RS-Prune: Training-Free Data Pruning at High Ratios for Efficient Remote Sensing Diffusion Foundation Models

TL;DR

This work tackles the data bottleneck in remote sensing diffusion foundation models by introducing a training-free, two-stage pruning strategy that rapidly yields a high-quality subset under high pruning ratios. Stage I uses grayscale entropy $H(I)$ to prune low-information images, while Stage II employs reference-guided, scene-aware clustering with centroid-prioritized sampling based on cosine similarity $s_k(x)$ to preserve diversity and representativeness. The method relies on a fixed reference bank of RS scene datasets to construct centroids via $K$-means on unit-normalized features $f(x)/\|f(x)\|_2$, enabling efficient assignment and sampling without full-dataset clustering or model training. Empirical results show significant convergence and generation quality gains, with FID improvements (e.g., down to $28.45$ on Git-10M) and robust downstream performance on super-resolution and semantic image synthesis, validating its practical impact for RS foundation-model development. By reducing data requirements and computation without sacrificing representativeness, this approach provides a scalable blueprint for building RS generative foundation models and guiding future data-centric RS research.

Abstract

Diffusion-based remote sensing (RS) generative foundation models are cruial for downstream tasks. However, these models rely on large amounts of globally representative data, which often contain redundancy, noise, and class imbalance, reducing training efficiency and preventing convergence. Existing RS diffusion foundation models typically aggregate multiple classification datasets or apply simplistic deduplication, overlooking the distributional requirements of generation modeling and the heterogeneity of RS imagery. To address these limitations, we propose a training-free, two-stage data pruning approach that quickly select a high-quality subset under high pruning ratios, enabling a preliminary foundation model to converge rapidly and serve as a versatile backbone for generation, downstream fine-tuning, and other applications. Our method jointly considers local information content with global scene-level diversity and representativeness. First, an entropy-based criterion efficiently removes low-information samples. Next, leveraging RS scene classification datasets as reference benchmarks, we perform scene-aware clustering with stratified sampling to improve clustering effectiveness while reducing computational costs on large-scale unlabeled data. Finally, by balancing cluster-level uniformity and sample representativeness, the method enables fine-grained selection under high pruning ratios while preserving overall diversity and representativeness. Experiments show that, even after pruning 85\% of the training data, our method significantly improves convergence and generation quality. Furthermore, diffusion foundation models trained with our method consistently achieve state-of-the-art performance across downstream tasks, including super-resolution and semantic image synthesis. This data pruning paradigm offers practical guidance for developing RS generative foundation models.

Figures (4)

• Figure 1: Illustration of our reference‐guided clustering and efficiency gains. (a) Direct clustering on the full unlabeled dataset is computationally expensive. (b) Our reference‐guided strategy selects a representative, scene-aware subset around reference centroids. (c) Compared with full-dataset clustering, our method achieves much lower FID at substantially reduced cost time. And compared with best baseline, ours requires 1.9× to 2.1× fewer training iterations across different pruning ratios.
• Figure 2: Overview of our multi-stage data pruning method for Remote Sensing generative foundation models.
• Figure 3: Comparison of generation performance (FID) across different pruning ratios on Git-10M.
• Figure 4: Results of entropy-based pruning at different pruning ratios on GiT-10M, reporting FID for the generation task.