Loading...
Please wait, while we are loading the content...
Similar Documents
Evolutionary Approach for Detection of Buried Remains Using Hyperspectral Images
| Content Provider | Semantic Scholar |
|---|---|
| Author | Dozal, León Silván-Cárdenas, José Luis Moctezuma, Daniela Siordia, Oscar Sánchez Naredo, Enrique |
| Copyright Year | 2018 |
| Abstract | Hyperspectral imaging has been successfully utilized to locate clandestine graves. This study applied a Genetic Programming technique called Brain Programming (BP) for automating the design of Hyperspectral Visual Attention Models (H-VAM.), which is proposed as a new method for the detection of buried remains. Four graves were simulated and monitored during six months by taking in situ spectral measurements of the ground. Two experiments were implemented using Kappa and weighted Kappa coefficients as classification accuracy measures for guiding the BP search of the best H-VAM. Experimental results demonstrate that the proposed BP method improves classification accuracy compared to a previous approach. A better detection performance was observed for the image acquired after three months from burial. Moreover, results suggest that the use of spectral bands that respond to vegetation and water content of the plants and provide evidence that the number of buried bodies plays a crucial role on a successful detection. Introduction Locating unmarked graves represents a complicated and timeconsuming forensic problem because their locations are often remote and the burial time is generally unknown (Siegel and Saukko, 2013). The research on the detection of clandestine graves through multi and hyperspectral images is incipient, yet has proven to be one of the most challenging forensic problems. This is an important area of work, since airborne hyperspectral data enable searching over a large area that is otherwise inaccessible by foot; especially because, in principle, any area of the Earth can be mapped by hyperspectral imaging, be it with aircraft or satellites (Ross et al., 2005). Several studies have tested the potential of multispectral and hyperspectral images with varying results. Kalacska and Bell (2006) were among the first that demonstrated the potential of remote sensing as a tool for locating heretofore unknown mass graves. Afterwards, Kalacska et al. (2009) analyzed the in situ and airborne spectral reflectance of a set of animal mass graves and identically constructed false graves. Their results indicated that the reflectance spectra of grave are readily distinguishable from false grave at both scales. In addition, they observed that vegetation regeneration was severely inhibited by cattle carcasses for up to a period of 16 months. Caccianiga et al. (2012) studied the effects of decomposition of buried swine carcasses on soil and vegetation structure and composition as a tool for detecting clandestine graves. They found that soil disturbance was the main factor affecting plant cover, while the role of decomposition seemed to be much less critical. Leblanc et al. (2014) performed a blind-test of the potential for airborne hyperspectral imaging technology to locate buried remains of pig carcasses. They were able to predict two single graves, within GPS error (10 m), whose location they did not know. Recently, Silván-Cárdenas et al. (2017) studied some methods for detecting clandestine graves using hyperspectral data collected on ground. Through a controlled experiment using buried carcasses of pigs, demonstrated that hyperspectral data have potential for detecting buried remains only after three months from burial. Furthermore, that the critical spectral regions for graves detection are the NIR and SWIR11 spectral regions, some of which were so narrow (10 nm) that stressed the need for hyperspectral sensing. The method of acquisition of hyperspectral images is equally important than the process of pattern recognition for detection of graves based on such information. In this sense, some techniques of evolutionary computation have been successfully applied for selection and combination of spectral bands aiming at different applications such as classification of vegetation species, soil mineral identification, synthesizing spectral indices, estimate pasture mass and quality, and precision farming, to mention just a few (Ross et al., 2005; Chion et al., 2008; Albarracín et al., 2016; Zhuo et al., 2008; Li et al., 2011, Kawamura et al., 2010, Puente et al., 2011, Ullah et al., 2012, Davis et al., 2006, Landry et al., 2006; Kawamura et al., 2010; Awuley and Ross, 2016). On the other hand, currently, visual attention models have been designed for the spatial and spectral analysis of hyperspectral images with applications such as detection of prominence, visualization and interpretation, and detection of objects (Le Moan et al., 2011and 2013; Wang, 2013; Liang et al., 2013; Cao et al., 2015; Zhang et al., 2017). In this study an evolutionary technique is proposed based on genetic programming, known as Brain Programming (BP), for optimizing a so-called Hyperspectral Visual Attention Models (H-VAM) for graves detection. Problem Statement The present work addresses the problem of detection of clandestine graves as a problem of classification of hyperspectral images. The image classification problem can be stated in formal terms as follows. Suppose