Urban land use change, particularly in rapidly developing or policy-driven cities, exerts a critical influence on carbon emission fluctuations. Yet, existing forecasting models often rely on static inventories or statistical regressions, lacking the capacity to capture temporal and spatial dynamics or adapt across different regions. This study addresses this research gap by developing a deep learning framework that combines a Time-Series Transformer architecture with transfer learning to predict monthly carbon emissions using remote sensing and emission inventory data. The model integrates Normalised Difference Vegetation Index (NDVI) derived from Sentinel-2 satellite imagery and city-scale CO2 emissions from the Carbon Monitor dataset. Two cities, Beijing and London, are selected to represent different urbanisation contexts and policy environments. The model is pretrained on Beijing data and fine-tuned on London to assess cross-city transferability. The results show that the model achieves high predictive accuracy, with the Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) indicating robust performance across cities. Attention-based interpretation reveals key spatial–temporal patterns in land use–emission coupling. This work demonstrates how fusing satellite-based ecological indicators with advanced sequence models can yield interpretable and generalisable tools for urban carbon emission forecasting. The proposed framework has strong potential to support policy-making in low-carbon urban development, enabling data-driven strategies for climate mitigation, green infrastructure planning, and sustainable land management.