Integration of Spatial and Spectral Information by Means of ...

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING

1

Integration of Spatial and Spectral Information by Means of Unsupervised Extraction and Classification for Homogenous Objects Applied to Multispectral and Hyperspectral Data Luis O. Jiménez, Jorge L. Rivera-Medina, Eladio Rodríguez-Díaz, Emmanuel Arzuaga-Cruz, and Mabel Ramírez-Vélez, Member, IEEE

Abstract—This paper presents a method of unsupervised enhancement of pixels homogeneity in a local neighborhood. This mechanism will enable an unsupervised contextual classification of multispectral data that integrates the spectral and spatial information producing results that are more meaningful to the human analyst. This unsupervised classifier is an unsupervised development of the well-known supervised extraction and classification for homogenous objects (ECHO) classifier. One of its main characteristics is that it simplifies the retrieval process of spatial structures. This development is specially relevant for the new generation of airborne and spaceborne sensors with high spatial resolution. Index Terms—Contextual classification, multispectral data analysis, multivariate image analysis, pattern recognition, remote sensing, spectral-spatial classification, unsupervised classification.

I. INTRODUCTION

C

URRENT developments in high-resolution imaging provide higher spatial resolution sensors by decreasing the instantaneous field of view (IFOV). Landsat’s Enhanced Thematic Mapper Plus (ETM+) sensor provides imagery with an IFOV of 15 m. IKONOS provides data with 0.7 m, and the Hyperspectral Digital Imagery Collection Experiment (HYDICE) is able to collect hyperspectral data, depending on the height of its carrier, with 3.5 m of spatial resolution. This

Manuscript received February 21, 2003; revised August 9, 2004. This work was supported in part by the U.S. Army Corps of Engineering Topographic Engineering Center under Contract DACA76-97-K-0007, in part by the National Science Foundation Engineering Research Centers Program under Grant EEC-9986821, in part by the National Aeronautics and Space Administration University Research Center Program under Grant NCC5-518, and in part by the National Imagery and Mapping Agency under Contract NMA2010112014. L. O. Jimenez and E. Arzuaga-Cruz are with the Electrical and Computer Engineering Department, University of Puerto Rico at Mayagüez, Mayagüez, PR 00681 (e-mail: [email protected]; [email protected]). J. Rivera-Medina is with Kodak Corporation, Rochester, NY 14607 USA (e-mail: [email protected]). E. Rodriguez-Diaz is with the Department of Electrical and Computer Engineering, Boston, University, Boston, MA 02215 USA (e-mail: [email protected]). M. Ramirez-Vélez is with the High Tech Tools and Toys Laboratory, Center for Subsurface Sensing and Imaging Systems, Electrical and Computer Engineering Department, University of Puerto Rico at Mayagüez, Mayagüez, PR 00681 (e-mail: [email protected]). Digital Object Identifier 10.1109/TGRS.2004.843193

current trend of development provides more information about spatial structures to the human analyst. At the same time, this type of dataset adds more complexity to the retrieval process that requires higher processing time. Further development in data processing is required to retrieve meaningful spatial information from multispectral and hyperspectral sensors with high spatial resolution. Known unsupervised classification algorithms have been developed for multispectral imagery, i.e., C-means and ISODATA. These types of algorithms only use the spectral content as information to discriminate between different unsupervised classes, also named clusters. Spatial information has been used in union of spectral information in supervised and unsupervised classification. In supervised classification, spatial information has been used to correct classification errors. Examples are the Gauss–Markov random field (GMRF)-based methods [1]–[3] and the supervised ECHO classifier [4], [5]. In terms of unsupervised classification, Baraldi et al. developed a nonadaptive Bayesian contextual clustering algorithm based on Markov random fields (MRFs) [6]. Usually, MRF-based unsupervised classification algorithms are highly complex, taking long periods of data processing. This paper presents a method developed for unsupervised enhancement of pixel spectral homogeneity in a local neighborhood. If there is strong evidence that pixels in a neighborhood are spectrally homogeneous, they are enforced to the same cluster [7]. The algorithm will estimate from the same dataset, without the input from the human analyst, the required threshold values required to evaluate the conditions of homogeneity of all the neighborhoods. This mechanism will enable an unsupervised contextual classification of multispectral data that integrates the spectral and spatial information producing results that are more meaningful to the human analyst. One of its main characteristics is that it simplifies the retrieval process of spatial structures. II. INTEGRATION OF SPECTRAL AND SPATIAL INFORMATION The extraction and classification for homogenous objects (ECHO) classifier is an example of a supervised classifier that integrates spatial and spectral information. ECHO classifies pixels in a neighborhood to a single supervised class after

0196-2892/$20.00 © 2005 IEEE

2


measuring the degree of homogeneity on the neighborhood [4]. It uses the probability density function of the neighborhoods

(1) is the th pixel in the neighborhood . The form of the where neighborhood and the parameters required are estimated from the labeled samples on a ground truth.

