Synthetic aperture radar image segmentation using supervised artificial neural network

Abstract

Synthetic Aperture Radar Image Segmentation has been a challenging task because of the presence of speckle noise. Therefore, the segmentation process can not directly rely on the intensity information alone, but it must consider several derived features in order to get satisfactory segmentation results. In this paper, it is attempted to use supervised information about regions for segmentation criteria in which ANN is employed to give training on the basis of known ground truth image derived. Three different features are employed for segmentation, first feature is the original image, second feature is the roughness information and the third feature is the filtered image. The segmentation accuracy is measured against the Difficulty of Segmentation (DoS) and Cross Region Fitting (CRF) methods. The performance of our algorithm has been compared with other proposed methods employing the same set of data.

Keywords

Image segmentation roughness feature artificial neural network noise removal SAR image supervised training

1. Introduction

Synthetic Aperture Radar (SAR) images are obtained by a hovering airplane or satellites in which the images are formed by the active illumination of the ground and the reflection from it. There are several bands of Radar and they differ mainly on the frequency used. Higher frequency bands such as X-band can be used for imaging vegetation whereas lower frequency bands such as C-band are having higher penetrability and can be used for imaging ground surfaces which are covered by vegetation [1]. The main advantage of SAR imaging over optical imaging is their application without regard to natural lighting and weather conditions [2]. So, they are extremely useful for imaging land and water surface topography mapping and also for studies related to object detections. Though the SAR images are very useful but they are usually affected drastically by speckle noise [3, 4, 5]. Speckle noise is resulted by the reflection of an active Radar beam by objects and can be seen in the image as alternate bright and dark spots. SAR images can be classified and differentiated by the number of looks wherein speckle noise effect is more in the case of less number of looks. However, more number of looks sacrifice the information contained in a particular resolution cell of SAR image. So, with enhanced algorithms, actual information can be retrieved more effectively by processing SAR images with less number of looks.

Object detection and different types of surface classification are the two major tasks of SAR image segmentation. Object detection can be achieved by segmenting the foreground object from the background, whereas different area segmentation can be achieved by segmentation based on the difference in the feature space which are potentially extracted from the SAR image. There are a number of image segmentation methods and they are broadly classified into – thresholding, region based, edge based, model based segmentation [6]. The simple thresholding based segmentation technique works by determining the threshold value and the result is a two class or binary segmentation. Region based segmentation starts with seed pixels and conditionally expands to the neighboring pixel until some given criteria is met. Edge based segmentation technique works by determining gradient and thereby finding the edges of regions that are used for segmentation into different regions [7, 8]. The model based segmentation tries to incorporate knowledge from the training data and then formulate mathematical formulas by approximating the known data.

In this work, an attempt is made to incorporate the capability of Artificial Neural Network (ANN) for supervised segmentation of SAR images. Since the intensity information alone can not be used for image segmentation as the individual numerical value does not possess sufficient information about the actual segmentation result that the area pertains. Apart from the above, we attempted to use other features such as local entropy and median filtered image and test the accuracy of segmentation for improved result of segmentation. The use of ANN requires the number of neurons in the input layer equals the number of input pixels that results in impractical usage of ANN. In this paper, the choice of using ANN requires a minimal set of input features, and in our example, it requires only three input neurons that are very light to use and practical also.

