Source-linked AI summary
Change Detection Methods for Remote Sensing in the Last Decade: A Comprehensive Review
Guangliang Cheng, Yunmeng Huang, Xiangtai Li, Shuchang Lyu, Zhaoyang Xu, Qi Zhao, Shiming Xiang
TL;DR
Remote-sensing change detection must identify temporal changes despite variable image quality, noise, registration errors, illumination, and heterogeneous scenes. This paper surveys the past decade by organizing methods across three taxonomies, reviewing datasets and benchmarks, and identifying future directions. Its benchmark synthesis reports representative state-of-the-art methods across several dominant datasets and concludes that deep learning has been effective in addressing change-detection challenges.
Problem
Remote-sensing change detection is important across applications but remains difficult because paired images differ in quality, noise, registration, illumination, and scene characteristics.
Method
The paper comprehensively surveys recent methods, introduces task foundations and datasets, classifies algorithms by granularity, supervision, and learning framework, and summarizes benchmark performance.
Results
The benchmark review identifies representative state-of-the-art methods across several dominant datasets, including DDNet, SFCNet, ShearNet, ESMOFCM, BIFLICM/D, and MTBIT on reported dataset-specific metrics.
Takeaways & Limitations
The survey provides an organized view of the field, highlights deep learning’s effectiveness in addressing change-detection challenges, and identifies future research directions.
Abstract
from arXiv · showhide
Change detection is an essential and widely utilized task in remote sensing that aims to detect and analyze changes occurring in the same geographical area over time, which has broad applications in urban development, agricultural surveys, and land cover monitoring. Detecting changes in remote sensing images is a complex challenge due to various factors, including variations in image quality, noise, registration errors, illumination changes, complex landscapes, and spatial heterogeneity. In recent years, deep learning has emerged as a powerful tool for feature extraction and addressing these challenges. Its versatility has resulted in its widespread adoption for numerous image-processing tasks. This paper presents a comprehensive survey of significant advancements in change detection for remote sensing images over the past decade. We first introduce some preliminary knowledge for the change detection task, such as problem definition, datasets, evaluation metrics, and transformer basics, as well as provide a detailed taxonomy of existing algorithms from three different perspectives: algorithm granularity, supervision modes, and learning frameworks in the methodology section. This survey enables readers to gain systematic knowledge of change detection tasks from various angles. We then summarize the state-of-the-art performance on several dominant change detection datasets, providing insights into the strengths and limitations of existing algorithms. Based on our survey, some future research directions for change detection in remote sensing are well identified. This survey paper will shed some light on the community and inspire further research efforts in the change detection task.
I. INTRODUCTION
Remote-sensing change detection compares images of the same area captured at different times, but varying acquisition conditions and scene complexity make reliable detection difficult. This survey updates the field’s evidence base through broad coverage, three complementary taxonomies, benchmark summaries, and future research directions.
- Change detection identifies changes between images of the same geographical area captured at different times and supports applications including urban development, agriculture, and land-cover monitoring.
- Paired images vary in quality, noise, registration, illumination, viewing angle, and landscape structure, creating significant detection challenges.
- Deep learning, particularly CNNs and attention-based models, has become prominent because of its representation, nonlinear characterization, and dependency-modeling capabilities.
- The survey provides a more comprehensive review of change-detection research from the past decade than previous surveys and depicts the current state of the field.
- It organizes representative algorithms by algorithm granularity, supervision modes, and learning frameworks, while summarizing benchmark performance and future research trends.
II. PRELIMINARY KNOWLEDGE
The preliminary section establishes the knowledge needed to study and compare remote-sensing change-detection methods, covering the task, datasets, evaluation, and transformer foundations.
- The survey introduces the change-detection task, popular datasets from diverse data sources, crucial evaluation metrics, and basic transformer knowledge.
A. Task Definition
Remote-sensing change detection compares temporally separated data from the same location to identify land-surface changes. Its pipeline includes preprocessing, registration, algorithm selection, and optional post-processing, while methods can be classified by fusion position and output type.
- Task Definition: Change detection compares two or more remotely sensed images of the same area at different times to identify land-cover, land-use, natural, or human-driven changes.
- Task Definition: Algorithms are categorized by fusion-module position as early-fusion, middle-fusion, or late-fusion methods, including both deep-learning and traditional approaches.
- Task Definition: The pipeline may include noise removal, geometric and radiometric correction, image enhancement, spatial-temporal alignment, algorithm selection, and optional post-processing.Post-processing removes noise and false positives and generates a final change map, but the paper states it can be skipped.
- Task Definition: The formal pipeline applies preprocessing and registration before a change-detection algorithm produces maps of changed areas.PR denotes preprocessing and registration, CD denotes the change-detection algorithm, and M represents output maps of changed areas.
