Dual-carbon goal oriented optimization of sustainable land use based on ST-GTNet model

Accurate estimation of carbon emissions is crucial for supporting China's strategic Dual-Carbon goals, which aim to reach a carbon peak by 2030 and achieve carbon neutrality by 2060, in line with the Sustainable Development Goals (SDGs). However, conventional estimation methods (land use coefficient-based and energy consumption-based approaches) significantly underestimate the potential of land use optimization in contributing to carbon neutrality. To address these limitations, a Land Use Carbon Emission Accounting and Spatial Pattern Assessment System (LUC-SPAS) framework is proposed. Firstly, the Spatio-Temporal GRU-Transformer Network (ST-GTNet) model was established, which couples historical carbon emission data and employs deep learning techniques within the neural network to establish a quantitative relationship between the land use pattern index and carbon emissions. Secondly, the Patch-based land use Simulation (PLUS) model projected land use from 2030 to 2060 under three scenarios: the Business-as-Usual Scenario (BAUS), the Crop Land Guarantee Scenario (CLGS), and the Sustainable Low-Carbon Development Scenario (LCDS). Finally, based on the trained ST-GTNet model and predicted land use pattern, future carbon emissions under three scenarios were predicted. Specifically, the ST-GTNet model achieves high accuracy (R 2 = 0.89) in carbon-emission forecasting by harmonizing the temporal memory of GRU and the global spatial attention of Transformer. The study found that, under LCDS, through ecological restoration and compact urban growth, a 9.2% reduction in regional carbon emissions by 2060 compared with BAUS is achieved. The Shanghai-Nanjing-Hangzhou corridor is identified as a key decarbonization hotspot. This study presents a replicable GeoAI framework for predicting carbon emissions in urban agglomerations. The ST-GTNet model establishes a new paradigm for spatio-temporal GeoAI, bridging mechanistic knowledge and data-driven learning.