2. Literature review

Several attempts were made for SAR image segmentation exploiting different approaches. Since the segmentation which is based only on the intensity value of the SAR image is infeasible because the pixel values barely attain features for a region, i.e., it usually results in over-segmentation. Effects of noise can be reduced by applying different low-pass spatial filters but these filters drastically reduce the information contained in each of the pixel cells. Yin and Yang [6] proposed to apply a level set method on SAR image segmentation by defining regions using contours. Though the result was satisfactory enough for multilook SAR images but it does not perform well on single look SAR images which are highly affected by speckle noise. Markov Random Field was recently applied for SAR image segmentation by Doulgeris et al. [10] and also by Liu et al. [11] and their results were satisfactory for object recognition purpose but performs very poor for speckle affected images. Clustering methods were also applied for segmentation in which similar or close pixels were grouped together to form clusters [12, 13]. There was also an option of coarse to fine grained hierarchical clustering based on the measure of homogeneity of different clusters. Fuzzy C-means (FCM) is a popular algorithm used for image segmentation based on clustering in which an objective function is minimized by updating fuzzy membership matrix and also cluster centers in each iteration and is used by [14, 15]. Though it is a good algorithm and results are good, huge computation time required for iteration is an issue. Chen and Zhang [16] proposed FCM S1 and FCM S2 in order to compensate for high computation involved in the original FCM. Their proposed algorithm utilizes local mean and median filtering techniques for speeding up the clustering method. Wan et al. [17] proposed robust SAR image segmentation based on FCM that utilizes the non-local spatial information. However, the downside of the aforementioned proposals was the requirement of setting different parameters manually in order to maintain effective suppression of noise and preservation of the actual content of the original SAR image. Krinidis and Chatzis [18] came up with a fuzzy local information C-means clustering algorithm (FLICM) that does not require manual parameter setting by adding a neighbourhood term to the objective function and automatically calculating the spatial distance between pixels in the neighbourhood. Singha et al. [19] proposed ANN based oil spill detection from SAR image in which the image segmentation was done in three different ways, such as segmentation using ANN, edge detection segmentation and adaptive thresholding. Their proposed feature extraction method works on the segmented images and 14 features were extracted for ANN based classification for oil spill detection. State of the art segmentation using application of deep learning methods were also proposed. Yaohua and Xudong [20] proposed the application of Densenet CNN for Oil spill image detection. Pai et al. [21] proposed Automatic Segmentation of River and Land in SAR Images using deep learning approach in which they proposed to use U-Net architecture of CNN in their work and several such works were carried out on SAR image segmentation [22, 23]. The main problem in using deep learning methods for SAR image segmentation is the requirement of large datasets for training without which the network could not perform well. Another downside is the requirement of resources for training in terms high computing power, space, and time consumption.

So, considering the requirements for SAR image segmentation, we proposed to contribute the followings through this paper:

•
Possibility and effectiveness of shallow ANN is studied.
•
Requirement of large dataset required for segmentation problem using Deep Learning is an issue especially in SAR image. So, minimal available SAR dataset has been incorporated with its own derived features to test its effectiveness.

The paper is organised as follows: the following section confines on the proposed methodology and Section 3 discusses the the implementation of the proposed method. Section 4 is about the discussion of the experimental results and performance comparison. The last section is dedicated for the conclusion and future scope of the proposed methodology.
3. Proposed method

The proposed methodology tries to combine different methods already available such as extraction of different features and then combine those features using ANN. The ANN we proposed to use here is a shallow perceptron network having single hidden layer of neurons. Though shallow ANNs are barely used for learning purposes, we proposed here because of the possible limited number of training images and also considering the resource constraint in terms of computing power, space requirements and training time. Again, the reason for choosing different features is that when the raw SAR image is to be segmented, the intensity information alone would not be sufficient to distinguish different regions in a SAR image. In order to have derived features based on the average intensity values of a particular area, segmentation of different regions would improve because of smoothing out speckle noise. However, doing so would hugely deteriorate the information contained in the original SAR image. It is also known that different regions in SAR images can be effectively segmented on the basis of local randomness of a pixel cell on the basis of knowledge that different surfaces have different reflectivity and would also appear different to SAR sensors [24]. With that being said, it would again deteriorate the actual content of the SAR image. So, with our proposed methodology, different features of SAR images would be extracted and then fed to the ANN for training. The trained network would then be used for automatic segmentation of SAR image into different trained regions. The advantage of the use of shallow ANN is its lightweight nature and requirement of less amount of training data to achieve acceptable segmentation result.

3.1 Feature extraction

In this paper, it is proposed to use 3 features for the purpose of segmentation. The first feature is the original image and the reason for choosing the original image is that it contains the information about the actual content of a pixel cell. The second feature is the median filtered image which could be either in $3\times 3$ , $5\times 5$ or maybe higher dimension also. The choice of smaller dimension kernel removes noises which are of less prominence whereas higher dimension kernels remove more prominent edges or noise. Tradeoff can be made on the choice of dimension of spatial filters on the basis of what kinds of regions to be segmented. The third feature that we may consider is the roughness information of SAR image which could be in the form of entropy.

