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

Dongyao Zhu; Ranga Raju Vatsavai

cs.CVAug 22, 2026

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

Authors: Dongyao Zhu, Ranga Raju Vatsavai

Organizations: Department of Computer Science North Carolina State University Raleigh, North Carolina, USA

Abstract

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 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 https://github.com/ddydyd32/fidelity-diversity-consistency.

Figures & tables

Fig. 1 : Higher image diversity correlates with better mIoU, still

\sim

33% of random coresets are competitive with diversity-driven DP methods.

Fig. 2 : The first CD examples chosen by different pruning methods. Loss-based EL2N and $\mathbb{D}^{2}$ picks examples with minimal or ambiguous changes. Label-based LC picks an example with substantial changes. Diversity-based K-Center and FD pick first examples similar to random selection.

Dataset	Desc.	5% Pruning		1% Pruning
Dataset	Desc.	$R^{2}$	RF	$R^{2}$	RF
LEVIR-CD	I1	0.654	0.479	0.077	0.069
	L1	0.413	0.750	0.000	1.727
	F1	0.441	0.012	0.027	0.072
	F2	0.119	0.003	0.036	0.063
REDD-Forest-CD	I1	0.711	0.005	0.528	0.023

TABLE I : Univariate

R^{2}

and random forest (RF) permutation importance of descriptors on random subsets vs. mIoU.

Fig. 3 : Higher change distribution fidelity (L1) correlates to better mIoU.

		LEVIR-CD			REDD-Forest-CD
Ratio	Pruner	F1	mIoU	$\kappa$	F1	mIoU	$\kappa$
100%	Full	76.9 ±0.7	80.0 ±0.5	0.76 ±0.01	58.3 ±1.4	67.2 ±0.7	0.55 ±0.01
10%	Random	68.1 ±0.8	74.2 ±0.4	0.67 ±0.01	58.8 ±1.3	66.9 ±0.7	0.55 ±0.01
	EL2N	7.2 ±10.1	49.5 ±2.9	0.07 ±0.09	0.9 ±0.6	45.2 ±0.2	0.01 ±0.01
	$\mathbb{D}^{2}$	24.6 ±18.5	55.3 ±6.1	0.23 ±0.18	55.9 ±1.3	65.6 ±0.6	0.52 ±0.01
	FA	65.4 ±1.6	72.2 ±1.1	0.63 ±0.02	52.4 ±1.3	64.0 ±0.6	0.49 ±0.01

TABLE II : Benchmark results (%, except for

\kappa

) with Tiny-CD on LEVIR-CD and REDD-Forest-CD . Best results per pruning ratio and dataset are in bold . FDC is competitive with all baselines and consistently stable relative to random selection.

		LEVIR-CD			REDD-Forest-CD
Ratio	Pruner	F1	mIoU	$\kappa$	F1	mIoU	$\kappa$
100%	Full	72.0 ±0.7	76.5 ±0.4	0.70 ±0.01	61.1 ±0.5	68.2 ±0.3	0.57 ±0.00
10%	Random	61.8 ±1.6	70.1 ±0.9	0.60 ±0.02	50.6 ±1.0	62.0 ±0.6	0.45 ±0.01
	EL2N	0.0 ±0.0	47.4 ±0.0	0.00 ±0.00	0.0 ±0.0	45.0 ±0.0	0.00 ±0.00
	$\mathbb{D}^{2}$	28.4 ±20.1	55.8 ±5.9	0.26 ±0.18	51.9 ±1.3	62.3 ±0.9	0.46 ±0.02
	FA	56.6 ±5.7	67.5 ±2.8	0.54 ±0.06	49.7 ±2.7	60.5 ±1.8	0.43 ±0.03

TABLE III : Benchmark results (%, except for

\kappa

) of BIT backbone on LEVIR-CD and REDD-Forest-CD . Best results per pruning ratio and dataset are in bold . FDC is competitive with all baselines and consistently stable against random selection.

