Elsevier

Applied Soft Computing

Volume 71, October 2018, Pages 698-714
Applied Soft Computing

Self-paced stacked denoising autoencoders based on differential evolution for change detection

https://doi.org/10.1016/j.asoc.2018.07.021Get rights and content

Highlights

  • We put forward a self-paced stacked denoising autoencoders model for change detection in radar images.

  • Every training sample is assigned with a weight, then deep network stacked denoising autoencoders is adopted to learn these weighted samples.

  • Self-paced learning is employed for alternately training stacked denoising autoencoders and updating the sample weights.

  • We adopt differential evolution to optimize the pace parameter used in the proposed model.

Abstract

Due to the existence of speckle noise in synthetic aperture radar images, the traditional unsupervised change detection methods do not need any prior information whereas cannot preserve details well. In order to improve change detection performance, change detection methods exploiting supervised classifier have been investigated recently. These methods require reliable labeled samples to train a robust classifier and these samples are always unavailable for image change detection. In this paper, we put forward a novel self-paced stacked denoising autoencoders model to address this issue. In the proposed model, stacked denoising autoencoders are adopted as the supervised classifier, and then self-paced learning is employed to improve it. During iterations, each training sample is associated with a weight and stacked denoising autoencoders are implemented to learn these weighted samples. Furthermore, in the original self-paced learning, it is difficult to determine the pace parameter for acquiring the desired classification performance. Therefore differential evolution is employed to acquire an appropriate pace parameter sequence. Experiments on five real synthetic aperture radar image datasets demonstrate the feasibility and availability of the proposed model. Compared with several other change detection methods, the proposed model is more robust to the speckle noise and can achieve better performance on high resolution synthetic aperture radar images.

Introduction

Synthetic aperture radar (SAR) image change detection aims at distinguishing the changed and unchanged areas in two SAR images of the same region but taken at different times. It has attracted widespread interests in a variety of applications [1], [2], [3], [4], [5], [6]. Because of its stability without affecting by climatic conditions, SAR becomes a popular and significant remote sensing imagery technology. In despite of the inherent existence of speckle noise in SAR images, change detection in multi-temporal SAR images plays an important role in many application fields such as urbanization construction [7], [8], glacier changes monitoring [9], and flood monitoring [10]. A great deal of measures and algorithms have been put forward for SAR image change detection. These methods mainly have the following three steps: image preprocessing, difference image generation, and difference image analysis [11].

The changed and unchanged pixels are distinguished in the process of difference image analysis. The unsupervised schemes include the thresholding methods (Kittler–Illingworth (KI) model [12] and expectation maximization (EM) model [13]), the clustering methods (the fuzzy c-means (FCM) model [14] and the fuzzy local information c-means clustering (FLICM) model [15]), and the level-set methods [16]. Conventional unsupervised change detection methods can be easily implemented, but they usually lack detailed change information and the robustness to speckle noise [17].

In order to preserve geometrical details and be robust to speckle noise, change detection methods using the supervised classifier have been proposed recently. The supervised classifier needs to be trained with labeled samples that are hard to acquire in change detection. Some researchers employed artificial operation to collect the required training samples. For example, Frate et al. chose multilayer perceptron as the classification algorithm for monitoring urban land cover and its changes. The training samples were ground truth information validated by human with heterogeneous information [18]. Artificial operation can guarantee the quality of training samples but takes time and energy. Some researchers utilized traditional unsupervised methods introduced above to collect labeled samples. Feng et al. proposed a change detection method based on neighborhood-based ratio and extreme learning machine (ELM). ELM is adopted as the classifier and hierarchical FCM clustering is designed to collect some labeled samples for ELM [19]. Liu et al. constructed a deep neural network using stacked Restricted Boltzmann Machines for representation learning of difference image, and then labeled samples were acquired through simply sifting the change detection result of traditional unsupervised method [20]. Gong et al. proposed a SAR change detection method without difference image generation step whereas they used joint classification based on FCM for pseudo labels generation [3].

As described above, it is simple and efficient to collect labeled samples by using unsupervised change detection methods. However, there exist many misclassification samples, i.e., unreliable training samples, in the training set. Therefore it is an urgent task to assist the supervised classifier to learn the reliable samples from the initial training set. In this paper, we develop a novel self-paced stacked denoising autoencoders (SPSDAE) model for SAR image change detection. On the one hand, stacked denoising autoencoders with a supervised classifier layer are utilized as the supervised classification algorithm. Denoising autoencoder (DAE) is a variant of the ordinary autoencoder, which is robust to the corrupted input and can learn useful over-complete representations [21]. Deep network built by stacking layers of DAE, named as SDAE, can learn higher-level and abstract representation of input for solving classification problems [22]. On the other hand, we adopt self-paced learning (SPL) to improve SDAE. Different from traditional objective function optimization process in which classifier trains all samples at one time, SPL is used in this paper to help SDAE to learn reliable samples progressively through an iterative optimization process.

