Evaluation of Causal-Forest Estimation for Computer Agriculture Irrigation Scheduling Under Confounded Sensor Deployments: An Artificial Intelligence Heterogeneity Audit for Water-Use Equity

Sensor-driven irrigation scheduling in computer agriculture relies on causal estimates of how water application affects crop response, yet sensor deployments are rarely randomised and confounding by field position, soil type and management history is pervasive. the present analysis evaluates causal-forest estimation as a heterogeneity-aware alternative to pooled regression for such settings, and audits whether the resulting heterogeneity is stable enough to support water-use equity claims. The analysis is methodological and theoretical: it specifies a causal-forest estimator for irrigation scheduling under confounded sensor deployment, derives a heterogeneity-audit bound on subgroup treatment effects, and examines the conditions under which the estimator's honesty guarantee survives sensor-induced confounding. To ground the audit in a real quantitative setting, the paper draws on a random-effects meta-analysis of the country-specific association between access to electricity (% of population) and under-five mortality (per 1,000 live births) across 24 African countries, estimated from World Bank World Development Indicators for 2000 to 2023. The pooled effect was -2.19361 (SE = 0.26513, 95% CI [-2.71326, -1.67395], z = -8.274, p = 2.22e-16), with Q = 2862.465 on 23 degrees of freedom, tau-squared = 1.45635 and I-squared = 99.2%. That extreme heterogeneity is used here as a calibration analogue for the irrigation case: it shows how a single pooled coefficient can conceal country-level variation of the kind that causal forests are designed to expose. Causal-forest estimation is defensible for irrigation scheduling only when the heterogeneity audit is reported alongside the point estimates, and it proposes a reporting template that makes water-use equity claims falsifiable.