		LEVIR-CD			REDD-Forest-CD
Ratio	Pruner	F1	mIoU	$\kappa$	F1	mIoU	$\kappa$
10%	Full	68.6 ±1.3	74.5 ±0.8	0.67 ±0.01	56.0 ±1.5	65.8 ±0.8	0.53 ±0.01
	- I1	66.6 ±0.6	73.3 ±0.3	0.65 ±0.01	52.9 ±1.9	64.2 ±0.9	0.49 ±0.02
	- L1	67.2 ±2.2	73.7 ±1.2	0.66 ±0.02	55.5 ±1.6	65.5 ±0.8	0.52 ±0.02
	- F1	68.6 ±0.9	74.4 ±0.6	0.67 ±0.01	54.9 ±2.1	65.3 ±1.1	0.51 ±0.02
	- F2	66.4 ±1.4	73.2 ±0.8	0.65 ±0.01	55.0 ±3.7	65.4 ±1.9	0.52 ±0.04

TABLE IV : Ablation results (%, except for

\kappa

) of removing each descriptor from sample scoring. CD Backbone: Tiny-CD.

Panel of Figure 8 — Fig. 4 : Sensitivity of FDC -P to $p$ and FDC -G to $g$ .

Dataset Pruning

Dataset pruning selects compact, informative training subsets to reduce computation while preserving or improving model performance.

See all 22 papers

ChangeFlow -- Latent Rectified Flow for Change Detection in Remote Sensing
May 14, 2026Blaž Rolih, Matic Fučka, Filip Wolf +1Remote Sensing Change Detection
University of Ljubljana, Faculty of Computer and Information Science, Slovenia
CoRegOVCD: Consistency-Regularized Open-Vocabulary Change Detection
Apr 2, 2026Weidong Tang, Hanbin Sun, Zihan Li +2Remote Sensing Change Detection Semantic Change Detection
China Agricultural University, College of Information and Electrical Engineering, Beijing, China · China Agricultural University, College of Science, Beijing, China
Make Some Noise: Unsupervised Remote Sensing Change Detection Using Latent Space Perturbations
Feb 23, 2026Blaž Rolih, Matic Fučka, Filip Wolf +1Synthetic Data Augmentation Remote Sensing Change Detection
University of Ljubljana, Faculty of Computer and Information Science, Slovenia

May 14, 2026J/K move · Enter open · S save

ChangeFlow -- Latent Rectified Flow for Change Detection in Remote Sensing

Blaž Rolih, Matic Fučka, Filip Wolf, Luka Čehovin Zajc

University of Ljubljana, Faculty of Computer and Information Science, Slovenia

Remote Sensing Change Detection

Open paper AskPDF arXiv

Remote sensing change detection (RSCD) aims to localise changes between two images of the same geographic region. In practice, change masks often follow region-level annotation conventions rather than purely local appearance differences, making them context-dependent and occasionally ambiguous. Most state-of-the-art methods utilise per-pixel discriminative classification, which produces a single prediction per input and fails to explicitly model the changed region as a coherent whole. A natural alternative is generative formulation, which can model a distribution of plausible masks, enabling sampling to capture ambiguity and encourage global consistency. However, existing generative RSCD approaches typically lag behind strong discriminative baselines due to the high computational cost of pixel-space generation and the complexity of their conditioning mechanisms. To address the limitations of prior discriminative and generative methods, we propose ChangeFlow, a generative framework that reformulates change detection as the synthesis of a change mask in latent space via rectified flow. ChangeFlow is guided by a structured yet lightweight conditioning signal, and its stochastic design naturally supports sampling-based prediction ensembling. Namely, aggregating multiple predicted change masks improves robustness, while sample agreement provides a practical confidence estimation that highlights ambiguous regions. Across four benchmarks, ChangeFlow achieves an average F1 of 80.4%, improving by 1.3 points on average over the previous best method, while maintaining inference speed comparable to recent strong baselines. Project page: https://blaz-r.github.io/changeflow_cd

Abstract

Dataset Pruning

Explore similar work

ChangeFlow -- Latent Rectified Flow for Change Detection in Remote Sensing

CoRegOVCD: Consistency-Regularized Open-Vocabulary Change Detection

Make Some Noise: Unsupervised Remote Sensing Change Detection Using Latent Space Perturbations