PaperlayerView exact revision on arXiv

Source-linked AI summary

Fidelity-Diversity-Consistency (FDC): Data Pruning for Remote Sensing Change Detection

Dongyao Zhu, Ranga Raju Vatsavai

arXiv:2608.21754v1cs.CV

TL;DR

Remote-sensing change-detection pruning lacks reliable evidence that existing methods improve on random selection, despite pruning’s success in related tasks. The paper benchmarks six methods, analyzes 540 random coresets, and proposes FDC, which consistently improves selection, especially at extreme pruning ratios. Its central finding is that preserving change-distribution fidelity is crucial, with diversity and consistency providing complementary signals.

  • Problem

    Remote-sensing change detection pruning must address extreme class imbalance and spatial redundancy, but existing data-pruning literature does not explicitly preserve change distribution fidelity.

  • Method

    The paper benchmarks existing criteria, regresses performance on four coreset descriptors, and develops two-stage FDC with fidelity as a hard constraint and diversity and consistency as soft objectives.

  • Results

    Existing methods do not reliably outperform random selection, whereas FDC-P and FDC-G consistently outperform random selection, particularly at extreme pruning ratios.

  • Takeaways & Limitations

    Coreset selection for remote-sensing change detection should preserve distributional properties governing change-learning dynamics rather than prioritize only difficult samples.

  • Takeaways & Limitations

    The study covers binary change detection on two benchmarks; broader datasets, alternative training paradigms, and learned descriptor weights remain future work.

Abstract

from arXiv · show

Despite the success of data pruning (DP) in reducing training data sizes and improving downstream model performance in classification and segmentation tasks, its potential in remote sensing change detection remains unexplored. For the first time, we benchmark six representative DP methods across building- and forest-change datasets, CNN- and transformer-based models, and three pruning budgets, and show that existing baselines yield no reliable advantage over random selection. Notably, even the strongest evaluated baseline, Feature Diversity, is matched or exceeded by $\sim$33\% of randomly sampled subsets. To understand the underlying mechanism, we conduct a systematic regression study over 540 randomly sampled data subsets, characterizing each with four descriptors covering label statistics, image diversity, and feature-space geometry. Random Forest models show that \emph{change distribution fidelity} is the most prominent factor in determining the quality of change detection data subsets, a property absent from the existing pruning literature. Our analyses further show that pixel-wise image diversity and label-feature consistency are secondary factors. We translate these findings into Fidelity-Diversity-Consistency (FDC), a simple two-stage pruning method that shows consistent improvements over existing baselines across change detection benchmarks and backbones, especially at lower pruning ratios. Code is available at \href{https://github.com/ddydyd32/fidelity-diversity-consistency}{https://github.com/ddydyd32/fidelity-diversity-consistency}.

I. Introduction

Remote-sensing change detection combines costly, noisy supervision with task-specific distributional challenges that existing data-pruning methods have not systematically addressed. This work argues that useful coresets must preserve change distributions alongside diversity and label-feature consistency.

  • Data collection scales poorly because marginal examples contribute less while processing costs remain constant.
  • Data pruning selects smaller, higher-quality coresets to reduce training time, energy, and storage while maintaining comparable accuracy.
  • Remote-sensing change detection often relies on noisy ancillary labels with systematic errors, generalized boundaries, and temporal mismatches.
  • Change detection pruning must preserve scarce changed samples and the dataset’s change distribution amid severe class imbalance and spatial redundancy.
  • Across benchmarks and backbones, existing pruning methods do not reliably outperform random selection, motivating analysis of which coreset properties matter.
  • Regression over 540 random coresets identifies change distribution fidelity as indispensable, while diversity and label-feature consistency provide additional signals for two-stage selection.

II. Related Work

Data pruning extends coreset selection to deep learning and has been studied in classification, segmentation, and self-supervised remote sensing, but not systematically in remote-sensing change detection.

  • Data pruning retains informative samples from large datasets to reduce computational costs without substantially sacrificing model performance.
  • Remote-sensing pruning requires domain-specific adaptations because satellite data involve high spectral dimensionality, class imbalance, top-down viewpoints, and noisy annotations.
  • Existing remote-sensing pruning work targets self-supervised learning and semantic segmentation, while remote-sensing change detection remains largely unexplored.

