Change Detection Methods for Remote Sensing in the Last Decade: A Comprehensive Review

I. INTRODUCTION

• Change detection identifies differences between images of the same area captured at different times and supports urban development,, agriculture,, land-cover monitoring –, disaster assessment, and military applications.Source: §I · p. 1
• Paired images can differ in quality, noise, registration, illumination, shadows, viewing angle, landscape dynamics, scale, and spatial heterogeneity, complicating change detection.Source: §I · p. 1
• Traditional classifiers, – face accuracy, generalization, and threshold-dependence limits; CNNs, –, attention –, and transformers, – show stronger change-detection performance.Source: §I · pp. 1–2
• Prior surveys cover general deep-learning methods and narrower topics, while the latest survey cited here covers algorithms only through 2021; recent attention, transformer, and diffusion work motivates an updated review.The introduction notes growing publication activity and identifies attention and transformer methods, [51], [52] and diffusion methods, as recent developments.Source: §I · p. 2
• The survey offers a more comprehensive overview of change-detection research across the past decade and systematically reviews current algorithms.Source: §I · p. 2
• It classifies representative methods by algorithm granularity, supervision mode, and learning framework, benchmarks dominant datasets, and discusses challenges and future trends.The survey also covers task definitions, datasets, evaluation metrics, and transformer basics.Source: §I · p. 2

II. PRELIMINARY KNOWLEDGE

• The preliminary section introduces popular change-detection datasets, evaluation metrics for benchmarking methods, and basic transformer knowledge used in attention- and transformer-based algorithms.It also previews a taxonomy of algorithms by granularity, supervision mode, and learning framework.Source: §II · p. 2

A. Task Definition

• Change detection compares two or more remotely sensed images of the same area captured at different times to identify land-surface changes, producing maps of changed areas.The inputs may be image or point-cloud data, and preprocessing and registration ensure paired inputs are aligned before the change-detection algorithm is applied.Source: A. Task Definition · pp. 2–3
• Early-, middle-, and late-fusion methods are distinguished by the position of the fusion module in the change-detection framework.This categorization applies to traditional methods as well as deep-learning methods.Source: A. Task Definition · p. 2, Figure 3 · p. 3
• The task pipeline includes image preprocessing to remove noise and correct distortions, spatial and temporal registration, and algorithm selection based on the application and image characteristics.Algorithm options include traditional, CNN, and transformer methods.Source: A. Task Definition · pp. 2–3
• Post-processing can remove noise and false positives to produce a final change map, but it is optional; the survey focuses primarily on change-detection algorithms.For other pipeline steps, the paper recommends existing works,,.Source: A. Task Definition · p. 3
• Inputs may come from homogeneous sources such as SAR, multispectral, hyperspectral, or point-cloud data, or from different sources as heterogeneous data.The passage defines P as image height × width for images and as the number of points for point clouds.Source: A. Task Definition · pp. 3–4
• Change-detection outputs are either binary, distinguishing changed from unchanged areas, or semantic, identifying changes in object categories or labels before and after change,, [60]–.Source: A. Task Definition · p. 4

SAR Data

• SAR works in all weather and through vegetation and clouds, but faces geometric distortion, interference, and speckle; the survey includes Yellow River and Bern.Its high-resolution imaging can identify small features.Source: B. Change Detection Datasets · p. 4

Multi-spectral Data

• Multispectral bands distinguish surface features such as vegetation, water, and urban areas; the survey introduces LEVIR-CD, CDD [73], WHU Building, and SECOND.Healthy vegetation typically reflects more near-infrared energy than other land-cover types.Source: B. Change Detection Datasets · p. 4

Hyperspectral Data

• Hyperspectral images capture hundreds of narrow, contiguous bands, enabling discrimination of materials with similar appearances; the survey uses River and Hermiston.The source describes applications including environmental monitoring, mineral exploration, and crop health assessment.Source: B. Change Detection Datasets · p. 4

Heterogeneous Data

• Heterogeneous imagery integrates sensors such as optical, SAR, and LiDAR for more comprehensive surface information, but requires data registration, normalization, and fusion; the survey includes California.California combines SAR and multispectral images.Source: B. Change Detection Datasets · pp. 4–5

3D Change Detection Data

• LiDAR-derived 3D point clouds support change detection of terrain and object heights and shapes over time; the survey uses 3DCD for benchmarking.The source notes applications to buildings, vegetation, and coastlines.Source: B. Change Detection Datasets · p. 5

