An explainable digital twin for balancing food production, resource use and post-harvest waste: a groundnut case study from Tamil Nadu
Food planners trade marketable output against water, fertilizer and post-harvest losses, and yield models predict output without pricing those trades. We built a computational digital twin of groundnut production in 36 Tamil Nadu districts. A Ridge regression surrogate predicts district production from cultivated area and previous production. SHapley Additive exPlanations (SHAP) attribute its predictions to the inputs, and the non-dominated sorting genetic algorithm II (NSGA-II) optimizes four objectives: marketable food, water use, a fertilizer-resource index and post-harvest waste. The district records sum to the official 2022–23 state totals within 0.001%. In repeated cross-validation the surrogate reached a coefficient of determination of 0.920±0.058 and a mean absolute error of 3,923 t, and adding weather variables lowered the score to 0.916. Previous production outweighed cultivated area, with mean absolute SHAP values of 15,321 t and 6,471 t. Dropping the waste objective cut the normalized hypervolume from 0.39177 to 0.32708. The balanced Pareto solution reduced modelled water use by 21.23% and waste by 60.60%, and it cost 20.81% of marketable food. Harvested output fell by 21.25%, so the water saving comes from producing less, and the waste saving comes from an assumed halving of the loss rate. All reductions are simulated and no field measurement supports them.
Authors
- Arshed Ayoob
- ArunPandian J
- Yoganand Selvaraj
Institutions
- Vellore Institute of Technology University (IN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-20
- DOI
- https://doi.org/10.21203/rs.3.rs-11589751/v1
- Primary Topic
- Smart Agriculture and AI
- Type
- preprint
Funders
- VIT University