Causal Identifiability of Artificial Intelligence Attribution Maps for Satellite-Observed Deforestation: Instrumental-Variable Bounds and Enforcement-Priority Fairness

Attribution maps produced by convolutional networks for satellite-based deforestation monitoring are routinely read as evidence of where forest loss occurs, yet their causal interpretation is rarely justified. the present analysis formalises the conditions under which such maps identify the causal effect of land-cover change on enforcement priority, and derives instrumental-variable bounds when those conditions fail. Using a structural causal model of satellite observation, atmospheric interference and enforcement allocation, the analysis shows that attribution maps are causally identifiable only when the image-formation process satisfies an exclusion restriction that standard saliency methods do not enforce. A bound on the enforcement-priority bias is derived from the instrumental-variable estimand, and the bound is shown to depend on the strength of the instrument and the degree of unmeasured confounding. The framework is illustrated with a reanalysis of NASA POWER monthly temperature data for Lagos, Nigeria (2000–2023; 312 observations; mean 26.414 °C; warming trend 0.1553 °C per decade), which serves as a worked example of the identifiability conditions rather than as a deforestation dataset. Fairness in enforcement priority requires bounding the causal estimand, not merely calibrating the attribution map, and it specifies the conditions under which the bound is tight enough to support allocation decisions.