MoEST-Net: mixture-of-experts spatio-temporal transformer network for long-term land-use change detection in high-resolution remote sensing imagery

Abstract

Long-term land-use change detection requires models that distinguish persistent land-cover transitions from seasonal appearance changes while remaining efficient over extended image sequences. MoEST-Net is a Mixture-of-Experts spatio-temporal Transformer designed for this setting. Its temporal-scale expert grid uses hierarchical routing to select a small subset of experts for sequences containing 3–8 observations, while a seasonal-disentanglement branch models vegetation phenology and an adaptive gate suppresses season-related responses in the change representation. A change-aware load-balancing objective further adjusts expert use to the spatial sparsity of changed regions. Evaluation on six change-detection benchmarks yields F1 scores of 93.52% on LEVIR-CD and 96.18% on WHU-CD, with larger gains on datasets containing stronger seasonal and scene variation. On long-term sequences, gradual-change recall increases from 62.35% for the bi-temporal repeat baseline to 85.42%. The model contains 22.4 M parameters and requires 18.5 GFLOPs, providing a practical balance between detection accuracy and active computation for long-term land-use monitoring.