C. Evaluation Metrics

• TP and TN count correctly identified changed and unchanged pixels, while FP and FN count pixels wrongly classified as changed or unchanged.These counts provide the basis for evaluating a change detection model.Source: C. Evaluation Metrics · p. 5
• Precision reflects accurate identification of changed pixels, recall reflects the proportion of ground-truth changes detected, and F1-score is their harmonic average.The section also describes IoU as the ratio of intersection to concatenation of the predicted map and ground truth, and OA as prediction accuracy.Source: C. Evaluation Metrics · p. 5
• KC measures classification accuracy from the confusion matrix, while RMSE and cRMSE are used for 3D change detection.cRMSE considers only errors in ground-truth pixels influenced by altitude; SCD uses mIoU, Sek, and Fscd, with further metric information in.Source: C. Evaluation Metrics · p. 5

D. Transformer Basics

• Attention weights parts of an input by relevance, enabling models to focus on useful information in text or images.The passage notes its growing use in NLP and computer vision.Source: D. Transformer Basics · p. 5
• Multi-head attention runs multiple attention operations in parallel, concatenates their outputs, and fuses them with a projection matrix.The attention map is multiplied by V; the scale factor dk represents the dimension of the keys.Source: D. Transformer Basics · p. 6
• MHSA uses queries and keys from the same sequences or tokens, whereas MHCA uses them from different sequences or tokens.Source: D. Transformer Basics · p. 6
• The Transformer architecture [52] has encoder and decoder layers with multi-head attention and feedforward networks, and adds positional encodings to capture position.Its feedforward network consists of two consecutive MLPs with a nonlinear activation; the decoder uses MHCA to attend across multimodal sequences.Source: D. Transformer Basics · p. 6

• The methodology organizes algorithms by granularity, supervision mode, and learning framework; the three taxonomies are chosen to be orthogonal and complementary.Source: §III · p. 6
• The survey also presents approaches by data type, reflecting differences in data quantity and characteristics across sources.Source: §III · p. 6

SAR Methods

• The taxonomy distinguishes pixel, region, and hybrid algorithms for each data source, illustrating SAR methods with pixel classifiers, superpixel or object-based methods, and pixel-region fusion.SAR examples include unsupervised filtering and clustering, CNN-based false-alarm screening, and region-level superpixel modeling.Source: A. Taxonomy based on Algorithm… · pp. 6–8

A. Taxonomy based on Algorithm Granularity

• Pixel-based methods label individual pixels from spectral characteristics; deep-learning receptive-field mechanisms can improve on traditional methods’ false positives and false negatives.Most end-to-end CNN models fall into this category.Source: A. Taxonomy based on Algorithm… · p. 6
• Region-based methods first group pixels into objects, superpixels, or bounding boxes, then use those regions as units for change detection and recognition.Grouping uses spatial, spectral, and contextual characteristics and can be followed by traditional or deep-learning methods.Source: A. Taxonomy based on Algorithm… · p. 6
• Hybrid methods combine pixel- and region-based techniques in parallel or succession to address individual methods’ limitations and provide a more robust change-detection approach.The survey describes hybrid approaches as potentially improving accuracy by analyzing changes from different perspectives.Source: A. Taxonomy based on Algorithm… · p. 6

Multi-spectral Methods

• Multispectral examples span pixel, region, and hybrid approaches; high-resolution imagery has also motivated object- and superpixel-based methods for homogeneous results and efficiency.The surveyed examples include differential and lightweight CNNs, object-oriented CRF and fuzzy-fusion approaches, and a dual-pathway network.Source: A. Taxonomy based on Algorithm… · p. 8

Hyperspectral Methods

• Hyperspectral examples include pixel-based CNN and transformer methods, alongside hybrid strategies that combine spectral distance measures or fuse multiscale changed features.The surveyed pixel methods include spectral-spatial attention and space-time encoding with self-attention.Source: A. Taxonomy based on Algorithm… · p. 8

Heterogeneous Methods and 3D Change Detection Methods

• Heterogeneous-image examples cover pixel, region, and hybrid methods, while 3D examples include pixel-level point-cloud or height-difference analysis and canopy-height detection followed by object-based refinement.The heterogeneous methods include coupled CNNs, graph convolution with multiscale objects, and adversarial fusion.Source: A. Taxonomy based on Algorithm… · p. 8

A. Taxonomy based on Algorithm Granularity

• Pixel methods are efficient and can detect small changes, but struggle in heterogeneous regions, omit spatial context, and are noise-sensitive; region methods capture spatial structure and resist noise but are computationally intensive and may miss small changes.Hybrid methods combine pixel and region strengths for small and large changes in complex scenes, but can inherit segmentation sensitivity and computational cost.Source: A. Taxonomy based on Algorithm… · p. 8