SPL is a modified regime of curriculum learning (CL) [23]. CL is inspired by the learning strategy of human in cognitive process, which learns the easier knowledge at first and then gradually learns difficult knowledge. Unlike CL whose curriculum is predefined, SPL simultaneously assigns sample weights and updates the model parameters by optimizing a biconvex objective function. Specifically, SPL adopts a sample weight variable to quantify the difficulty level of samples and an increasing pace parameter to control the sample weight distribution. Besides, SPL has acquired quite good developments in theoretical research. For example, Jiang et al. proposed a self-paced learning with diversity model by considering both easiness and diversity in learning process [24] and later combined CL and SPL called self-paced curriculum learning which took into account both prior information known before training and the learning situation during training [25]. Meng et al. explored the theoretical insight of SPL according to the non-convex regularized penalty [26]. SPL also has many successful applications such as face identification [27], road segmentation [28], and object detection [29].

Chen et al. have proposed a polarimetric SAR image classification method using multilayer autoencoders and self-paced learning [30]. However, in this paper, we combine SPL with SDAE for solving the existing issues in SAR image change detection. Moreover, in current SPL regimes, it is difficult to determine the pace parameter sequence used in the iterative optimization process. In this paper, we adopt differential evolution (DE) algorithm to optimize a pace parameter sequence with a adjustable sequence length. DE can adaptively adjust the gap between two pace parameters to generate a reasonable pace parameter sequence [31].

The contributions of this paper are threefold: (1) In order to address the issue that the training set has some unreliable samples, this paper uses a weight to measure the reliability of the samples and incorporates SPL into SDAE to learn the weighted samples in an iterative manner. (2) In order to automatically generate the pace sequence, this paper adopts DE to acquire the pace parameter sequence instead of using a predefined sequence. (3) The superiority of the proposed SPSDAE model is substantiated by the experiments on five real SAR image datasets.

The rest of the paper is organized as follows. Section 2 introduces the background knowledge of the proposed SPSDAE model as well as the motivations. Section 3 describes the proposed SPSDAE model. Section 4 shows the experimental studies on five real SAR image datasets. Section 5 concludes this paper.

Section snippets

Background knowledge and motivation

Given two co-registered SAR images of the same area: I1 = {I1(a, b)|1 ≤ a ≤ A, 1 ≤ b ≤ B} and I2 = {I2(a, b)|1 ≤ a ≤ A, 1 ≤ b ≤ B} which are taken at the different times t1 and t2 with the same size of A × B, we can acquire a difference image (DI) by log-ratio operator: DI=logI1+1I2+1={DI(a,b)|a{1,2,,A},b{1,2,,B}}. The training sample set is written as {(Di, yi), 1 ≤ i ≤ m}, where m = A × B is the total number of pixels, Di denotes ith sample extracted from the DI and yi ∈ {0, 1}. In this

Methodology

In this section, we will describe the self-paced stacked denoising autoencoders model in detail. First, the proposed SPSDAE model is established for solving change detection problem. Then an iterative majorization minimization algorithm is depicted for solving the SPSDAE model. Finally, DE based pace sequence generation is used to generate a pace parameter sequence.

Experimental study

To demonstrate the feasibility and availability of the proposed method, five high-resolution and real SAR images are investigated in the experiments. In this section, the datasets and the evaluation criteria are introduced in detail. Then the experiments are conducted for evaluating the proposed model. Note that the experimental parameters are described in the supplementary material, such as the neighborhood size, corruption method, and the network architecture.

Conclusion

The change detection methods based on supervised classifier are able to extract detailed change information by learning labeled samples. However, it is difficult to obtain the required high-quality labeled samples in change detection task. In this paper, we put forward a self-paced stacked denoising autoencoders model to address the above issue. Different from the existing change detection methods using supervised classifier, in the proposed model, SDAE is trained with weighted training

References (48)

  • W. Liu et al.

    A novel method of unsupervised change detection using multi-temporal PolSAR images

    Remote Sens.

    (2017)
  • F. Chen et al.

    Differential evolution algorithm with learning selection strategy for SAR image change detection

    Proceedings of Evolutionary Computation

    (2017)
  • Y. Ban et al.

    Multitemporal spaceborne SAR data for urban change detection in China

    IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.

    (2012)
  • O. Yousif et al.

    Improving SAR-based urban change detection by combining MAP-MRF classifier and nonlocal means similarity weights

    IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.

