# Replication Package for Modelling the Global Potential of Smart Farming Innovations

*Shahrear Roman

Email: roman@uni-bonn.de

 

Description:

This dataset is the replication package for Roman, S., Hadi, H., Oluoch, W., Zilberman, D., Finger, R., & Wüpper, D. (2025), "Modelling the Global Potential of Smart Farming Innovations," Under Review: Proceedings of the National Academy of Sciences (PNAS).

The package provides all data, code, and model outputs needed to reproduce the manuscript results. It includes privacy-preserving training data (presence locations spatially displaced 5–10 km, constrained to cropland), ensemble prediction rasters for five smart farming innovations — (Semi-) Automated Application of Inputs, Autonomous Weeding Machines, Application Mapping for Precision Agriculture, GPS Steering Assistance Systems, and Autonomous Irrigation Systems — and aggregate innovation potential maps under three Environmental Performance Index (EPI) scenarios (Normal, Low, High) for 2025, 2035, and 2045.

Predictions were generated using an ensemble of five geospatial machine learning algorithms (GLM, Random Forest, SVM, MaxEnt, BRT) fitted within a 5-fold spatial block cross-validation framework and combined via AUC-weighted averaging across 25 sub-models. All global rasters are provided as GeoTIFF files at 5 km resolution (WGS84, EPSG:4326) and can be cropped or aggregated to any administrative or ecological unit by users. Analysis scripts are written in R (v4.4.x) with package versions locked via renv.

Keywords:

Smart Farming Innovations, Digital Technology, Sustainability Challenges, Technology Fit

 

Replication package for Roman et al. (2025), "Modelling the Global Potential of Smart Farming Innovations," PNAS. Contains privacy-preserving training data (5–10 km spatial noise), ensemble prediction rasters for five individual smart farming innovations and an aggregate index, scenario projections under three EPI trajectories (Normal/Low/High) for 2025–2045, and all R analysis scripts. Predictions use an AUC-weighted ensemble of five algorithms (GLM, RF, SVM, MaxEnt, BRT) with 5-fold spatial block cross-validation. Global GeoTIFFs at 5 km resolution (WGS84). Package versions locked via renv.