1) SAR Methods

• For SAR, scarce public datasets make unsupervised clustering and fixed-feature networks prevalent, alongside semi-supervised methods using limited labels and supervised spatial-temporal feature models.The survey also describes unsupervised difference-image and generative approaches, and semi-supervised pretraining or self-ensembling.Source: B. Taxonomy based on Supervision Modes · p. 10

2) Multi-spectral Methods

• Multispectral methods span unsupervised feature learning and image coregistration, semi-supervised pseudo-labeling and contrastive learning, and supervised differential-pyramid fusion and metric learning.The described supervised approaches use side-out fusion or aggregate multiscale and global context for discriminative features.Source: B. Taxonomy based on Supervision Modes · p. 10

3) Hyperspectral Methods

• Hyperspectral methods include unsupervised binary-to-multiclass detection, semi-supervised classification from pseudo-labels, and supervised networks modeling spatial-spectral dependencies or refining multipath features.The supervised examples also use channel attention to refine features across paths.Source: B. Taxonomy based on Supervision Modes · p. 10

4) Heterogeneous Methods

• For heterogeneous imagery, methods align images through adversarial translation or prior-weighted losses, transfer features in semi-supervised learning, and use multiscale or semantic-style separation in supervised models.The described adversarial approach combines structural-consistency and adversarial losses to encourage a common domain and consistent image style.Source: B. Taxonomy based on Supervision Modes · p. 10

5) 3D Change Detection Methods

• 3D methods use unsupervised tree clustering or polar change-vector analysis, while supervised approaches segment buildings from point clouds or detect and categorize changes in one step.The building-segmentation approach reduces point clouds to a smaller representation while retaining necessary information.Source: B. Taxonomy based on Supervision Modes · pp. 10–11

Taxonomy based on Supervision Modes

• Unsupervised methods avoid labels but can be hard to interpret and suffer dimensionality problems; semi-supervised methods use unlabeled data but depend on its quality, while supervised methods need accurate labels and may not generalize.With sufficient labeled data, supervised methods can achieve high accuracy; scarce labels and complex patterns limit their use.Source: B. Taxonomy based on Supervision Modes · pp. 9–11

Learning-framework taxonomy

• Traditional methods use clustering or classification, CNNs learn spatial features, and attention or transformers model relevant regions and long-range dependencies.The survey reports that CNN methods outperform traditional methods, while attention or transformer methods are the new state of the art,,,,,.Source: C. Taxonomy based on Learning Frameworks · p. 11

SAR Methods

• SAR methods span traditional unsupervised approaches, staged CNN pipelines, and attention-based or jointly optimized unsupervised frameworks.Yu et al. combine likelihood-ratio similarity, Shannon-entropy differencing, and MRF segmentation; Liu et al. use prior detections to select random-forest training data. Vinholi et al. use four stages, while DDNet fuses spatial and frequency features; MSDC and ASGF use unsupervised clustering, and Li et al. use multiscale attention.Source: C. Taxonomy based on Learning Frameworks · p. 11

Multi-spectral Methods

• Multispectral methods range from traditional fusion and MRF approaches to CNN feature learning and transformer-based temporal, cross-scale, and cross-attention modeling.Shao et al. use Dempster–Shafer fusion and a modified CRF, while Fang et al. combine weighted change-vector analysis with an improved MRF. CNN examples include ECFNet and semantic-aware pretraining; transformer examples include DMATNet, a bitemporal image transformer, PA-Former, and ACAHNet.Source: C. Taxonomy based on Learning Frameworks · p. 11

Hyperspectral Methods

• Hyperspectral methods include anomaly detection and subspace learning, CNN-based spectral-spatial feature learning, and band-selection and feature-fusion frameworks.ACDA is described as outperforming other state-of-the-art approaches, while SMSL preserves major image-pair information with a sketched representation. MMSRC targets multiclass change detection; SFBS-FFGNET selects effective bands and fuses features.Source: C. Taxonomy based on Learning Frameworks · pp. 11–12
• Hyperspectral attention methods integrate spatial, spectral, and temporal information or emphasize correlation differences and multiscale features.SST-Former [155] jointly considers spatial, spectral, and temporal information; Dong et al. highlight subpixel-level correlation differences, CSANet integrates joint spatial-spectral-temporal features, and DPMs2raN mines multilevel spatial-spectral features for complex changed regions.Source: C. Taxonomy based on Learning Frameworks · p. 13

Heterogeneous Methods

• Heterogeneous methods address cross-source differences through signal decomposition, image regression, autoencoders, domain adaptation, or topology-coupled attention.SDA-HCD separates regressed and changed signals; Sun et al. use structure-consistent regression and superpixel segmentation. CNN examples include CAE, TVRBN, and DA-MSCDNet, which aligns optical and SAR feature spaces; TSCNet transforms heterogeneous feature spaces and adds wavelet, channel, and spatial attention.Source: C. Taxonomy based on Learning Frameworks · p. 13

