🤖 AI Summary
To address the high training costs and limited generalization of large models in Earth observation (EO) data mining, this paper proposes a lightweight paradigm based on compositional pre-trained foundation models. Instead of training from scratch, our method fuses domain-specific remote sensing models (e.g., Prithvi) with general-purpose vision models (e.g., Hiera, DOFA) via a feature-level ensemble architecture, and transfers the ensemble knowledge to a compact student model through knowledge distillation. Evaluated across 11 multi-resolution, multi-sensor, and multi-task benchmarks in GEO-Bench, our approach matches or surpasses individual large models in performance while significantly reducing training time and computational overhead. The core contribution is the empirical validation that synergistic small-model ensembles outperform monolithic large models—establishing a scalable, cost-efficient, and highly generalizable pathway for EO AI.
📝 Abstract
Foundation models are rapidly transforming Earth Observation data mining by enabling generalizable and scalable solutions for key tasks such as scene classification and semantic segmentation. While most efforts in the geospatial domain have focused on developing large models trained from scratch using massive Earth Observation datasets, an alternative strategy that remains underexplored is the reuse and combination of existing pretrained models. In this study, we investigate whether foundation models pretrained on remote sensing and general vision datasets can be effectively combined to improve performance across a diverse set of key Earth Observation tasks. Using the GEO-Bench benchmark, we evaluate several prominent models, including Prithvi, Hiera, and DOFA, on eleven datasets covering a range of spatial resolutions, sensor modalities, and task types. The results show that feature-level ensembling of smaller pretrained models can match or exceed the performance of much larger models, while requiring less training time and computational resources. Moreover, the study highlights the potential of applying knowledge distillation to transfer the strengths of ensembles into more compact models, offering a practical path for deploying foundation models in real-world Earth Observation applications.