guest ::
| Home > Publications database > Coreset Selection with K-CREST: Evaluation of a Sustainable Approach on Remote Sensing Data |
| Home > Publications database > Coreset Selection with K-CREST: Evaluation of a Sustainable Approach on Remote Sensing Data |
| Journal Article | FZJ-2026-03983 |
; ; ; ; ; ;
2026
IEEE
New York, NY
This record in other databases:
Please use a persistent id in citations: doi:10.1109/ACCESS.2026.3719777 doi:10.34734/FZJ-2026-03983
Abstract: The environmental impact of training machine learning models is of critical concern as model sizes and dataset volumes continue to grow. Coreset selection offers a principled approach to reducing training-related carbon emissions by decreasing the amount of training data while maintaining model performance, directly addressing computational sustainability. In this paper, we present a comprehensive evaluation of coreset selection strategies for satellite remote sensing data.We evaluate multiple approaches, including random sampling, clustering-based methods such as K-means, and gradient-based selection. We also propose a hybrid approach that combines K-means clustering with gradient-based selection, which we dub K-CREST. We introduce a modular framework that enables systematic comparison of model performance, computational efficiency, and environmental impact across multiple node configurations. We experiment at fine-tuning scale as it provides a practical proving ground for coreset selection techniques, allowing us to develop and validate unsupervised selection strategies without the prohibitive computational costs of experimentation at pre-training scale. The label-free design of approaches like K-CREST positions them as promising candidates for self-supervised pre-training scenarios, to be explored in our future works. Results demonstrate that structured selection strategies achieve up to 97% reduction in carbon emissions while often maintaining around 80%-90% of full dataset performance, establishing coreset selection as a viable path toward sustainable machine learning in Earth observation applications. Notably, K-CREST achieves the highest efficiency score (6.00 Accuracy/Emissions at 1% data) by combining K-Means’ global structure awareness with CREST’s gradient-based informativeness. Looking ahead, K-CREST and parallel selection methods are considered strong contenders for future self-supervised pre-training scenarios in ML, aiming to further minimize environmental impacts.
; 
; 
; 
; Article Processing Charges ; Clarivate Analytics Master Journal List ; Current Contents - Electronics and Telecommunications Collection ; Current Contents - Engineering, Computing and Technology ; DOAJ Seal ; Essential Science Indicators ; Fees ; IF < 5 ; JCR ; SCOPUS ; Science Citation Index Expanded ; Web of Science Core Collection
|
The record appears in these collections: |
PDF