by soil environment as influence by soil characteristic, topography, and climate. We also found that the machine learning model built on remotely sensed variables performed as well as when direct site level measurements were available. This finding supports the potential of using remotely sensed data to build machine learning models to characterize soil N2O emissions without the need for intensive soil measurements for entity level assessments.