PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning for Change Detection in Remote Sensing Images: Comprehensive Review and Meta-Analysis

Lazhar Khelifi, Max Mignotte

arXiv:2006.05612v1cs.CV

TL;DR

Remote-sensing change detection has attracted substantial deep-learning research, but its recent progress lacked a focused comprehensive synthesis. This paper combines a systematic meta-analysis with a technical review, categorizes methods by supervision and transfer-learning strategy, and identifies future research directions while noting limitations in weakly supervised localization.

  • Problem

    Prior reviews addressed deep learning across major remote-sensing tasks, but no work had extensively studied deep learning specifically for change detection.

  • Method

    The study systematically searched published deep-learning change-detection studies, conducted a meta-analysis, reviewed frequently used deep models, and categorized methods as fully supervised, fully unsupervised, or transfer learning-based.

  • Results

    The reviewed literature showed rapid growth in publications and citations, with CNNs the most widely applied model for remote-sensing change detection.

  • Takeaways & Limitations

    The review provides a basis for further research and identifies deep reinforcement learning and weakly supervised change detection as promising directions.

  • Takeaways & Limitations

    Existing weakly supervised change-detection methods remain unsatisfactory for accurately locating changes, including building changes.

Abstract

from arXiv · show

Deep learning (DL) algorithms are considered as a methodology of choice for remote-sensing image analysis over the past few years. Due to its effective applications, deep learning has also been introduced for automatic change detection and achieved great success. The present study attempts to provide a comprehensive review and a meta-analysis of the recent progress in this subfield. Specifically, we first introduce the fundamentals of deep learning methods which arefrequently adopted for change detection. Secondly, we present the details of the meta-analysis conducted to examine the status of change detection DL studies. Then, we focus on deep learning-based change detection methodologies for remote sensing images by giving a general overview of the existing methods. Specifically, these deep learning-based methods were classified into three groups; fully supervised learning-based methods, fully unsupervised learning-based methods and transfer learning-based techniques. As a result of these investigations, promising new directions were identified for future research. This study will contribute in several ways to our understanding of deep learning for change detection and will provide a basis for further research.

1 Introduction

Deep learning has become influential in remote sensing, including automated change detection, but a focused synthesis of this subfield was lacking. This review addresses that gap through a meta-analysis and technical classification of deep learning change-detection studies.

  • Deep learning learns hierarchical data representations through multiple processing layers and has been applied across computer vision, speech recognition, information retrieval, and remote sensing.
  • Change detection identifies differences between images of the same geographical area acquired at distinct times.
  • Change detection supports applications including disaster assessment, environmental monitoring, land management, and urban change analysis.
  • Extreme disasters associated with climate change create a need for more effective automated change-detection methods, motivating deep learning applications in remote sensing.
  • Because prior reviews covered broad remote-sensing applications rather than change detection specifically, this study conducts a meta-analysis and technical review of the subfield.

2 Change detection in remote sensing

Remote-sensing change detection compares temporally separated imagery of the same area to produce a binary map of changed and unchanged pixels. Its accuracy is challenged by acquisition variability and atmospheric or illumination effects that can create irrelevant differences.

  • Change detection distinguishes changed and unchanged pixels in bi-temporal or multi-temporal, co-registered images acquired over the same area at different times.
  • In the classic monomodal case, the input images share dimensions W × L, cover the same region, use the same sensor, and are acquired at times t1 and t2.
  • The primary system output is a binary change map in which each pixel receives a changed or unchanged label.
  • Traditional methods commonly compute a difference image, such as DI = |I1 − I2|, before analyzing the result.
  • Change detection is used in agricultural, military, civil, and environmental applications, including deforestation monitoring, damage assessment, and urban development control.
  • Clouds, fog, dust, changing sun angles, and shadows can introduce irrelevant differences and reduce the quality of the resulting change map.

3 Brief Overview of deep learning

This section introduces deep learning models used in remote-sensing change detection, from foundational architectures to CNNs and RNNs. It emphasizes hierarchical feature learning, reconstruction, convolutional processing, and sequential hidden-state updates.

  • Models for change detection: The review organizes deep network models for remote-sensing change detection around DBNs, SAEs, GANs, CNNs, and RNNs.These models are presented as widely designed architectures for the task.
  • Deep learning foundations: Deep networks learn hierarchical representations through multiple processing layers, progressively transforming inputs into higher-level features.The section describes hidden layers as intermediate transformations between inputs and outputs.
  • Deep belief networks: DBNs are constructed by stacking RBMs, whose visible and hidden units form a bipartite graph without within-layer connections.RBMs use learnable weights and biases to define probabilities over joint visible-hidden configurations through an energy function.
  • Stacked autoencoders: Autoencoders encode an input into a hidden representation and decode it into a reconstruction, while SAEs connect successive hidden layers across multiple autoencoders.Training generally minimizes Euclidean reconstruction distance between the input and reconstructed output.
  • Convolutional neural networks: CNNs process grid-like image data by applying convolutions and nonlinear activations, followed by pooling and fully connected layers that produce class probabilities.Pooling summarizes local neighborhoods and gradually reduces feature-map spatial size; softmax outputs normalized probabilities.
  • Recurrent neural networks: RNNs process sequences by combining each current observation with a recurrent hidden state from the previous time step.The same recurrent matrix is reused across time steps, although standard RNNs are described as vulnerable to vanishing gradients.