- Task Definition: Tasks are divided into binary change detection, which separates changed from unchanged areas, and semantic change detection, which identifies category or label changes.
B. Change Detection Datasets
The survey organizes dominant change-detection datasets by data source, covering SAR, multispectral, hyperspectral, heterogeneous, and 3D point-cloud data. Each source offers distinct capabilities and limitations relevant to change detection.
- SAR Data: SAR datasets support imaging in all weather conditions and through vegetation and clouds, but can suffer geometric distortion, interference, and speckle noise.
- Multi-spectral Data: Four primary multispectral datasets introduced are LEVIR-CD, CDD, WHU Building, and SECOND.
- Hyperspectral Data: Hyperspectral images use hundreds of narrow, contiguous spectral bands to distinguish materials with different spectral signatures.The survey uses the River and Hermiston datasets for hyperspectral data.
- Heterogeneous Data: Heterogeneous data combines images from multiple sensors, while integration introduces registration, normalization, and fusion challenges.The survey includes the California dataset, combining SAR and multispectral images.
- 3D Change Detection Data: 3D point-cloud data represents object or terrain surfaces in three-dimensional space and supports detecting changes in heights and shapes.The 3DCD dataset is used to benchmark state-of-the-art performance.
C. Evaluation Metrics
The survey introduces standard change-detection evaluation quantities and transformer fundamentals used to understand modern attention-based methods. Metrics quantify classification outcomes, while attention mechanisms emphasize relevant image regions and combine information across heads.
- Evaluation Metrics: Evaluation begins with TP and TN for correctly identified changed and unchanged pixels, and FP and FN for misclassified pixels.
- Evaluation Metrics: Precision measures accurate identification of changed pixels, whereas recall measures the proportion of ground-truth changed pixels detected.
- Evaluation Metrics: IoU measures overlap between predicted and ground-truth maps, OA measures prediction accuracy, and F1 is the harmonic average of precision and recall.
- Attention Basics: Attention assigns different importance weights to input parts so models can focus on task-relevant information.
- Attention Basics: Given Q, K, and V, attention produces an attention map that is multiplied with V to focus on relevant regions.
- Attention Basics: Multi-head attention performs parallel attention operations, concatenates their outputs, and fuses them with a projection matrix.
- Transformer Basics: Transformers combine encoder-decoder layers containing MHA and FFNs, with decoder cross-attention and positional encoding for positional information.
III. METHODOLOGY: A SURVEY
The methodology surveys change-detection algorithms through three complementary, orthogonal taxonomies: algorithm granularity, supervision modes, and learning frameworks. It also categorizes approaches by data type to account for differing data quantities and characteristics.
- The survey reviews methodologies by algorithm granularity, supervision modes, and learning frameworks.
- The three taxonomies are selected to be orthogonal so they complement one another.
- Approaches are additionally introduced by data type because data sources differ in quantity and characteristics.
A. Taxonomy based on Algorithm Granularity
The survey classifies change detection methods by algorithm granularity into pixel-based, region-based, and hybrid approaches, examining their use across remote sensing data sources. These categories trade computational efficiency, spatial context, noise robustness, and sensitivity to small or complex changes.
- Data-source coverage: The survey applies the granularity taxonomy across SAR, multispectral, hyperspectral, and 3D change detection methods.Representative literature is organized by data source and includes pixel-based, region-based, and hybrid approaches.
- Pixel-based methods: Pixel-based methods assign labels to individual pixels using spectral characteristics for image segmentation and change detection.They are easy to implement, computationally efficient, and effective for small changes in homogeneous regions, but they lack spatial context and are sensitive to noise.
- Region-based methods: Region-based methods first group pixels into meaningful objects, superpixels, or bounding boxes using spatial, spectral, and contextual characteristics.They capture spatial properties and are robust to noise and heterogeneous regions, but segmentation choices can be computationally intensive and may miss small changes.
- Hybrid methods: Hybrid methods combine pixel-based and region-based techniques to detect small and large changes in complex scenes.The combination improves robustness to noise and supports heterogeneous regions, although hybrid approaches can be computationally intensive and sensitive to segmentation choices.
B. Taxonomy based on Supervision Modes
The survey categorizes change detection methods by supervision mode into unsupervised, semi-supervised, and supervised learning. The modes differ in labeling requirements, accuracy potential, interpretability, and robustness to noise or distribution changes.
- Taxonomy: The survey organizes existing algorithms into the three supervision modes for different remote sensing data sources.This taxonomy supports detailed comparison of learning requirements and practical limitations.
- Unsupervised learning: Unsupervised learning identifies patterns without labels and accommodates diverse data sources, but difficulty distinguishing true changes from noise restricts performance in complex environments.The survey also notes that high dimensionality can cause computational inefficiency or inaccurate results.