3.1.1 Low-pass smoothing filter

The reason for choosing low-pass smoothing filter for SAR image is that due to speckle noise effect, the image appears as alternate dark and bright spots [25]. When viewed from a distance different regions appear different from the others and can be visually distinguishable. But when it comes to segmentation on the basis of the intensity information alone, oversegmentation occurs [26]. So, in order to avoid false segmentations, the image needs to be filtered using low-pass filters. There can be a choice of having different low-pass filters such as mean filters, median filters and mode filters. There are also Butterworth, Gaussian low pass filters as well. Usage of a smaller kernel results in a smaller window for convolution and speeds up the computation. Whereas a large window kernel is likely to remove small edges while giving more importance to more prominent edges. Low pass filters are good for producing average intensity values in the area but also remove sharp edges and usually produce blurred boundaries. The convolution operation is used for spatial filters as below:

$\displaystyle O=I\times M$ (1)

where $O(r,c)$ is output, $I(r,c)$ is input, and $M(r,c)$ is mask.

3.1.2 Roughness feature extraction using local entropy

Entropy is the measure of impurity, disorder or uncertainty in a bunch of dataset. In a situation where an image is to be segmented but contaminated by noise severely, determination of the segments may not be easy. So, determining the different regions in an image may be done on the basis of randomness. In a SAR image, there could be different types of reflection properties in which different types of region provide different pattern of reflections [27]. For instance, water bodies reflect less than forest regions and there would normally be less randomness in the case of water than in forest regions. Using this difference, we could determine the randomness of a region and use this information to segment the regions.

The local entropy can be determined using the following formula:

$\displaystyle\textit{Entropy}=-\sum p(X)\log_{2}p(X)$ (2)

where $p(x)$ is the probability of an event $x$ happening, and $x$ is the random variable.

Whereas it is also possible to use other statistical models to calculate texture and randomness, this particular feature provides more differentiation of the model under consideration, i.e., to segment forest cover areas, water bodies and settlement areas.

Since the end product of SAR image is basically a bitmap grayscale image contaminated by speckle noise that appears in the form of salt and pepper effect. So, the correlation between the actual pixel value and the actual targeted segment is very low. Therefore, it is required to have another feature that is not only the pixel intensity value, but possesses the possibility that it differentiates on that basis. So, it is proposed here to obtain the roughness feature for each of the pixels in the image.

3.2 Artificial neural networks

Artificial Neural Networks (ANN) is the soft computing technique that mimics the working of human brain [28]. In a Biological Neural Network (BNN), signal is passed from one neuron to another neuron through the synapses and axons. The brain learns by adjusting the weights of connection, i.e., synapses. When connection is strong between neurons, more information can pass through between them. The ANN is derived from the working principle of BNN wherein the neurons are represented by nodes, and the synapses weights are represented by weighted edges [29]. In practical application, inputs are given from one layer and outputs can be received by sum of product of inputs and weights, and by applying the activation function on the output.

In this paper, three features are proposed for image segmentation purposes and they are the original intensity, roughness information and the average intensity. Though more number of features can be selected, it is avoided here because it would increase computation time. The selection of the original intensity image is that it retains the original information in the image. As was mentioned earlier, this would not be sufficient as it carries very less information about the segment where it actually belongs. SAR images are heavily contaminated with speckle noise, so, eventually it exposes the nature of the surface that reflects the signal transmitted by the Radar system. Water bodies are known to reflect less amount of signal as compared to other types of surface like forests. The randomness of signals reflected by these surfaces would also be less. Human settlements have materials that reflect more amounts of signals and also more back-scatterings that would make the pixel cell brighter. Also that would also increase the speckle effect and there are likely more alternate black and dark spots increasing the entropy or randomness in that area. So, the average brightness and the local entropy in the area that we want to segment would likely be having more differentiability.

