Prune Before You Adapt: Efficient Earth Observation Foundation Models for Climate Monitoring

Earth observation foundation models (EOFMs) enable broad reuse across climate-monitoring tasks, but their scale makes repeated downstream adaptation and deployment costly. We present MUSE, the first unified framework for pre-adaptation structured pruning of EOFMs, designed to reduce this cost before task-specific fine-tuning begins. MUSE performs dependency-correct physical pruning of attention heads and MLP units, producing compact dense encoders that remain compatible with standard training and inference kernels. To preserve task-relevant capacity, it further introduces output-aware progressive pruning, repeatedly updating structural importance as the compressed model recovers rather than relying on a fixed one-shot ranking. We evaluate MUSE across four heterogeneous EOFMs and four dense EO tasks spanning burn-scar segmentation, land-cover mapping, and biomass estimation. At 50\% sparsity, MUSE achieves the state-of-the-art performance across all tasks, with several compact models matching or even exceeding their dense counterparts. These results show that substantial EOFM redundancy can be removed before adaptation, improving both training- and deployment-efficiency while preserving climate-monitoring utility.