Explainable AI for Economic–Environmental Trade-offs and Regional Heterogeneity in Land-Use Composition

Abstract

Tropical forests face accelerating pressure from land-use transformation, yet whether economic development necessarily entails environmental degradation remains contested. Resolving this tension requires moving beyond aggregate analyses to examine how land-use composition simultaneously shapes both dimensions across structurally diverse local contexts. This study therefore examines how land-use compositions shape economic and environmental outcomes at the municipal level in Pará, Brazil, a state where extensive land conversion has generated disproportionate greenhouse gas emissions alongside persistent economic underdevelopment. Employing a framework that integrates machine learning model comparison, SHAP-based interpretation, Gaussian Mixture Model （GMM） clustering, and spatial autocorrelation analysis, the results reveal substantial nonlinear and heterogeneous effects. Pasture and mining drive upward environmental pressure without commensurate economic returns, while perennial crops tend toward more environmentally favorable configurations. A key finding is that economic and environmental outcomes more frequently move in the same direction than in opposition, suggesting that land-use composition reflects broader structural conditions shaping both dimensions concurrently. The classic conservation-development trade-off is therefore not inevitable; pathways exist in which environmental protection and economic performance are mutually compatible, contingent on local structural context. The resulting municipal typologies further display significant positive spatial autocorrelation, indicating that similar trade-off structures cluster geographically across the state. Building on these findings, this study proposes an empirically grounded decision-support framework that translates municipal trade-off profiles into spatially differentiated planning typologies, highlighting the importance of context-sensitive strategies over uniform average-effect approaches.