Spatiotemporal deep learning for national-scale land-use and land-cover projection using swin transformer and convLSTM

Accurate projection of land-use and land-cover (LULC) dynamics remains challenging due to spatial heterogeneity and path-dependent processes. This study proposes a spatiotemporal deep learning framework integrating a Swin Transformer spatial encoder with a ConvLSTM module to model cumulative land-use dynamics. Using Malawi as a case study, Landsat-derived LULC maps (2010–2022) and spatial drivers were used to predict the 2024 land-cover state and generate projections for 2029 and 2034. Model performance was evaluated using strict forward temporal validation and compared with multiple baselines. Results show that while no model consistently outperforms across all metrics, the Swin-ConvLSTM provides more balanced class-wise performance, particularly for transition-driven classes. Projections indicate urban expansion, forest decline, and moderate agricultural contraction, with changes concentrated in peri-urban areas. Entropy analysis reveals generally stable patterns, with higher uncertainty in transition zones and hotspots. These findings highlight the importance of explicit temporal modeling for robust LULC forecasting.