So, the backpropagation perceptron ANN is designed to take three inputs and one output. Two of the inputs namely, the original image and the filtered image are expected to have the range 0–255 as they are in the form of grayscale image. On the other hand, the local entropy for the test images have the range from 0 to 4.7 and hence, it has to be normalized in the range 0–255. Target output is a single neuron output in the range 0–255. Depending on the surfaces that we want to classify, we may set the target values as 0 and 255 for two region segmentation, 0, 127 and 255 for three regions segmentation and so on. For the sake of performance, the mean is shifted towards 0 by subtracting 128 from the normalized input.

4. Implementation

The implementation is described as follows: First, a known area is sliced off and each feature is extracted from it, namely roughness feature in the form of local entropy and smoothness feature in the form of applying smoothing filter, apart from the original image. The experiment uses a dataset from [30] which is a single look image data in HH polarimetry. The image consists of settlement area, forest or vegetation cover area and water body. From the image, rectangular regions are selected for training purposes and are shown in Fig. 1a. The corresponding google earth image is also given in Fig. 1b. The steps of implementation is shown in Fig. 3.

Figure 1.

a. Single look SAR image and areas showing known areas, b. Google image of the same area.

Figure 2.

a–c. Areas of interests, d–f. Local entropy images, g–i. Filtered images.

Figure 3.

Steps followed in the proposed methodology for training and testing.

4.1 Feature extraction

From each of the different areas extracted from the source image in Fig. 1a, the selected regions for training and the first features are shown in Fig. 2a–c. The roughness features for each of them are calculated and normalized. The calculated and scaled local entropy is given in Fig. 2d–f for each of the regions.

The filtered image is determining the average brightness in the region thereby allowing the differentiation amongst different surface covers. The filtered image of each of the regions are given in Fig. 2g–i.

4.2 Training of ANN

Before training the network, the boundary pixels are deleted in order to remove possible anomalies due to padding effect. For each of the filters used, the amount of edge pixels to be removed can be determined as per the following equations. Padding requirement depends on the size of the filters used.

$\displaystyle p_{r}=\textit{floor}\left(\frac{f_{r}}{2}\right)$ (3) $\displaystyle p_{c}=\textit{floor}\left(\frac{f_{c}}{2}\right)$ (4)

where $f_{r}$ is the number of rows and $f_{c}$ is the number of columns of the spatial filter used, $p_{r}$ and $p_{c}$ are the number of paddings required to get the same size output of the filters. The same number of rows and columns are affected by the paddings, so they need to be removed from the training set in order to reduce inconsistent training data.

Training is done for the three features and targets are set such that each region targets are set apart as far as possible. For the training, the Lavenberg-Marquardt algorithm is used. As expected, the training algorithm does have difficulty in convergence due to the speckled nature of the SAR data. However, when the training is conducted, it somehow takes inputs from the combined features and gives results which are satisfactory as compared to the recent proposed methods.

5. Experimental results and performance evaluation

The actual performance evaluation against the ground truth of real SAR images is infeasible because of the difficulty of obtaining the ground truth and the non-deterministicity of the problem. However, visual judgement plays an important role in the accuracy of the solutions. Braga et al. [31] proposed the use of Arithmetic-Geometric distance of the segmented areas for performance evaluation. The Arithmetic-Geometric distance measure is used for finding the Error of Segmentation (EoS) as well as Difficulty of Segmentation (DoS) parameters that signifies the performance.

Their proposed Arithmetic-Geometric distance is given as:

$\displaystyle S_{AG}=\frac{1}{2}\int_{R^{+}}(f_{G_{A}^{0}}(z_{i},\alpha_{i},y_% {i},n)+f_{G_{A}^{0}}(z_{j},\alpha_{j},y_{j},n)){}\times\log\left(\frac{f_{G_{A% }^{0}}(z_{i},\alpha_{i},y_{i},n)+f_{G_{A}^{0}}(z_{j},\alpha_{j},y_{j},n)}{% \sqrt{f_{G_{A}^{0}}(z_{i},\alpha_{i},y_{i},n)\times f_{G_{A}^{0}}(z_{j},\alpha% _{j},y_{j},n)}}\right)dz$ (5)

Applying the similar idea, the performance can be evaluated for our proposed method.

