NeXt2Former-CD: Efficient Remote Sensing Change Detection with Modern Vision Architectures

Yufan Wang; Sokratis Makrogiannis; Chandra Kambhamettu

cs.CVFeb 21, 2026

NeXt2Former-CD: Efficient Remote Sensing Change Detection with Modern Vision Architectures

Authors: Yufan Wang, Sokratis Makrogiannis, Chandra Kambhamettu

Organizations: University of South Florida · Delaware State University

Abstract

State Space Models (SSMs) have recently gained traction in remote sensing change detection (CD) for their favorable scaling properties. In this paper, we explore the potential of modern convolutional and attention-based architectures as a competitive alternative. We propose NeXt2Former-CD, an end-to-end framework that integrates a Siamese ConvNeXt encoder initialized with DINOv3 weights, a deformable attention-based temporal fusion module, and a Mask2Former decoder. This design is intended to better tolerate residual co-registration noise and small object-level spatial shifts, as well as semantic ambiguity in bi-temporal imagery. Experiments on LEVIR-CD, WHU-CD, and CDD datasets show that our method achieves the best results among the evaluated methods, improving over recent Mamba-based baselines in both F1 score and IoU. Furthermore, despite a larger parameter count, our model maintains inference latency comparable to SSM-based approaches, suggesting it is practical for high-resolution change detection tasks.

Figures & tables

Figure 1: Overview of the proposed change detection network. It consists of a weight-sharing DINOv3 [ 15 ] backbone, Feature Rectify Modules (FRM) and Feature Fusion Modules (FFM) at multiple scales, and a Mask2Former head for final prediction. Query-level class logits and masks are aggregated into dense predictions via log-sum-exp over queries.

CNN-based Methods:
		WHU-CD [ 10 ]			LEVIR-CD [ 4 ]			CDD [ 11 ]
Method	Extra training data	F1 ( $\uparrow$ )	IoU ( $\uparrow$ )	OA ( $\uparrow$ )	F1 ( $\uparrow$ )	IoU ( $\uparrow$ )	OA ( $\uparrow$ )	F1 ( $\uparrow$ )	IoU ( $\uparrow$ )	OA ( $\uparrow$ )
SNUNet [ 8 ]	None	0.835	0.717	98.7	0.882	0.788	98.8	0.839	0.721	96.2
IFNet [ 19 ]	ImageNet 1k	0.834	0.715	98.8	0.881	0.788	98.9	0.840	0.719	96.0
CNN + Attention based Methods:
DT-SCN [ 12 ]	ImageNet 1k	0.914	0.842	99.3	0.877	0.781	98.8	0.921	0.853	98.2

Table 1: Comparison of NeXt2Former-CD with state-of-the-art change detection methods. F1 and IoU denote the F1 score and Intersection-over-Union for the change class, and OA denotes overall accuracy. For M-CD [ 14 ] , we re-evaluated the officially released models and report the corrected F1 scores (IoU/OA match the M-CD paper; see supplementary material). The best result is indicated in bold , and the second-best result is underlined .

Fusion Strategy	F1 $\uparrow$	IoU $\uparrow$	OA $\uparrow$
Cross-Attn	0.919	0.8498	99.321
Deformable-Attn	0.921	0.8531	99.330

Table 2: Ablation study on the use of deformable attention in the Feature Fusion Module (FFM) on LEVIR-CD [ 4 ] with batch size 64, evaluated on the validation split.

Figure 2: Qualitative results on three public datasets. White represents true positives, black represents true negatives, green represents false positives and red represents false negatives.

Loss Configuration	F1 $\uparrow$	IoU $\uparrow$	OA $\uparrow$
Cross Entropy Loss	0.916	0.8456	99.294
Mask2Former Set Loss	0.917	0.8474	99.316
Hybrid Loss (ours)	0.921	0.8531	99.330

Table 3: Ablation study on loss components of our method on LEVIR-CD [ 4 ] with batch size 64, evaluated on the validation split.

Method	Trainable Params (M)	GFLOPs	Time (ms)
M-CD	69.81	28.23	33.84
Ours	392.00	159.54	36.79

Table 4: Efficiency comparison between M-CD [ 14 ] and our proposed method. Inference time is measured per image pair on an RTX 5090 GPU.

Figure 3: Validation IoU curves for M-CD [ 14 ] and the proposed method. The vertical axis shows IoU for the change class.

	WHU-CD	LEVIR-CD	CDD
Reported in M-CD	0.953	0.921	0.982
Re-evaluated	0.954	0.919	0.981

Table S1: F1 scores reported in M-CD [ 14 ] and those obtained by re-evaluating the officially released checkpoints.

