The impact of climate change on land use is particularly complex in fragmented land use regions, and accurate simulation of land use spatial patterns in such regions is crucial for national spatial planning. This study, using Kunming as the research area, addresses the insufficient adaptability of traditional models by constructing a conditional GAN land use prediction model, SFiLM-GAN, based on FiLM feature modulation. This model uses U-Net as the encoding and decoding backbone, embeds FiLM to implement scenario constraints, and combines a multi-scale PatchGAN discriminator to enhance the rationality of spatial patterns. The model's results are validated, and predictions are made using the CMIP6 multi-carbon emission-driven climate scenario. Results show that the SFiLM-GAN model achieves OA, Kappa, and mIoU of 94.25%, 91.58%, and 88.32%, respectively, with an AD of only 4.56%. It demonstrates accurate land use identification and strong ability to distinguish land uses with overlapping landscape features. From 2000 to 2020, forest land in Kunming remained stable, urban land expanded by a cumulative 3.37%, cultivated land and grassland decreased by 1.32% and 1.58%, respectively, water areas increased slightly, and unused land remained stable. Under different carbon emission scenarios in 2030, urban land expansion intensity is the core differentiator in land use patterns. Scenario SSP585 shows an expansion of 2.85%, while scenario SSP119 shows the mildest expansion of 1.73%. Cultivated land and grassland are preserved more intact, while other land types show no significant differences across scenarios. The landscape pattern in the study area continues to fragment, with PD increasing from 0.33 to 0.38, reaching 0.41 for SSP585 in 2030. LSI is rising, while AI and CONNECT are declining. Low-carbon development scenarios can mitigate this deterioration. The model constructed in this study can provide scientific support and decision-making basis for accurate land use simulation in complex geomorphological areas, as well as regional spatial planning and sustainable development. • Deep learning GAN networks are successfully used for land use simulation in fragmented Kunming areas. • The proposed SFiLM-GAN model effectively solves the challenge of simulating fragmented land use. • SFiLM-GAN significantly outperforms four comparative models, including U-Net and PLUS. • SFiLM-GAN integrates multiple scenarios from CMIP6, providing support for urban climate change mitigation.
Wang et al. (Fri,) studied this question.