Assume that the segmentation aims at $n$ number of region segmentation, then let the regions be defined as $\Omega_{1},\Omega_{2},\ldots,\Omega_{n}$ and let the scaled histogram $H_{i}$ are in region $\Omega_{i}$ . The distance between region $\Omega_{i}$ and region $\Omega_{j}$ can be given as a set of distance between each pair of histogram intensity belonging to region $\Omega_{i}$ and region $\Omega_{j}$ respectively as follows:

$\displaystyle D(\Omega_{i},\Omega_{j})=\sqrt{\sum{(H_{i}-H_{j})}^{2}}$ (6)

After defining the distance between two regions, the following terms can be derived, which are Difficulty of Segmentation (DoS) and Cross-region Fitting (CRF). DoS is quantified as the difficulty of segmentation between two regions $\Omega_{i}$ and $\Omega_{j}$ and it depends on the contrast between regions as follows:

$\displaystyle\textit{DoS}(\Omega_{i},\Omega_{j})=\frac{1}{D(\Omega_{i},\Omega_% {j})}$ (7)

Cross Region Fitting (CRF) combines the segmentation distance and the DoS and is used for quantifying the ability to segment the regions correctly. It is given as:

$\displaystyle\textit{CRF}=\frac{1}{1+\sqrt{\textit{DoS}\times|D(\Omega_{i}^{r}% ,\Omega_{j}^{s})-D(\Omega_{i}^{s},\Omega_{j}^{r})|}}$ (8)

where $\Omega_{i}^{r}$ is the referenced region $i$ and $\Omega_{i}^{s}$ is the segmented region $i$ .

The proposed method has been tested with real SAR images which are captured in HH polarimetry and are single looked images. They are shown in Fig. 4 along with their corresponding google images for visual understanding of the images.

Table 1

Performance evaluation: Difficulty of Segmentation (DoS)

Image tag	Original image			Proposed algorithm
	$\textit{DoS}_{1,2}$	$\textit{DoS}_{2,3}$	$\textit{DoS}_{1,3}$	$\textit{DoS}_{1,2}$	$\textit{DoS}_{2,3}$	$\textit{DoS}_{1,3}$
Image 1	2.09	2.18	2.10	1.96	2.11	2.02
Image 2	2.04	2.12	1.98	1.96	2.16	2.01
Image 3	2.14	2.14	2.10	1.93	2.13	1.99
Image 4	1.97	2.17	1.97	1.97	2.10	1.99
Image 5	2.10	1.97	1.92	2.01	2.15	2.05
Image 6	1.99	2.17	2.01	1.74	1.96	1.84

Table 2

Performance evaluation: Cross Region Fitting (CRF)

Image tag	Original image			Proposed algorithm
	$\textit{CRF}_{1,2}$	$\textit{CRF}_{2,3}$	$\textit{CRF}_{1,3}$	$\textit{CRF}_{1,2}$	$\textit{CRF}_{2,3}$	$\textit{CRF}_{1,3}$
Image 1	0.81	0.85	0.93	0.81	0.93	0.81
Image 2	0.78	0.77	0.83	0.77	0.93	0.78
Image 3	0.92	0.79	0.74	0.78	0.85	0.77
Image 4	0.71	0.82	0.76	0.77	0.90	0.78
Image 5	0.84	0.66	0.66	0.83	0.91	0.91
Image 6	0.76	0.91	0.80	0.76	0.79	0.71

Table 3

Comparative accuracy with other methods in the literature

Image tag	MLS [32]	MoLC [9]	MoRE [24]	Proposed method
Image 1	0.7743	0.3708	0.7644	0.8717
Image 2	0.5679	0.5363	0.5872	0.7850
Image 3	0.4820	0.6180	0.6474	0.5019
Image 4	0.9039	0.6548	0.7582	0.8746
Image 5	0.7843	0.7006	0.7276	0.7923
Image 6	0.7025	0.5761	0.6970	0.7651

Figure 4.

