Uncovering spatiotemporal impacts of landscape patterns on urban carbon emissions: Evidence from machine learning and case study in Wuhan

Global climate change has intensified the need to optimize urban landscape patterns (ULP) for achieving carbon neutrality. This study investigates the spatiotemporal relationship between ULP and carbon emissions (CE) in Wuhan's central urban area from 2000 to 2020. Integrating CE and land-use data within a 1 km grid, we selected the Geographic Random Forest (GRF) model through multi-model comparison. Following a three-tier index screening that identified nine core ULP metrics, we applied interpretable machine learning (ML) to identify nonlinear relationships and threshold effects. Results reveal a dynamic, stage-dependent evolution in dominant drivers: shifting from urban scale (Percentage of Landscape, PLAND) during early rapid expansion to spatial compactness (Aggregation Index, AI) in the mature phase, and finally to landscape complexity and irregularity in the current restructuring stage. Crucially, we identified a compactness paradox where emissions rebound when AI exceeds 75.0, challenging the linear denser is better assumption. Additionally, local geography moderates these relationships; specifically, extensive water bodies (ratio >0.2) elevate emissions by acting as physical barriers that hinder transport connectivity. These findings demand a shift from monolithic pro-compactness policies to differentiated strategies such as functional decompression in the core and morphological regularization in transitional zones.