High-Resolution Land-Use Reconstruction Using Swin Transformers

Uncertainty in the terrestrial carbon cycle remains a major constraint in climate projections, partly driven by imperfect land surface representations and inter-model spread across Earth system models. To address this limitation, we propose a Swin Transformer framework for high-resolution reconstruction of land use, integrating coarse-resolution land-use data with climate classification and geophysical features to produce annual land-use maps at 1 km resolution. We compare the Swin backbone against persistence, linear, and convolutional baselines, and evaluate autoregressive and non-autoregressive variants. The non-autoregressive Swin model gives the best overall reconstruction, while the autoregressive variant is marginally better on the pixels that change from year to year but degrades rapidly once rolled out on its own predictions. This work lays the methodological foundation for a forthcoming temporally continuous land-use product, with the aim of better constraining land-surface-driven uncertainty and improving the predictive power of next-generation climate simulations.