Each rows are different SAR images numbered from 1 to 6. a. Original SAR image, b. Corresponding google scene, c. Manually segmented ground truth, d–f. Output of segmented image proposed by Braga et al. [32], Rogrigues et al. [9] and Nobre et al. [24] respectively, g. Our proposed method.

The performance of the algorithm has been measured for different SAR images and against the original image using Difficulty of Segmentation (DoS) and Cross Region Fitting (CRF). As can be seen in Tables 1 and 2, the output of the program offers less DoS as well as improved CRF.

Moreover, the algorithm is compared with other algorithms that work on the same set of data. The proposed methods used for comparison of performance are a median regularized level set for hierarchical segmentation of SAR images proposed by Braga et al. [32] which they named it as Median-regularised Level Set (MLS) method, SAR Image Segmentation Using the Roughness Information proposed by Rodrigues et al. [9] in which they used Method of Log Cumulants (MoLC) method for parameter estimation and SAR image segmentation with Renyi’s entropy proposed by Nobre et al. [24] and their proposed method relies on Matrix of Renyis Entropy (MoRE). The dataset has been provided by Nobre et al. [30]. The segmentation output of the proposed mentioned methods along with our proposed method are shown in Fig. 4. The accuracy of each algorithms are also shown in Table 3.

6. Conclusion

The paper proposed a shallow multilayer perceptron ANN for different areas segmentation of SAR image. The proposed method takes the advantage of shallow ANN and possibly effective features, i.e., original SAR image, roughness information in terms of local entropy and smoothness value in terms of median filtering. The main advantage of our method being non-requirement of large dataset which is usually required in the case of deep learning and also the simplicity of the algorithm. Also as compared to the methods highlighted in Table 3, our proposed method required much less time in the order of 1 minute as it does not require parameter estimation which is required in each of the previously proposed methods and it might take approximately 7–10 minutes. Though quite being simple, it yields very good result which is comparable to other state of the art algorithms proposed for SAR image segmentation of different surface cover areas. The method can further be improved with the use of larger dataset with more hidden layers added to the ANN or by the use of Convolutional Neural Networks.

Footnotes

Author’s Bios

R. Lalchhanhima completed his Bachelor of Engineering degree in Computer Science and Engineering from Coimbatore Institute of Technology and Master degree from Bits-Pilani. He currently works in the Department of Information Technology Department, Mizoram University as Assistant Professor. Prior to joining the academic line, he worked in the Information and Communication Technology which is a State Government department for a period of 5 years. He is currently pursuing Ph.D from the Department of Information Technology, North Eastern Hill University, Shillong, Meghalaya.His area of research is SAR image Segmentation and Soft Computing.

	Goutam Saha is currently holding a post of Professor in the Department of Information Technology, of North Eastern Hill University (NEHU), Meghalaya. He has done B.E. (Electrical Engineering) and M.E. (Electronics and Telecommunication Engineering) both under University of Calcutta. He received his Ph.D degree from IIT Kharagpur and he also completed his Post Doctoral Research work from Ben-Gurion University, Israel. Prior to joining the University, he holds several technical posts in different companies and his work career starts from 1985. He published lots of scientific papers and is actively involving in teachings and research. Presently, he is carrying out research work on the aspects of Genetic Network Modelling and Drug Design.
	Debdatta Kandar Debdatta Kandar: Born in 1977 at Deulia, Kolaghat, Purba Medinipur, West Bengal, INDIA. He has 18 years (approx.) of professional experience. Currently holding the post of Associate Professor in the department of Information Technology, North-Eastern Hill University, Shillong, India. Before joining NEHU, he worked at S.K.P. Engineering College, Tiruvannamalai, Tamil Nadu and Sikkim Manipal Institute ofTechnology, Sikkim, India. He worked in a DRDO sponsored project and he has several years of industry experience too. He has successfully guided several PhD, M. Tech and BTech students. Four of his students already awarded PhD degree. He has national and international collaborations for carrying out research works. He has delivered talks in different workshop, seminar, FDP etc. He successfully organized several national and international Seminar, Conferences etc. He also chaired several sessions in different conferences. He has published approximately 90 research papers in different national and international journals, conference proceedings and book chapters. He also published edited book etc. He has completed his PostDoc from Department of Electrical and Electronic Engineering
Technology, University of Johannesburg, Johannesburg, South Africa. He received his PhD (Engg.) from Department of electronics and Telecommunication Engineering, Jadavpur University, Kolkata, West Bengal, India, in the year 2011. He has been awarded ‘Young Scientist’ award from Union Radio Science International (URSI GA-2005) at Vigyan Bhaban, Delhi, for his research work. President of India, Dr. A.P.J. Abdul Kalam invited him at his residence on that occasion. His research interests include Wireless Mobile Communication, Artificial Intelligence, Soft Computing and Radar Operation etc.
	Morrel V.L. Nunsanga received his M.Tech (IT) degree from Tezpur University and BE (CSE) Degree from Dr Ambedkar Institute of Technology, Bangalore. He was working as System Administrator in the Department of Information and Communication Technology under Govt of Mizoram untill he joined Mizoram University as Assistant Professor in the Department of Information Technology in the year 2010. He is a member of Mizo Academy of Science (MAS) and he is also presently pursuing PhD from Mizoram University.

