cs.CVAug 28, 2026

A comprehensive and trustworthy benchmark of AI methods for change detection in Earth observation

Authors: Tadej Tomanič, Alice Baudhuin, Jan Sotošek, Jure Brence, Panče Panov, Nikola Simidjievski, Dragi Kocev

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Change detection in Earth observation (EO) is critical for monitoring land surface transformations, yet recent research in the field is constrained by inconsistent evaluation protocols and a narrow focus on predictive accuracy without regard for computational efficiency. To address this, we present a standardized, open-source benchmark for evaluating state-of-the-art (SOTA) deep learning methods for Earth observation change detection. We conduct a comprehensive analysis of ten representative model architectures, ranging from convolutional networks (CNNs) to vision transformers (ViTs), across ten heterogeneous change detection datasets. We rigorously evaluate these models with identical experimental protocols, comparing models trained from scratch against those utilizing pre-trained weights. Furthermore, we evaluate predictive performance alongside computational efficiency, including parameter counts and inference latency. Our findings reveal that well-optimized classical architectures, such as Siamese U-Nets, frequently outperform more complex contemporary models when computational efficiency is factored in, and that pre-training consistently provides a significant performance boost with no additional inference cost. To ensure complete transparency and reproducibility, all experimental resources, including standardized data splits, training scripts, training logs, and model checkpoints are publicly available and adhere to FAIR principles (Findable, Accessible, Interoperable, and Reusable).

Figures & tables

DatasetImage sourceImage pairsBandsPatch sizeGSD/ Resolution
BANDON ( Pang et al., 2023 )Off-nadir aerial2,28332048×20482048\times 20480.60.6 m
CLCD ( Liu et al., 2022 )GaoFen-26003512×512512\times 5120.50.5 – 22 m
DSIFN ( Zhang et al., 2020 )Google Earth3,9883512×512512\times 5120.50.5 – 22 m
EGY-BCD ( Holail et al., 2023 )Google Earth6,0913256×256256\times 2560.250.25 m
LEVIR-CD+ ( Chen and Shi, 2020 )Google Earth98531024×10241024\times 10240.50.5 m
MSBC ( Li et al., 2022 )GaoFen-23,7693256×256256\times 2560.80.8 – 1010 m
Table 1: Key characteristics of the change detection datasets used for the benchmarking experiments.
Figure 1: Example image pair ( T1,T2T_{1},T_{2} ) and ground truth change mask from the Season-Varying CDD dataset ( Lebedev et al., 2018 ) .
ModelYearParams (M)GFLOPs
BIT ( Chen et al., 2022 )20213.5010.88
CGNet ( Han et al., 2023 )202333.6887.55
ChangeFormerV6 ( Bandara and Patel, 2022 )202241.03138.77
ChangeViT ( Zhu et al., 2026 )202620.6626.36
CSSM ( Ghazaei and Aptoula, 2026 )20252.592.78
HRNet SiamConc ( Sun et al., 2019 )201921.489.55
Table 2: Summary of the representative model architectures considered in this benchmark. FLOPs and parameter counts are reported for bi-temporal input tensor pairs of size 2×3×256×2562\times 3\times 256\times 256 .
Figure 2: Ontological schema used by the benchmark.
(a) Trained from scratch
DatasetBITCGNetChangeFormerV6ChangeViTCSSMHRNet SiamConcSiamCRNNSTANetTinyCDU-Net SiamConcAverage
BANDON0.50150.47850.56610.55450.47870.52310.55110.49210.53060.48400.5160
CLCD0.69200.65930.65690.66370.45960.63750.69500.61200.66120.70480.6442
DSIFN0.77110.81090.74530.79190.68290.64090.78630.72240.71330.75930.7424
EGY-BCD0.77830.82140.75940.77700.70050.76190.79290.75980.76190.77620.7689
LEVIR-CD+0.63780.64120.65610.71810.57850.67660.69070.54490.72010.72010.6584
Table 3: Mean Intersection over Union (mIoU) benchmark results across ten change detection datasets, comparing models trained from scratch versus pre-trained backbones. The best performance per dataset within each training setup is indicated in bold , and the second-best is underlined .
Figure 3: Comparison of a) mIoU and b) macro F1F_{1} -score for the benchmarked change detection models. Points represent the metric values averaged across the ten benchmark datasets for each architecture. Error bars denote the 95% confidence interval, indicating the performance variance across the datasets. The models are evaluated under two different training regimes: trained from scratch (orange circles, dashed line) and initialized with pre-trained weights (blue squares, solid line). Figures c) and d) present diverging bar charts illustrating the absolute performance difference ( Δ\Delta ) between the pre-trained models and models trained from scratch. Blue bars extending to the right indicate a performance gain from pre-training, while orange bars to the left indicate a degradation. Models lacking pre-trained weights (ChangeFormerV6 and CSSM) are excluded from the delta comparisons.
ModelActivationLatencyThroughputAverage train time (h)
 memory (GB)(ms)(FPS)ScratchPre-trained
BIT8.3911.9283.892.21 ±\pm 2.571.72 ±\pm 2.06
CGNet11.516.02166.110.93 ±\pm 1.020.15 ±\pm 0.18
ChangeFormerV614.7822.3244.801.53 ±\pm 1.68/
ChangeViT10.4024.4740.872.12 ±\pm 2.401.58 ±\pm 1.81
CSSM1.6460.6316.491.79 ±\pm 2.13/
Table 4: Resource profiling and computational efficiency of the benchmarked models. Peak activation memory is estimated for a training batch of size 16. Latency and throughput are measured on a single GPU with bi-temporal input tensor of size 2×3×256×2562\times 3\times 256\times 256 (batch size 1 for inference). Training times represent the average wall-clock duration to reach convergence across all benchmark datasets.
Figure 4: Trade-off between predictive performance and computational efficiency for the benchmarked change detection models. a) Theoretical efficiency displaying average mIoU against GFLOPs, with shape size representing the total number of parameters (in millions). b) Practical hardware efficiency, showcasing average mIoU against inference latency (ms), with bubble size indicating peak activation memory (GB). Marker colors correspond to the specific model architectures, while marker shapes denote the training type: random initialization from scratch (circles) and ImageNet-1K pre-training (squares). Models positioned toward the top-left with smaller shape sizes represent the optimal model architecture, yielding high performance at low resource costs.
Figure 5: Qualitative comparison of model predictions on the Season-varying CDD dataset. The figure displays the bi-temporal input imagery, ground truth masks, and predicted change maps for three models trained from scratch: a) ChangeViT, b) U-Net SiamConc, and c) CGNet.
Figure 6: Qualitative comparison of model predictions (flood mapping) on the OMBRIA dataset. The figure displays the bi-temporal input imagery, ground truth masks, and predicted change maps for three models trained from scratch: a) ChangeFormerV6, b) U-Net SiamConc, and c) STANet.
More in this topic