Backbone	#Params	Pretrain	LEVIR-CD IoU $\uparrow$
ViT-L/16	300M	SAT	0.8300
ViT-H+/16	840M	LVD	0.8330
ConvNeXt-L	198M	LVD	0.8515

Table S2: Effect of encoder backbone and pretraining data for DINOv3 [ 15 ] models on LEVIR-CD [ 4 ] . All models are trained with batch size 64 using DINOv3 encoders, the original FRM and FFM with cross-attention, and a Mask2Former [ 7 ] decoder with the original set-based loss. LVD denotes LVD-1689M (web-scale), and SAT denotes SAT-493M (satellite-domain) pretraining. Results are reported on the validation split.

Backbone	#Params $\downarrow$	WHU-CD IoU $\uparrow$
ConvNeXt-Tiny	29M	0.8966
ConvNeXt-Small	50M	0.9042
ConvNeXt-Base	89M	0.9048
ConvNeXt-Large	198M	0.9193

Table S3: Effect of encoder backbone capacity for DINOv3-pretrained ConvNeXt models on WHU-CD [ 10 ] . All models use DINOv3 encoders with identical FRM and FFM configurations, a Mask2Former [ 7 ] decoder, and the same training protocol. Only the ConvNeXt backbone scale (Tiny / Small / Base / Large) is varied. Results are reported on the validation split.

Explore similar work

ChangeT $^3$ : Evaluating Vision Test-Time Training for Remote Sensing Change Detection
Oct 11, 2026Sharukshan Niranjan, Oshadha Samarakoon, Nichula Wasalathilaka +3Remote Sensing Change Detection Test-Time Training
University of Peradeniya, Peradeniya, Sri Lanka
CoSA: Correlation-Guided Change Attention with Learnable Residual Gating for Remote Sensing Change Detection
Jun 20, 2026Abdirashid Omar, Jonghyuk ParkRemote Sensing Change Detection Attention Mechanisms
Department of Data Science, Graduate School of Kookmin University Seoul 02707, Republic of Korea
Beyond Quadratic: Linear-Time Change Detection with RWKV
Mar 20, 2026Zhenyu Yang, Gensheng Pei, Tao Chen +4Remote Sensing Change Detection
School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, China · State Key Laboratory of Intelligent Manufacturing of Advanced Construction Machinery, China · Department of Electrical and Computer Engineering, Sungkyunkwan University, Suwon, Korea

Oct 11, 2026J/K move · Enter open · S save

ChangeT $^3$ : Evaluating Vision Test-Time Training for Remote Sensing Change Detection

Sharukshan Niranjan, Oshadha Samarakoon, Nichula Wasalathilaka, Roshan Godaliyadda+2

University of Peradeniya, Peradeniya, Sri Lanka

Remote Sensing Change Detection Test-Time Training Vision Transformer

Open paper AskPDF arXiv

Bitemporal remote sensing change detection requires representations that preserve spatial detail while capturing contextual information relevant to changes between acquisitions. Visual test-time training offers an alternative to explicit token-to-token attention by adapting a compact inner model to each image. This letter describes ChangeT

^3

, a vanilla Siamese configuration combining the existing hierarchical ViT

^3

-Tiny encoder with the multiscale ChangeFormer decoder. The two image streams share learned encoder parameters, while their fast weights are computed independently from their respective spatial features. Each encoder block retains the inherited global simplified SwiGLU and local adaptive depthwise-convolution branches. Temporal comparison is performed by the decoder through learned fusion of paired features at four resolutions, followed by coarse-to-fine aggregation and binary prediction. This separation provides a direct setting for examining image-conditioned inner updates without introducing an additional temporal interaction module. Trained from scratch on LEVIR-CD+ with the same decoder, recipe, and budget as a ChangeFormer control, ChangeT

^3

reaches an F1 of 80.61

\pm

0.90 against 78.84

\pm

0.21 for ChangeFormer (three seeds each); as in the standard protocol for this dataset, checkpoints are selected on the test split. Trained with both inner updates disabled, the same encoder reaches 78.63

\pm

0.20. Neither model transfers zero-shot to SYSU-CD, DSIFN-CD, or WHU-CD, and ChangeT

^3

is the slower of the two at small inputs despite similar parameter and operation counts.

Abstract

Explore similar work

ChangeT3^33: Evaluating Vision Test-Time Training for Remote Sensing Change Detection

CoSA: Correlation-Guided Change Attention with Learnable Residual Gating for Remote Sensing Change Detection

Beyond Quadratic: Linear-Time Change Detection with RWKV

ChangeT $^3$ : Evaluating Vision Test-Time Training for Remote Sensing Change Detection