References

Dey

Zhang

and Zhong

, A review on image segmentation techniques with remote sensing perspective, Vienna, 2010.

Tomiyasu

, Tutorial review of synthetic-aperture radar (sar) with applications to imaging of the ocean surface, Proceedings of the IEEE 66(5) (1978), 563–583.

Sivakumar

M.V.K.

Roy

P.S.

Harmsen

and Saha

S.K.

, Satellite remote sensing and GIS applications in agricultural meteorology, Proceedings of the Training Workshop in Dehradun, India, AGM-8, WMO/TD, 2004.

Lang

Yang

Shi

and Wei

, Mean-shift-based speckle filtering of polarimetric sar data, IEEE Transactions on Geoscience and Remote Sensing 52(7) (2014), 4440–4454.

Zhang

Liu

Jiao

Hou

Wang

and Shang

, Sar image despeckling using edge detection and feature clustering in bandelet domain, IEEE Geoscience and Remote Sensing Letters 7(1) (2010), 131–135.

Yin

and Yang

, A modified level set approach for segmentation of multiband polarimetric sar images, IEEE Transactions on Geoscience and Remote Sensing 52(11) (2014), 7222–7232.

Taneja

Ranjan

and Ujjlayan

, A performance study of image segmentation techniques, in: Reliability, Infocom Technologies and Optimization (ICRITO) (Trends and Future Directions), 2015 4th International Conference, 2015, pp. 1–6.

Kong

and Zhou

, A new method of sar image reconstruction and segmentation, in: 2009 International Asia Conference on Informatics in Control, Automation and Robotics, IEEE, 2009, pp. 249–253.

Rodrigues

F.A.A.

Neto

J.R.

Marques

R.C.P.

Medeiros

F.N.S.

and Nobre

J.S.

, Sar image segmentation using the roughness information, IEEE Geoscience and Remote Sensing Letters 13(2) (2016), 132–136.

10.

Doulgeris

A.P.

, An automatic-distribution and markov random field segmentation algorithm for polsar images, IEEE Transactions on Geoscience and Remote Sensing 53(4) (2015), 1819–1827.

11.

Liu

Luo

Loy

and Tang

, Deep learning markov random field for semantic segmentation, IEEE Transactions on Pattern Analysis and Machine Intelligence 40(8) (2017), 1814–1828.

12.

Xiang

Tang

and Su

, A kernel clustering algorithm with fuzzy factor: application to sar image segmentation, IEEE Geoscience and Remote Sensing Letters 11(7) (2014), 1290–1294.

13.

Shang

Tian

Jiao

Stolkin

Feng

Hou

and Zhang

, A spatial fuzzy clustering algorithm with kernel metric based on immune clone for sar image segmentation, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 9(4) (2016), 1640–1652.

14.

