SC-LUFormer: A Scenario-Conditioned Transformer for Long-Term Land-Use Prediction from Remote Sensing Data in Mountainous Urban Agglomerations

Long-term land-use prediction based on remote sensing-derived data remains challenging in mountainous urban agglomerations because fragmented landscapes, complex terrain constraints, and scenario-dependent differences jointly affect spatial allocation and future projection reliability. To address these challenges, this study proposes a unified scenario-conditioned Transformer framework (SC-LUFormer) for long-term land-use prediction in mountainous urban agglomerations. Rather than introducing an entirely new neural primitive, the proposed framework integrates non-local spatial representation, scenario-aware feature modulation, and change-oriented evaluation into a single modeling framework, in which future socio-environmental scenarios are incorporated into the deep feature space via feature-wise linear modulation (FiLM). Multi-temporal land-use maps derived from remote sensing observations, together with multi-source biophysical and socioeconomic drivers, were used to validate the model through a 2020 backcasting experiment and to generate scenario-based land-use projections for 2040–2100 under multiple CMIP6-based Shared Socioeconomic Pathway (SSP) scenarios in the Central Yunnan Urban Agglomeration, China. Comparative experiments against representative convolutional neural network, machine-learning, and rule-based models show that SC-LUFormer achieves better backcasting performance, with an overall accuracy of 94.88%, a mean intersection-over-union of 85.71%, an area deviation of 3.17%, and a figure of merit of 0.2917 in reconstructing the year 2020. Further interpretability analysis using Grad-CAM and SHAP provides post-hoc indications that SC-LUFormer focuses on urban expansion boundaries and transition zones and produces scenario-responsive driver-attribution patterns under alternative future assumptions.