we want to classify each pixel in an image into one of N classes, let say C1, C2, ..., CN. Then, decision rules must be established to enable assignment of any given pixel to these classes (Varshney and Arora, 2004). When working with hyperspectral images, some issues arise due to the high dimensionality of this type of images, e.g., Hughes phenomenon, high information redundancy in spectral and spatial domains, need for finding features that increase discrimination between classes and high computational resources required in the classification process. For this reason, a compelling need to reduce the dimension of data exists. The methods for reduction of dimensionality can be roughly divided into two categories: feature extraction Centro de Investigación en Geografía y Geomática “Ing. Jorge L Tamayo” A.C., Circuito Tecnopolo Norte No. 117, Fracc. Tecnopolo II, Pocitos, Aguascalientes, CP 20313, Mexico (leon.dozal@gmail.com). Photogrammetric Engineering & Remote Sensing Vol. 84, No.7, July 2018, pp. 19–xxx. 0099-1112/18/19–xxx © 2018 American Society for Photogrammetry and Remote Sensing doi: 10.14358/PERS.84.7.xxx 1. The abbreviations used in this paper are summarized in Table 1. PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING J u l y 2018 19 and feature selection (Li et al., 2014). The former refers to a process to transform the high-dimensional feature space to a low-dimensional space through linear or no linear combinations of original bands. Whereas, the latter refers to a selection of an optimal subset of features, through a combinatorial optimization (Yoon and Park, 2015). At this point, it should be noted that classification accuracy is highly impacted by feature selection or extraction methods. In this study, we tested the BP algorithm as a feature selection method. Much research has been done about feature selection from an evolutionary computation approach (see the Brain Programming Implementation Section ), but little work has been done around feature extraction methods incorporating both spatial and spectral information through such approach. In order to integrate spectral and spatial information in our solution to the grave detection problem, we use an HVAM optimized with a BP algorithm. The H-VAM allows building a saliency map that corresponds to a classified image. Visual Attention from a Computational Standpoint The concept of visual attention or visual saliency comes from the study of the human visual system and this consists in the ability to efficiently find objects or regions that stand out relative to their neighboring parts in a scene (Borji and Itti, 2013). Therefore, visual attention is an inherently active and selective process by which people attend to a subset of the available information for further processing along the visual pathway (Zhang et al., 2008). For this reason, when it comes to computer science, the notion of saliency is closely related to visual attention in color or gray scale images (Le Moan et al., 2013). In a broad sense, the concept of saliency is not exclusive of the vision process. Recently, Le Moan et al. (2013) coined the term “spectral saliency” to name the extent to which a certain group of pixels stands out in an image, in terms of reflectance rather than in terms of colorimetric attributes. In principle, any kind of data can be considered, although, until now, only a few studies have used other than visible spectrum images. In 1998, the first computer models of visual attention appeared. The most influential model has been that of Itti et al. (1998); they introduced the concept of “saliency map” that is an image in gray scales where the larger the value of a pixel, the more prominent it is. Since then, the saliency map has been utilized in various applications, such as object recognition, object detection, segmentation, and tracking. The reader is referred to (Borji and Itti, 2013) for a comprehensive review of the state of the art in visual attention modeling. There is a growing interest in using saliency attention models for multi and hyperspectral images. Considering the high dimensionality of hyperspectral images, traditional Visual Attention Model (VAM), such as that of Itti et al. (1998), cannot be directly applied to hyperspectral imagery (Cao et al., 2015). Nowadays, there are few VAMs dealing with the spectral saliency (Le Moan et al., 2011and 2013; Wang, 2013; Liang et al., 2013; Cao et al., 2015; Zhang et al., 2017). Most of the methods adopted saliency detection based on spectral signatures. In this way, feature extraction is only operated in the spectral domain, but the spatial distribution of targets has not been fully explored. However, the spatial distribution has proved to be very valuable for image analysis in remote sensing and computer vision communities (Cao et al., 2015). Introduction to Brain Programming Brain Programming (BP) is an evolutionary computing technique based on Genetic Programming (GP), |
| Starting Page | 435 |
| Ending Page | 450 |
| Page Count | 16 |
| File Format | PDF HTM / HTML |
| DOI | 10.14358/pers.84.7.435 |
| Volume Number | 84 |
| Alternate Webpage(s) | http://mid.geoint.mx/documentos/20180919_162152_anexo_006n_2018_dozal_pers.pdf |
| Alternate Webpage(s) | https://doi.org/10.14358/pers.84.7.435 |
| Language | English |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Article |