III. Existing Pruning Criteria Fail to Consistently

The study evaluates representative pruning criteria across change-detection datasets, model architectures, and budgets to test whether they outperform random selection.

  • Six pruning methods are evaluated across two change-detection datasets, two backbones, and three pruning budgets.
  • The benchmarks comprise LEVIR-CD building changes and REDD-Forest-CD forest changes, with 1054 REDD examples retained from Bolivia and Brazil for label consistency.
  • Training uses 1%, 5%, and 10% subsets, corresponding to approximately 5, 23, and 46 LEVIR-CD samples.
  • TinyCD and BIT represent lightweight CNN-based and CNN–Transformer-based change-detection architectures.
  • Performance is reported using F1, mean Intersection over Union, and Cohen’s κ, averaged over three runs.

C. Baseline Data Pruning Methods

The benchmark spans error-, image-, and label-based pruning criteria, using common image embeddings for methods that require feature representations.

  • The baselines comprise EL2N and D2 error-based methods, FA, FD, and K-Center image-based methods, and LC label-based pruning.
  • All embedding-based baselines use the same frozen ImageNet-1k-pretrained ResNet-18 encoder to isolate differences between pruning methods.
  • D2 combines change-weighted difficulty with graph-based diversity, using k-nearest-neighbor message passing and difference-image feature distances.
  • Feature Activation scores examples from mean and standard deviation statistics of pretrained image embeddings, with higher scores indicating greater diversity and less noise.
  • Feature Density clusters embeddings with K-Means and selects centroid-near samples across clusters, whereas K-Center greedily maximizes minimum feature-space distance.
  • Label Complexity ranks examples by change-mask entropy, with higher scores indicating more uniform pixel-class distributions.

D. Benchmark Results

Existing pruning baselines provide no reliable advantage over random selection in remote sensing change detection. Their inconsistent behavior reflects mismatches between the selected examples and change-detection supervision.

  • Existing baselines yield no reliable advantage over random selection across the reported change-detection benchmarks and backbones.
  • EL2N and D2 select minimally changing or ambiguous examples, while LC selects a narrow band of frequent change ratios and can destroy distributional coverage.
  • FD reaches the highest baseline score at 5% on LEVIR-CD, with F1 64.5 versus random 62.6.
  • At the 1% budget, K-Center surpasses random on both datasets, reaching F1 44.2 on LEVIR-CD and 51.8 on REDD-Forest-CD versus 23.8 for random.
  • The findings motivate examining intrinsic properties of random subsets that drive change-detection performance.

IV. What Makes a Good Change Detection Coreset?

The study characterizes random coresets with four training-free descriptors spanning change labels, image diversity, and feature-space geometry. These descriptors quantify fidelity to the full dataset and different forms of subset diversity and consistency.

  • The regression study samples random coresets at 5% and 1% ratios and records Tiny-CD mIoU after training each subset for 100 epochs across three seeds.
  • Four descriptors are computed directly from raw images, labels, and model features without training a change-detection model.
  • L1 Change Distribution Fidelity measures deviation between the selected subset’s average change ratio and the full dataset’s change ratio.
  • I1 Intra-Set Image Diversity uses mean pairwise pixel-space distance, with larger values indicating lower visual redundancy.
  • F1 measures feature-space diversity, while F2 measures the correlation between feature similarity and labeled change ratio.

C. Analysis on coreset descriptors vs. model performance

Regression analyses identify change distribution fidelity as the dominant predictor of change-detection coreset quality, with image and feature diversity providing dataset-dependent complementary signals. The relationship is strongest when subsets preserve the full dataset’s change-ratio distribution.

  • Two complementary analyses—univariate linear regression and random-forest permutation importance—are applied at both pruning ratios and on both datasets.
  • Change distribution fidelity is the single most important predictor across datasets and pruning regimes, while I1 provides a dataset-dependent complement.
  • Distribution-dominated regime: On REDD-Forest-CD, L1 accounts for R2 = 0.889 at 5% and 0.823 at 1%, dominating RF importance in both regimes.
  • Higher change distribution fidelity correlates with better mIoU, and at lower pruning ratios fidelity becomes harder to satisfy while diversity gains importance on heterogeneous datasets.