3D Change Detection Methods

• 3D methods include object-based and point-comparison approaches alongside deep networks for point-cloud change detection and categorization.Dai et al. combine segmentation, clustering, and bidirectional object analysis; Liu et al. set adaptive thresholds using neighborhood distance and local point-cloud density. ChangeGAN combines Siamese and U-net-like extraction with spatial transformation, while Siamese KPConv performs change detection and categorization in one step.Source: C. Taxonomy based on Learning Frameworks · p. 13

1) SAR Benchmarks

• On Yellow River, DDNet leads FP and KC while SFCNet leads FN and OA; on Bern, ShearNet and ESMOFCM attain 99.68% OA, and BIFLICM/D leads KC.Source: §IV · p. 13, Table VI · p. 14

2) Multi-spectral Benchmarks

• Across multispectral benchmarks, P2V-CD leads F1 on LEVIR-CD, CDD, and WHU Building; ChangeStar leads LEVIR-CD IoU, ChangeFormer its OA, and SSCD/SCanNet lead SECOND's OA and mIoU, Sek, and Fscd, respectively.On WHU Building, DASNet achieves the second-best F1; on CDD, P2V-CD also leads precision and recall.Source: §IV · p. 13, Table VII · p. 14

3) Hyperspectral Benchmarks

• On River, MSDFFN leads Recall, F1, OA, and KC while SSA-SiamNet leads precision; on Hermiston, MSDFFN leads precision, recall, F1, and OA, and GETNET leads KC.Source: §IV · p. 13

4) Heterogeneous Benchmarks

• On the California heterogeneous dataset, DPFL-Net-4 leads OA, KC, and AUC, with particularly strong performance in KC.Source: §IV · pp. 13–14

5) 3D Change Detection Benchmarks

• On 3DCD, MTBIT achieves state-of-the-art performance on both RMSE and cRMSE.The section presents representative models for this 3D change-detection benchmark.Source: §IV · p. 14

Generalization of the change detection algorithms

• Transfer learning, domain adaptation, multisensor fusion, and explainable AI are proposed to help change detectors generalize across locations, sensors, and environmental conditions.The goal is to detect changes in new and unobserved data while improving overall performance.Source: §V · p. 14

Learning with few samples

• Few-shot learning [195], [196] could support detection of new change classes from few examples through meta-learning, generative models, and domain generalization.Training on varied domains is proposed to improve performance in new locations, with different sensors, and under changing environmental conditions.Source: §V · p. 15

Deep dive in the transformer-based algorithms

• Transformer-based change detection [52], – remains an emerging direction, with proposed work on attention, hybrid models, unlabeled-data learning, multimodal inputs, and continual adaptation.The proposed modalities include optical, radar, and LiDAR data; online and continual learning, could adapt models to new data streams.Source: §V · p. 15

Efficient models for practical applications

• Sparse and lightweight models – and compression methods such as pruning, quantization, and knowledge distillation, are proposed to reduce computation and memory demands.The aim is to retain high accuracy while making change detection models more practical and scalable.Source: §V · p. 15

Generate synthetic data for joint training

• Diffusion-based generation models, could produce synthetic images and masks with fewer domain gaps than real images for training without relying on real data.Synthetic samples could be tailored to few-shot or long-tail tasks, reducing dependence on substantial manual data collection and annotation.Source: §V · p. 16

Integration of multi-source/modal and multiple datasets

• Combining datasets across sources could improve performance and generality, while reducing domain gaps and exploring cross-domain multi-source training, including multispectral-to-SAR knowledge distillation.The section also proposes image-text datasets for selected-area change detection to combine visual and textual information.Source: §V · p. 16

Exploration based on foundation models

• Foundation models such as Segment Anything [218], Painter, and SegGPT [220] motivate larger aerial-specific datasets and model integration to address domain-adaptation gaps.Segment Anything is highlighted as versatile for remote sensing, especially with high-resolution images.Source: §V · p. 16

VI. CONCLUSIONS

• The survey organizes fundamental knowledge and existing algorithms to provide a detailed understanding of the current state of remote-sensing change detection.It reviews advances achieved over the past decade.Source: §VI · p. 16
• Its summary of state-of-the-art performance on several datasets demonstrates the effectiveness of deep learning in addressing change-detection challenges.Source: §VI · p. 16
• The survey identifies future research directions as potential avenues for advancing the field and aims to inspire and guide further research.Source: §VI · p. 16