    (2014)
  • A. Muhuri et al.

    Seasonal snow cover change detection over the Indian Himalayas using polarimetric SAR images

    IEEE Geosci. Remote Sens. Lett.

    (2017)
  • J. Avendano et al.

    Flood monitoring and change detection based on unsupervised image segmentation and fusion in multitemporal SAR imagery

    Proceedings of International Conference on Electrical Engineering, Computing Science and Automatic Control

    (2015)
  • L. Bruzzone et al.

    An adaptive semiparametric and context-based approach to unsupervised change detection in multitemporal remote-sensing images

    IEEE Trans. Image Process.

    (2002)
  • A.P. Dempster et al.

    Maximum likelihood from incomplete data via the EM algorithm

    J. R. Stat. Soc.

    (1977)
  • S. Krinidis et al.

    A robust fuzzy local information C-means clustering algorithm

    IEEE Trans. Image Process.

    (2010)
  • H. Li et al.

    A local statistical fuzzy active contour model for change detection

    IEEE Geosci. Remote Sens. Lett.

    (2015)
  • Y. Ban et al.

    Change Detection Techniques: A Review

    (2016)
  • F. Del Frate et al.

    Monitoring urban land cover in Rome, Italy, and its changes by single-polarization multitemporal SAR images

    IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.

    (2008)
  • F. Gao et al.

    Change detection from synthetic aperture radar images based on neighborhood-based ratio and extreme learning machine

    J. Appl. Remote Sens.

    (2016)
  • J. Liu et al.

    Difference representation learning using stacked restricted Boltzmann machines for change detection in SAR images

    Soft Comput.

    (2016)
  • Cited by (19)

    • Deep learning with small datasets: using autoencoders to address limited datasets in construction management

      2021, Applied Soft Computing
      Citation Excerpt :

      This study addresses this knowledge gap by investigating autoencoders as an alternative approach to address transformation variant data. Autoencoders are used for many machine learning applications including dimensionality reduction (e.g. [56]), semantic hashing (e.g. [57]), data denoising (e.g. [58]), clustering and classification (e.g. [59,60]), change detection [61], and demand forecasting [62]. Also, autoencoders have been used to improve model performance by pretraining the models, that is by learning features in the underlying structure of the data to provide good domain generalisation (e.g. [63,64]).

    • Robust unsupervised small area change detection from SAR imagery using deep learning

      2021, ISPRS Journal of Photogrammetry and Remote Sensing
      Citation Excerpt :

      The major components of unsupervised approaches include: 1) image pre-processing, 2) difference image (DI) generation, and 3) analysis of the DI and the classification of pixels into changed and unchanged classes (Li et al., 2020; Wang et al., 2020). The image pre-processing step is used to co-register the SAR images (Lei et al., 2019) and filter speckle noise (Gong et al., 2012b, 2016; Li et al., 2018). In SAR images, speckle noise severely degrades the quality of images, thus affecting the performance of subsequent tasks.

    • α-cut induced Fuzzy Deep Neural Network for change detection of SAR images

      2020, Applied Soft Computing Journal
      Citation Excerpt :

      An accelerated Genetic Algorithm strategy was implemented in parallel by Cai-Hong Mu et al. [40] to analyze and optimize undetermined pixels through search space decomposition. Li, H., Gong [41] reported a stacked denoising autoencoders model with self-paced learning. The major contributions of our paper are as follows:

    • A stacked convolutional sparse denoising autoencoder model for underwater heterogeneous information data

      2020, Applied Acoustics
      Citation Excerpt :

      Meanwhile, a deep denoising autoencoder model was introduced to learn the statistical difference between pure speech and noisy speech, it can effectively improve speech clarity [26]. Based on the consideration of the denoising performance of stacked denoising autoencoder, Li et al. [27] combined adaptive learning with neural network model to complete strong denoising of image data. Furthermore, to avoid the blurred edges and over-enhancement of denoising results, Fan et al. [28] developed SSDA to learn the low-level structure features of the image.

    • A wavelet broad learning adaptive filter for forecasting and cancelling the physiological tremor in teleoperation

      2019, Neurocomputing
      Citation Excerpt :

      Auto-encoder is one of the unsupervised learning algorithm, which learns latent features of samples. It has great performance and potential in signal processing and machine vision [38–41]. However, the diversity of samples makes it remaining a challenging task to learn a component autoencoder and existing auto-encoders lack the ability to extract time-frequency features [42].

    View all citing articles on Scopus

    This work was supported by the National Natural Science Foundation of China (Grant No. 61772393), the National Program for Support of Top-notch Young Professionals of China, and the National Key Research and Development Program of China (Grant No. 2017YFB0802200).

    View full text