Counterfactual Data Augmentation for Artificial Intelligence Crop-Yield Prediction Under Structural Climate Regime Shifts: Identifiability Bounds and Food-Security Targeting
Authors/Creators
- 1. Agostinho Neto University, Luanda
Description
Crop-yield prediction models trained on historical climate data degrade when the climate regime shifts, because the joint distribution of covariates and yields changes in ways that no amount of resampling from the source domain can repair. Counterfactual data augmentation has been proposed as a remedy, but its validity depends on assumptions that are rarely stated and never tested. the present analysis formalises the problem as one of identifiability under a structural regime shift and derives bounds on what counterfactual augmentation can and cannot recover. Using NASA POWER monthly reanalysis air temperature at 2 m for Kampala, Uganda (0.35°N, 32.58°E) over 2000–2023 (312 monthly observations, mean 22.443 °C, standard deviation 1.043 °C, linear trend −0.1564 °C per decade), the paper shows that the observed warming signal is small relative to interannual variability, which tightens the identifiability bound considerably. Three results follow. First, counterfactual augmentation is identifiable only when the regime-shift mechanism is restricted to a known, low-dimensional set of climate variables. Second, the augmentation gain is bounded above by the mutual information between the shift mechanism and the yield outcome, which for the Kampala temperature record is modest. Third, food-security targeting under augmented predictions inherits the identifiability bound, so targeting rules that appear robust in validation can misallocate under shift.
Notes
Files
article.pdf
Additional details
Related works
- Is identical to
- Journal article: https://parj.africa/ajai_artificial_intelligen_25/article/view/counterfactual-data-augmentation-for-artificial (URL)
References
- Ben-David, S., Blitzer, J., Crammer, K., Kulesza, A., Pereira, F., & Vaughan, J. W. (2010). A theory of learning from different domains. Machine Learning, 79(1–2), 151–175.
- Bühlmann, P. (2020). Invariance, causality and robustness. Statistical Science, 35(3), 404–426.
- Coady, D., Grosh, M., & Hoddinott, J. (2004). Targeting of transfers in developing countries: Review of lessons and experience. World Bank Regional and Sectoral Studies.
- Grosh, M., & Baker, J. L. (1995). Proxy means tests for targeting social programs: Simulations and speculation. World Bank Living Standards Measurement Study Working Paper No. 118.
- Leonard, M., Westra, S., Phatak, A., Lambert, M., van den Hurk, B., McInnes, K., Risbey, J., Schuster, S., Jakob, D., & Stafford-Smith, M. (2014). A compound event framework for understanding extreme impacts. WIREs Climate Change, 5(1), 113–128.
- Pearl, J. (2009). Causality: Models, reasoning, and inference (2nd ed.). Cambridge University Press.
- Peters, J., Bühlmann, P., & Meinshausen, N. (2016). Causal inference by using invariant prediction: Identification and confidence intervals. Journal of the Royal Statistical Society: Series B, 78(5), 947–1012.
- Peters, J., Janzing, D., & Schölkopf, B. (2017). Elements of causal inference: Foundations and learning algorithms. MIT Press.
- Polyanskiy, Y., & Wu, Y. (2022). Information theory: From coding to learning. Cambridge University Press.
- Zscheischler, J., Westra, S., van den Hurk, B. J. J. M., Seneviratne, S. I., Ward, P. J., Pitman, A., AghaKouchak, A., Bresch, D. N., Leonard, M., Wahl, T., & Zhang, X. (2018). Future climate risk from compound events. Nature Climate Change, 8(6), 469–477.