都市の炭素隔離を最大化するための深層学習ベースの代理モデリングと都市森林配分の最適化の統合

Urban afforestation has emerged as a key strategy for enhancing carbon sequestration to mitigate climate change. However, afforestation strategies driven by area-expansion targets can overlook the spatial heterogeneity of carbon uptake and the synergistic effects of ecological connectivity. This study presents an integrated optimization framework that incorporates a deep learning-based surrogate model, combining an Artificial Neural Network (ANN) with a Genetic Algorithm (GA), to optimize afforestation for maximum Net Primary Productivity (NPP). The ANN model, trained on topographic, climatic, land-use, and landscape variables, demonstrated high predictive accuracy (R² = 0.82 on test data; R² = 0.87 on training data) in estimating NPP, effectively capturing the nonlinear relationships influencing carbon sequestration in urban areas. SHapley Additive exPlanations (SHAP) analysis revealed that proximity to existing forests was the most influential factor in NPP enhancement in Suwon, South Korea. The GA results, which identified optimal afforestation patches under varying area constraints, indicated that afforestation near existing green spaces significantly improved total NPP through enhanced landscape connectivity. The GA optimization identified afforesting 20 % of bare land (0.8 km²) as the most efficient scenario, yielding a projected increase of 354,385 kgC/yr, whereas expanding to 30 % (1.2 km²) increased sequestration to 530,036 kgC/yr but with diminishing efficiency. Spatial sensitivity analysis showed that optimal afforestation locations varied across scenarios, highlighting the need for flexible planning that adapts to spatial heterogeneity and land-use dynamics. This data-driven, interpretable framework supports more effective and adaptive urban green-infrastructure strategies for a carbon-neutral future.