- Semi-supervised learning: Semi-supervised learning trains with limited labeled data and many unlabeled samples, which is advantageous when labels are scarce or expensive.Its implementation can be difficult, and performance depends on the quality and quantity of labeled data.
- Supervised learning: Supervised learning uses labeled input-output data to learn predictions and can achieve high accuracy with sufficient labels.It requires accurate and unbiased labeled data, and models may struggle when test data differ from training data.
C. Taxonomy based on Learning Frameworks
The survey organizes learning frameworks into traditional, CNN-based, and attention- or transformer-based methods. It reviews representative designs spanning conventional models, convolutional feature learning, and mechanisms for spatial, spectral, temporal, or global relationships.
- Framework taxonomy: The learning-framework taxonomy divides change detection algorithms into traditional, CNN-based, and attention or transformer-based categories.Traditional methods commonly use conventional clustering or classification algorithms, whereas CNN and attention-based approaches use learned representations.
- CNN-based methods: CNN-based methods include supervised architectures that combine difference-image formation, semantic segmentation, clustering, and change classification.Other CNN designs use spatial and frequency-domain features for SAR change detection.
- Attention and transformer methods: Attention and transformer methods use attention mechanisms to model relationships across image features and support change detection in complex data.Reviewed examples address spatial, spectral, and temporal information, including hyperspectral change detection.
- Data-source coverage: Representative frameworks cover SAR, hyperspectral, and 3D point-cloud change detection literature.The survey tables organize methods by learning framework and data source, while examples include anomalous change detection and hyperspectral models.
IV. BENCHMARK PERFORMANCE
The survey summarizes representative state-of-the-art methods across dominant SAR, multispectral, and 3D change detection benchmarks. Performance leaders vary by dataset and metric, with some methods achieving strong results on specific benchmark measures.
- SAR benchmarks: 99.68% OA is achieved by ShearNet and ESMOFCM on the Bern SAR dataset.BIFLICM/D achieves the best KC metric on Bern, while DDNet and SFCNet lead different metrics on Yellow River.
- SAR benchmarks: DDNet leads FP and KC on Yellow River, while SFCNet achieves the best FN and OA metrics.These results show that the best method depends on the evaluation metric and dataset.
- Multispectral benchmarks: P2V-CD outperforms other methods on the LEVIR-CD multispectral dataset across all three reported metrics.The passage specifically notes a significant margin in KC.
- 3D benchmarks: MTBIT achieves state-of-the-art performance on both RMSE and cRMSE for the 3DCD dataset.The benchmark comparison covers representative models for 3D change detection.
V. FUTURE TRENDS
Future work targets broader generalization, reduced labeling demands, stronger multimodal integration, transformer advances, and foundation-model integration for remote-sensing change detection.
- Generalization of the change detection algorithms: Generalization research targets new data, locations, sensors, and environmental conditions through transfer learning, domain adaptation, multisensor fusion, and explainable AI.These directions are intended to enhance generalization and overall performance.
- Learning with few samples: Few-shot learning aims to train change detection models with limited examples using meta-learning, generative models, and domain generalization.These approaches target adaptation to new classes, domains, locations, sensors, and environmental conditions.
- Deep dive in the transformer-based algorithms: Transformer research emphasizes improved attention, integration with other deep learning models, unlabeled or partially labeled learning, multimodal data, and online continual learning.Target modalities include optical, radar, and LiDAR data.
- Synthetic data generation: Diffusion-based synthetic datasets may reduce reliance on real labeled data by producing high-quality images and masks for task-specific change detection needs.The passage highlights few-shot and long-tail applications as examples.
- Integration of multi-source/modal and multiple datasets: Future dataset research combines multiple datasets and modalities while addressing domain gaps, including cross-domain multisource data and image-text pairs.One proposed direction distills knowledge from multispectral change data into SAR data.
- Exploration based on foundation models: Foundation-model research proposes automatic dataset generation for aerial-specific models and foundation-model integration to address domain-adaptation gaps.Examples include Segment Anything, Painter, and SegGPT.
VI. CONCLUSIONS
The paper surveys a decade of remote-sensing change-detection advances, organizing foundational knowledge and algorithms, reviewing dataset performance, and identifying future research directions.
- VI. CONCLUSIONS: The survey provides an organized understanding of the field through fundamental knowledge, algorithm classification, and a review of recent advancements.It covers change detection for remote sensing images over the past decade.
- VI. CONCLUSIONS: Its dataset-based summary demonstrates the effectiveness of deep learning techniques in addressing change-detection challenges.The conclusion refers to state-of-the-art performance on several datasets.
- VI. CONCLUSIONS: The identified future directions offer potential avenues for further advancement and are intended to guide subsequent research.The authors state that the survey aims to contribute to current understanding and inspire research efforts.