V. Method: Fidelity, Diversity, Consistency

FDC operationalizes the descriptor findings through a two-stage procedure that starts from a random candidate and iteratively improves its quality. Its weighting reflects whether a dataset is diversity-sensitive or distribution-dominated.

  • FDC initializes a candidate pool from competitive random selection and refines it using descriptor-based scoring.
  • FDC-P oversamples the target pool and removes low-quality samples, whereas FDC-G starts smaller and adds high-quality samples.
  • At 10% and 5% pruning FDC-P adds and removes pn samples, while at 1% FDC-G adds gn samples after starting gn below target size.
  • Both variants combine I1, L1, F1, and F2 after normalization, with marginal utility recomputed after each candidate addition or removal.
  • LEVIR-CD uses α = 1.0 and β = 1.0 for diversity and fidelity, while REDD-Forest-CD prioritizes fidelity with α = 0.1 and β = 1.0.

VI. Experiments

Across Tiny-CD and BIT benchmarks, FDC is competitive with or improves over existing pruning baselines and remains consistently stable relative to random selection, particularly at low budgets.

  • FDC-P improves LEVIR-CD performance at 10% pruning, raising F1 from 68.1 to 68.6 and mIoU from 74.2 to 74.5.
  • At 5% pruning, FDC-P achieves the best reported performance on LEVIR-CD and REDD-Forest-CD, including F1 56.4 and mIoU 65.9 on REDD-Forest-CD.
  • FDC-G achieves the best mIoU on LEVIR-CD and the best overall performance on REDD-Forest-CD with Tiny-CD.
  • On BIT, FDC-G recovers random-selection performance from F1 13.9 to 35.7 on LEVIR-CD and from 6.3 to 44.6 on REDD-Forest-CD.
  • At 1% pruning, FDC-G’s fidelity-aware scoring helps prevent failures caused by omitting scarce change-class examples.

B. Ablation: Component Contributions

Ablations show that descriptor importance depends on dataset and pruning budget, while sensitivity experiments indicate that preserving random structure improves performance stability.

  • Component Contributions: At 10% pruning, removing I1 causes the largest F1 drops on LEVIR-CD and REDD-Forest-CD: −2.0 and −3.1, respectively.
  • Component Contributions: At 10% pruning, F2 is the next most impactful component, with removal reducing F1 by 2.2 on LEVIR-CD and 1.0 on REDD-Forest-CD.
  • Component Contributions: At 1% pruning on LEVIR-CD, removing L1 raises mean F1 from 41.1 to 43.9 but substantially increases variance from ±2.6 to ±8.1.
  • Component Contributions: At 1% pruning on REDD-Forest-CD, L1 is indispensable: removing it lowers F1 from 53.2 to 49.9, while removing I1 has no effect.
  • Sensitivity Analysis: Varying p and g shows that preserving more of the random structure can improve performance at moderate pruning and reduce performance variance during sample addition.

VII. Limitations and Future Work

The study evaluates RS-CD coreset selection across two benchmarks and finds that existing methods do not reliably beat random selection. Its analysis motivates FDC, while broader dataset validation, alternative training paradigms, and automated descriptor weighting remain future directions.

  • Future work: Validation across a broader range of datasets remains necessary for the proposed taxonomy of diversity-limited and distribution-shift settings.The paper also identifies semi-supervised and self-supervised pruning, plus learned descriptor weights, as future directions.
  • Evaluation scope: Two RS-CD benchmarks and two backbones reveal no existing pruning method that reliably outperforms random selection.The failures are described as structural rather than architecture-dependent.
  • Descriptor analysis: Descriptor priorities are dataset-dependent: intra-set diversity dominates on LEVIR-CD, whereas change distribution fidelity is near-sufficient on REDD-Forest-CD.At 1% budgets, label-feature consistency provides a complementary discriminating signal.
  • Proposed method: Fidelity-Diversity-Consistency enforces change distribution fidelity as a hard constraint while treating diversity and consistency as soft objectives.FDC-P and FDC-G consistently outperform random selection, particularly at extreme pruning ratios.
  • Implication: The study reframes RS-CD coreset selection around preserving distributional properties that govern change-learning dynamics rather than identifying the hardest samples.This conclusion follows from the observed limitations of existing pruning criteria and the descriptor analysis.
Loading 2608.21754v1…