Tong

and Cheng

, Image segmentation using fcm optimized by quantum immune clone algorithm, in: 2014 Fifth International Conference on Intelligent Systems Design and Engineering Applications, IEEE, 2014, pp. 537–540.

15.

Seljuq

, Synthetic aperture radar (sar) image segmentation by fuzzy c-means clustering technique with thresholding for iceberg images, Computational Ecology and Software 4(2) (2014), 129–134.

16.

Chen

and Zhang

, Robust image segmentation using fcm with spatial constraints based on new kernel-induced distance measure, IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) 34(4) (2014), 1907–1916.

17.

Wan

Zhang

Xiang

and You

, A robust fuzzy c-means algorithm based on bayesian nonlocal spatial information for sar image segmentation, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 11(3) (2018), 896–906.

18.

Krinidis

and Chatzis

, A robust fuzzy local information c-means clustering algorithm, IEEE Transactions on Image Processing 19(5) (2010), 1328–1337.

19.

Singha

Bellerby

T.J.

and Trieschmann

, Satellite oil spill detection using artificial neural networks, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 6(6) (2013), 2355–2363.

20.

Yaohua

and Xudong

, A sar oil spill image recognition method based on densenet convolutional neural network, in: 2019 International Conference on Robots & Intelligent System (ICRIS), IEEE, 2019, pp. 78–81.

21.

Pai

Mehrotra

Aiyar

Verma

and Pai

, Automatic segmentation of river and land in sar images: A deep learning approach, in: 2019 IEEE Second International Conference on Artificial Intelligence and Knowledge Engineering (AIKE), IEEE, 2019, pp. 15–20.

22.

Hang

Ding

Jiang

M.H.

and Kot

A.C.

, Deepimaging: A ground moving target imaging based on cnn for sar-gmti system, IEEE Geoscience and Remote Sensing Letters, 2020, pp. 1–5.

23.

Natteshan

N.V.S.

and Kumar

N.S.

, Effective sar image segmentation and classification of crop areas using mrg and cdnn techniques, European Journal of Remote Sensing (2020), 1–15.

24.

Nobre

R.H.

Rodrigues

F.A.A.

Marques

R.C.P.

Nobre

J.S.

Neto

J.F.

and Medeiros

F.N.S.

, Sar image segmentation with renyi’s entropy, IEEE Signal Processing Letters 23(11) (2016), 1551–1555.

25.

Liu

and Cheng

, Polarimetric sar speckle filtering for high-resolution sar images using radarsat-2 polsar slc data, in: Computer Vision in Remote Sensing (CVRS), 2012 International Conference on, IEEE, 2012, pp. 329–334.

26.

Xie

Pierce

L.E.

and Ulaby

F.T.

, Sar speckle reduction using wavelet denoising and markov random field modeling, IEEE Transactions on Geoscience and Remote Sensing 40(10) (2002), 2196–2212.

27.

Jin

Yin

Zhou

and Yang

, Level set segmentation algorithm for high-resolution polarimetric sar images based on a heterogeneous clutter model, IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 10(10) (2017), 4565–4579.

28.

Atkinson

P.M.

and Tatnall

A.R.L.

, Introduction neural networks in remote sensing, International Journal of Remote Sensing 18(4) (1997), 699–709.

29.

Mohamad

and Hassoun

, Fundamentals of artificial neural networks, MIT press, 1995.

30.

Nobre

R.H.

Rodrigues

F.A.A.

Rosa

Medeiros

F.N.

Feitosa

Estevão

A.A.

and Barros

A.S.

, Grss sar/polsar database, 2017.

31.

Marques

R.C.P.

Medeiros

F.N.

and Nobre

J.S.

, Sar image segmentation based on level set approach and ga0 model, IEEE Transactions on Pattern Analysis and Machine Intelligence 34(10) (2012), 2046–2057.

32.

Braga

A.M.

Marques

R.C.P.

Rodrigues

F.A.A.

and Medeiros

F.N.S.

, A median regularized level set for hierarchical segmentation of sar images, IEEE Geoscience and Remote Sensing Letters 14(7) (2017), 1171–1175.