Keywords

1 Introduction

The value of information/intelligence obtained is time-bound. Image interpretation from huge remote sensing data by qualified Image Analysts is time consuming in the absence of reliable automatic analysis tools. Hence an objective approach to the fundamental image analysis problem is rather an intelligent decision-making process with a required degree/probability of success. The observation of maritime activity is the need of time which has improved since inception of synthetic aperture radar (SAR) imagery from space and aerial platforms. The information derived from remote sensing imagery provides decision-makers, the valuable time with automatic methods to ease the human image interpretability.

1.1 Objective

The objective is to study the usability of microwave remote sensing in the detection of ships and evaluate the potential of SVMs in improving the semi-automatic detection accuracy of ships. The research limits use of TerraSAR-X spotlight image, ERDAS Imagine, and MATLAB for analysis.

2 Background

Microwave SAR remote sensing offers cost-effectiveness with better efficiency and flexibility for the identification of ship under all weather conditions. Large amounts of image data generated by SAR systems can quickly overburden a human observer. Recent developments in this field have introduced different nonlinear procedures and techniques of statistical learning [1–5]. Using statistical learning techniques [6–8], SVMs have reached the minimum of the upper bound in the error of probability of a classifier. This paper discusses a robust method of image analysis for visualization and classification of image using SVM to assess data toward detection of ships and ascertain the accuracy of feature detection in proposed method.

3 Study Area and Datasets Used

VHR SAR images (Fig. 1) of Visakhapatnam Naval Port, Andhra Pradesh, India have been used as sample images of TerraSAR-X spot light mode for study of a robust method by using SVMs for the image classification and the corresponding Electro Optical (EO) image by Google Earth (Fig. 2).

Fig. 1
figure 1

VHR SAR imagery by TERRA SAR

Full size image
Fig. 2
figure 2

EO image by Google earth

Full size image

TerraSAR-X is specifically optimized to meet the requirements of commercial users around the globe, who require, high-quality and precise Earth observation data readily available. Image specification is as follows (Table 1):

Table 1 Dataset specification
Full size table

4 Methodology

The architecture of the proposed enterprise level, methodology used for the image analysis and implementation of the Intelligent Image Interpreter (Fig. 3) is restricted to the conceptualization. The initial part of the implementation was carried out by ERDAS Imagine and SVMs train and classification in MATLAB.

Fig. 3
figure 3

Flow chart of the proposed methodology and implementation of intelligent image interpreter

Full size image

4.1 Implementation of Algorithm

The following steps were utilized for implementation: The database is codified as two classes Ship (+1) and Water (–1). Two sets of training images for ship pixel and water pixels were selected in a folder/container. Read Training Images from the folder/container automatically (two each for each class in this case). SVM train. Test Image formation for ship detection from the same folder/container automatically. SVM Classify. Classified Image formation. Accuracy Assessment.

5 Results and Analysis

For this particular SAR image, the quality of the speckle suppression at the trade-off at smoothening the image for the purpose of the interpretation the Lee sigma filter is better than other because it has not only reduced the speckle but also highlighted the other features in surroundings. The semi-automatic implementation using MATLAB (Statistics and Machine Learning toolbox) has revealed that SVMs of using linear, quadratic basis revealed more error rate in classification when compared to the Gaussian radial basis function and multi layer perceptron models (Fig. 4). Hence it is deduced that while RBF produces greater accuracy with higher time cost which can be reduced by cross-validation quadratic gives a robust accuracy in lesser time (Fig. 5). It is interesting to note that linear function basis gives better visual appreciation of the finer details including portions of ever much smaller potential ship pixels; while MLP gives a clear indication of the ships by eliminating even portions of the jetty. Also RBF gives higher accuracy with reduced probability of false alarm and hence modification of the constraint, sigma, and with cross-validation, it will produce the best results. Finally, the Quadratic function basis gives the robust accuracy making it useful for detection and for further analysis it is the RBF which will effectively give the best results for learning about the target of interest for future detection.

Fig. 4
figure 4

SVMs classified images a Linear function. b Quadratic function. c Gaussian radial basis function. d Multi layer perceptron function

Full size image
Fig. 5
figure 5

Accuracy assessment

Full size image

6 Accuracy Assessment

The number of ships identified on the processed image with the original image and assessment is tabulated.

7 Discussion

This section discusses some solutions that may help improve this method further, so that detection of naval features is carried out more accurately. The problem handled in this paper is a generic detection of ships in sea. The trade-offs have to be formulated to match the aim of the user specific to area of interest. Considerations of the Target include type, class, size, and velocity of target to be detected in the given user area. Environmental considerations of surveillance area include sea state and wind conditions. Finally time constraints and revisit rate against available computing resources is the bottleneck for decision-maker. Considerations of radar include frequency, resolution, incidence angle, and polarization [1]. The radar parameters would be chosen to match the user target and area of interest. Therefore taking all the considerations the ship detection algorithms are designed to match the data used for analysis. This paper is aimed to provide base to choose the best SVM techniques for ship detection in SAR imagery.

7.1 Human Supervision

As the importance of intelligence is time sensitive for the decision-maker, it is vital to include expert to reduce to false detections. This semi-automated algorithm’s capability can be enhanced by the human supervision.

8 Conclusion

This paper presents improved methods for semi-automatic naval feature extraction using microwave remote sensing. For this purpose, Erdas Imagine software was used. This software was explored to define the best filters available for processing the microwave image. In this study, very high-resolution image of Terra SAR X is used and by using this image and by applying SVMs-based analysis of images for the semi-automatic interpretation of ships is explored.

8.1 Future Scope

SVM selection and parameter optimization using cross-validation of fit and avoidance of the over/under fitting to optimize the classification with trade-off between computational and time complexity for a given resources. Ship constructed with material like small wooden boats and fiber glass boats may also be picked up by the microwave imagery. Object-based image analysis may be carried out on the processed image for image segmentation for better extraction of naval features. Capability enhancement of ship detection algorithms can be done using of circular transmit, linear receive (CTLR) mode of RISAT 1 SAR data.