Remote Sensing Change Detection

Remote sensing change detection finds and interprets differences across Earth-observation images acquired at different times.

See all 89 papers

Explore similar work

CardsList
Mar 20, 2026

Beyond Quadratic: Linear-Time Change Detection with RWKV

Existing paradigms for remote sensing change detection are caught in a trade-off: CNNs excel at efficiency but lack global context, while Transformers capture long-range dependencies at a prohibitive computational cost. This paper introduces ChangeRWKV, a new architecture that reconciles this conflict. By building upon the Receptance Weighted Key Value (RWKV) framework, our ChangeRWKV uniquely combines the parallelizable training of Transformers with the linear-time inference of RNNs. Our approach core features two key innovations: a hierarchical RWKV encoder that builds multi-resolution feature representation, and a novel Spatial-Temporal Fusion Module (STFM) engineered to resolve spatial misalignments across scales while distilling fine-grained temporal discrepancies. ChangeRWKV not only achieves state-of-the-art performance on the LEVIR-CD benchmark, with an 85.46% IoU and 92.16% F1 score, but does so while drastically reducing parameters and FLOPs compared to previous leading methods. This work demonstrates a new, efficient, and powerful paradigm for operational-scale change detection. Our code and model are publicly available.

Zhenyu Yang, Gensheng Pei, Tao Chen +4

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

Remote Sensing Change Detection
AskPDF
Apr 15, 2026

Decoding the Delta: Unifying Remote Sensing Change Detection and Understanding with Multimodal Large Language Models

While Multimodal Large Language Models (MLLMs) excel in general vision-language tasks, their application to remote sensing change understanding is hindered by a fundamental "temporal blindness". Existing architectures lack intrinsic mechanisms for multi-temporal contrastive reasoning and struggle with precise spatial grounding. To address this, we first introduce Delta-QA, a comprehensive benchmark comprising 180k visual question-answering samples. Delta-QA unifies pixel-level segmentation and visual question answering across bi- and tri-temporal scenarios, structuring change interpretation into four progressive cognitive dimensions. Methodologically, we propose Delta-LLaVA, a novel MLLM framework explicitly tailored for multi-temporal remote sensing interpretation. It overcomes the limitations of naive feature concatenation through three core innovations: a Change-Enhanced Attention module that systematically isolates and amplifies visual differences, a Change-SEG module utilizing Change Prior Embedding to extract differentiable difference features as input for the LLM, and Local Causal Attention to prevent cross-temporal contextual leakage. Extensive experiments demonstrate that Delta-LLaVA decisively outperforms leading generalist MLLMs and specialized segmentation models in complex change deduction and high-precision boundary localization, establishing a unified framework for earth observation intelligence.

Xiaohe Li, Jiahao Li, Kaixin Zhang +5

Beijing, China Aerospace Information Research Institute, CAS · Beijing, China Space Engineering University · Beijing, China Capital Normal University

Remote Sensing Change DetectionRemote Sensing VQAMultimodal Large Language ModelsRemote Sensing Image SegmentationRemote Sensing Benchmarks
AskPDF
Feb 21, 2026

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

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.

Yufan Wang, Sokratis Makrogiannis, Chandra Kambhamettu

University of South Florida · Delaware State University

Remote Sensing Change DetectionHybrid CNN-Transformer ArchitecturesConvolutional Neural Networks
AskPDF