Identifying the drivers and predicting the outcome of conservation agriculture globally

Conservation agriculture (CA) is a potentially viable system for sustainable intensification across diverse agroecological and socio-economic landscapes. This analysis applied machine-learning techniques to a wealth of published data to create a predictive model of the agronomic outcome of CA relative to conventional practice (CP) based on 21 variables. The impact of different management scenarios were modeled by manipulating model input values for residue retention and N application rate, CA duration, and the ratio of CA to CP plant stand. Subsequently, the model was used to rank the importance of these variables in determining model outcome, and to create global maps of CA:CP outcomes for rainfed maize, wheat and soybean. Results showed that over-yielding of CA relative to CP was driven by a complex of climate, soil, geographic and management variables, and cannot be predicted accurately from precipitation amount or aridity index alone. Success of CA greatly increases with mean air temperature from 20 °C and with duration of CA for up to 13 years. Predictive maps showed that CA has the potential to increase system productivity in the humid tropics and sub-tropics given good plant stand establishment. Finally, this research demonstrates the ability of predictive modeling techniques to overcome the knowledge bottleneck created by the sole use of descriptive models within the CA literature to date. Predictive models can be used as an important tool by policy-makers and funding organizations to target financial resources to those regions where CA adoption will have the greatest impact on productivity.