Change detection (CD) and security plays a crucial role in remote sensing applications. The proposed change detection approach focuses on detecting the changes in synthetic aperture radar (SAR) images. The SAR images suffer from speckle noise which affects the classification accuracy. The proposed approach focuses on improving the model's accuracy by removing speckle noise with k-means clustering and an improved threshold approach based on curvelet transform and designing a stacked U-Net model. The stacked U-Net is designed with the help of a 2-dimensional convolutional neural network (2D-CNN). The proposed change detection strategy is evaluated via performing extensive experiments on three SAR datasets. The obtained results reveal that the proposed approach achieves better results than the several state-of-art works in terms of percentage of correct classification (PCC), overall error (OE), and kappa coefficient (KC).
Top1. Introduction
Change detection is a technique for identifying the dissimilarities in the images taken at different times. CD method is applied in various fields like remote sensing, disaster evaluation, medical diagnosis, etc. The change detection scheme's outcomes help provide information in policymaking, area planning, etc. (Kalaiselvi & Gomathi, 2020). The change detection on the earth's surface involves monitoring and comparing two images obtained at different times in the same area. In remote sensing, SAR images acquired for change detection are divided into microwave remote sensing and optical image sensing (Mu et al., 2019). SAR images in remote sensing have advantages over optical images since they can penetrate through clouds and work without light. The optical image resolution is affected by natural factors such as clouds, rain, fog, etc. However, the problem with SAR images is speckle noise (Mu et al., 2019), (Jiang et al., 2007). In conventional approaches, several methods are employed to remove speckle noise. Conventional methods to remove speckle noise with a filter process are spatial adaptive filtering, Lee filter, Pre-test filter, wiener filter, etc. (Ahmed et al., 2010; Chen et al., 2011; Fracastoro et al., 2020; Lee, 1980). These approaches have encountered a loss of resolution and speckle reduction (Fracastoro et al., 2020). The wavelet-based approaches improve the speckle rejection, but problems such as visible filtering artifacts like ringing occur (Fracastoro et al., 2020), (Bianchi et al., 2008). Signal artifacts are observed in methods that model image de-speckling as an optimization problem (Aubert & Aujol, 2008; Bioucas-Dias & Figueiredo, 2010; Shi & Osher, 2008).
Moreover, the conventional approaches to removing the speckle noise cannot be applicable in fields where the number of data is significant. Change detection with the neural network has been a primary area of research in recent years. The deep neural network (DNN) has been widely used in SAR image CD (Minematsu et al., 2017). A deep belief network (DBN) has been utilized to detect the changes in SAR images (Li et al., 2018). In (Keshk & Yin, 2019), an automatic CD method for flood mapping with SAR data with an Expectation maximization-based Gaussian mixture model (GMM) was proposed. Change detection of SAR images with deep learning focused on identifying the changes with the CNN approach (Touati et al., 2019). The neural network approaches have improved the basic models and detected the image changes for enhanced accuracy.
In this research, we proposed a Stacked U-Net where a 2D-CNN is used for SAR change detection. In this approach, SAR images are de-speckled with k means clustering and Curvelet Transform-based threshold method. The de-speckling and change detection are handled separately, so the deep learning method has to focus only on detecting the changes, significantly reducing the complexity. The two-dimensional CNN has been used to learn the features in depth. The contributions of the approach are as follows.
- •
First, the images are pre-processed with k- means clustering and an improved thresholding approach to remove the speckle noise.
- •
Second, the 2D CNN has been used in the encoder part of the Stacked U-Net to obtain better features.
- •
Finally, the experiments on the proposed approach are conducted with various datasets to reveal the proposed strategies' effectiveness on SAR change detection.
The paper is organized as follows. In Section 2, the related works are discussed. The proposed methodology is discussed in Section 3. In Section 4, the results and discussion of the experiments are provided. Finally, the Section ends with a conclusion.