Core-Set Selection for Data-efficient Land Cover Segmentation

TL;DR

Large Earth Observation datasets suffer from redundancy, noise, and high computational costs. The authors propose six model-agnostic core-set selection methods (LC, FD, LC/FD, FA, CB, FA/CB) to identify informative subsets for land-cover segmentation and benchmark them on three high-resolution datasets using two architectures. They show core-sets can match or exceed full-dataset performance while reducing training data and time, with label-diversity playing a key role in gains. The study underscores the potential of data-centric learning in remote sensing and suggests future integration with self-supervised and foundation-model approaches.

Abstract

The increasing accessibility of remotely sensed data and their potential to support large-scale decision-making have driven the development of deep learning models for many Earth Observation tasks. Traditionally, such models rely on large datasets. However, the common assumption that larger training datasets lead to better performance tends to overlook issues related to data redundancy, noise, and the computational cost of processing massive datasets. Effective solutions must therefore consider not only the quantity but also the quality of data. Towards this, in this paper, we introduce six basic core-set selection approaches -- that rely on imagery only, labels only, or a combination of both -- and investigate whether they can identify high-quality subsets of data capable of maintaining -- or even surpassing -- the performance achieved when using full datasets for remote sensing semantic segmentation. We benchmark such approaches against two traditional baselines on three widely used land-cover classification datasets (DFC2022, Vaihingen, and Potsdam) using two different architectures (SegFormer and U-Net), thus establishing a general baseline for future works. Our experiments show that all proposed methods consistently outperform the baselines across multiple subset sizes, with some approaches even selecting core sets that surpass training on all available data. Notably, on DFC2022, a selected subset comprising only 25% of the training data yields slightly higher SegFormer performance than training with the entire dataset. This result shows the importance and potential of data-centric learning for the remote sensing domain. The code is available at https://github.com/keillernogueira/data-centric-rs-classification/.

Figures (6)

• Figure 1: General overview of core-set selection. An input data set is first processed by a core-set selection algorithm that, based on some criteria, prioritizes certain examples over others (represented by the size of the blue circles). Based on this, it is possible to select the core-set data depending on the amount of data one would like to retain (illustrated by red, orange, and green dotted circles). Finally, the selected core-set is used to train a machine learning model, thus reducing the training time while maintaining, or even improving, task performance.
• Figure 2: Example images (first row) of the DFC2022 dataset hansch20222022, their DEM data (second row), and the respective reference data (third row).
• Figure 3: Example images (first row) of the Vaihingen and Potsdam datasets isprs, their DSM data (second row), and the respective reference data (third row).
• Figure 4: Visualizations of the proposed methods' rankings. The line represents the average rank position for each patch across all proposed approaches, while the shaded area represents the standard deviation.
• Figure 5: Correlation of methods according to Kendall Tau coefficient. A high correlation value means that the methods produce similar rankings.