4 Methods and data used to review DL for CD in remote sensing images

The review assembled a screened corpus of deep-learning change-detection studies and used publication, citation, venue, and model or image-type distributions to characterize the field’s growth and composition.

  • Data collection: 160 unique studies were collected from Web of Science using controlled vocabulary for deep learning, change detection, and remote sensing.The corpus included 110 journal articles, 47 conference papers, two early-access papers, and one editorial material paper.
  • Publication venues: 82 of the 110 peer-reviewed journal papers appeared across the ten journals identified as most relevant to deep-learning change detection and remote sensing.The five leading venues listed were Remote Sensing, TGRS, IEEE Access, IEEE JSTARS, and IEEE GRSL.
  • Screening: 20 journal papers were excluded after content review because they did not cover the study’s subject, while conference papers were excluded from the meta-analysis.The authors excluded conference papers because many had later been expanded into journal papers.
  • Publication venues: The topic was well represented at major international remote-sensing conferences, with most of the 47 conference papers published by IGARSS and SPIE.Conference papers were not included in the meta-analysis.
  • Field growth: Publication and citation counts increased markedly after 2015, with projections exceeding 100 papers and 1000 citations in 2020.The study interprets this growth as evidence of rapidly increasing interest in deep-learning change detection.

5 Deep learning for change detection in remote Sensing images

The review organizes deep-learning change-detection methods into supervised, unsupervised, and transfer-learning approaches, while relating their use to data availability and remote-sensing image types. CNNs were the most widely applied models, and SAR images were the most commonly used data type.

  • Method classification: Deep-learning change-detection methods are classified into fully supervised, fully unsupervised, and transfer-learning groups.The classification provides the framework for the detailed review of prior methods.
  • Image-type distribution: SAR images were the most commonly used remote-sensing image type, followed by multispectral, aerial, optical, heterogeneous, and hyperspectral images.The cited distribution summarizes the image modalities represented in the reviewed studies.
  • Transfer learning: Transfer learning applies knowledge from one source task to an associated target task to reduce fine-tuning data requirements or improve performance.The review describes feature extraction from pretrained networks and network fine-tuning as two common approaches.
  • Data constraints: Insufficient annotated change-detection data and the high effort required to construct ground-truth maps constrain fully supervised methods.These conditions motivate unsupervised learning and transfer learning as alternatives or complements.

6 Promising research directions

The paper proposes deep reinforcement learning and weakly supervised learning as two directions for advancing remote-sensing change detection under limited or costly supervision. It frames reinforcement learning as sequential action refinement and weak supervision as promising but currently insufficiently accurate.

  • 6.1 Deep reinforcement learning: Deep reinforcement learning is proposed to address limited labeled databases that can impoverish learned feature representations.DRL combines deep neural networks with reinforcement learning, allowing agents to learn through exploration, feedback, and rewards.
  • 6.1 Deep reinforcement learning: Change detection can be formulated as an action-decision problem in which sequential actions refine changed-region size between two input images.This direction follows an aircraft-detection framework combining a CNN with reinforcement learning.
  • 6.2 Weakly supervised change detection: Weakly supervised learning is motivated by expensive, unreliable annotation and difficult change-detection imagery containing small changed regions, cluttered backgrounds, and differing sensors.The paper distinguishes incomplete, inexact, and inaccurate supervision according to label quantity, coarseness, and noise.
  • 6.2 Weakly supervised change detection: Weakly supervised change detection is feasible in geospatial object detection but remains unsatisfactory for remote-sensing change localization.Existing methods may fail to yield accurate change positions, including for building-change detection, motivating more efficient accuracy-improvement methods.

7 Conclusion

The study systematically analyzes deep learning publications in remote sensing and specifically reviews deep learning-based change detection. It categorizes methods into supervised, unsupervised, and transfer-learning groups and identifies two future research directions.

  • 7 Conclusion: The study combines a meta-analysis of deep learning publications with a focused review of deep learning-based remote-sensing change detection.It also describes commonly used deep models and surveys existing change-detection approaches.
  • 7 Conclusion: Deep learning change-detection methods are classified into fully supervised, fully unsupervised, and transfer-learning approaches.The classification organizes the reviewed methods by their learning setting.
  • 7 Conclusion: The paper proposes two promising future research directions for deep learning-based change detection.The directions are deep reinforcement learning and weakly supervised change detection.
Loading 2006.05612v1…