, assigns all The second proposition, the pixels in the neighborhood to the th cluster if the degree of membership (degree of homogeneity) to that particular th cluster is the largest (smallest ). An advantage of this mechanism is that it is fast and integrates at the same time , which is the the spatial and the spectral information. degree of heterogeneity and can be understood as a degree of membership deficiency, is inversely proportional to the degree of homogeneity and cluster membership. It is defined as

A. Unsupervised ECHO: UnECHO

(3)

The unsupervised version of ECHO proposed in this paper (referred from now on as UnECHO) uses the concept of degree of membership of a neighborhood to a single cluster. This degree of membership verifies the degree of heterogeneity, or homogeneity, of a particular neighborhood. Let be the set of pixels is a measurement that are members of a neighborhood. of the neighborhood’s (which contains the set of pixels in ) runs degree of heterogeneity to a particular th cluster. inversely proportional to the degree of homogeneity. A large signifies a high heterogeneity, which implies that the set belongs to different clusters. A low signifies a high belong to the homogeneity, which implies that all pixels in th cluster. UnECHO has two stages. The first stage consists of a conventional clustering algorithm that classifies the image on a pixel-by-pixel basis, i.e., C-means. The second stage uses both, the classification results of the first stage and the spectral content of every pixel on the image, to take into consideration the spatial context and the spectral information. In this unsupervised version, due to the unsupervised mechanism of the algorithm, the objects to be extracted are clusters that represent regions in the imagery. UnECHO divides the whole multispectral or hyperspectral image into a number of nonoverlapping neighborhoods. The algorithm will measure the degree of heterogeneity of all neighis spectrally homogeborhoods. If the th neighborhood neous, which implies low heterogeneity, the spectral content of all the pixels in the neighborhood is similar. As a consequence, will be forced to be one cluster according to the following rule. is The complete set of pixels of the th neighborhood classified as part of the th cluster

if

and

(2)

The first proposition, , verifies if the degree of homogeneity of the th neighborhood is large enough. If the is large enough, then that parspectral homogeneity of ticular neighborhood can be classified as a single cluster acis cording to the second proposition. If that is not the case, spectrally heterogeneous and has a high degree of membership deficiency to a single cluster. On this condition, the algorithm does not change the classification performed on a pixel-by-pixel basis for that particular neighborhood that resulted from the first stage.

is the th pixel in the th neighborhood and where is the number of pixels in that neighborhood. is the distance of the pixel to the th cluster. Different metrics could be used to measure these distances. Here we will mention three possibilities. 1) Euclidean distance: (4) 2) Mahalanobis distance: (5) 3) Maximum likelihood: (6) is the expected value, and is the covariance of where the th cluster. The development and implementation of the algorithm are independent of the type of measurements that are found in (4)–(6). The algorithm works for any multivariate measurement, e.g., remote sensing reflectance. This accomplishes the objective of analyzing any multivariate imagery without a priori knowledge. Let us explain a general guideline for the choice of one among the three possible distances. Usually, maximum likelihood, under the condition of a good estimation of the covariance matrices , provides better results due to its direct relationship with the minimization of the probability of error. If the eigenvalues of the estimated covariance matrices are close to zero, the Euclidean distance should be used, to avoid problems with matrix inversion. is the average of the distances of every pixel to the th cluster. If is too large ( ), then this implies that the pixels in the th neighborhood are spectrally heterogeneous, hence not spectrally homogeneous. As a consequence, the pixels in the neighborhood have a high probability of beis small ( ), then longing to different clusters. If the th neighborhood is very homogeneous; hence, all the pixels are likely to belong to one cluster, although they could in have been classified to different unsupervised classes at the first stage.

JIMÉNEZ et al.: INTEGRATION OF SPATIAL AND SPECTRAL INFORMATION

3

C. Estimation of the Threshold

Fig. 1. UnECHO possible neighborhoods. (a) The 2 and (b) 3 3 neighborhood.

2

2 2 pixels neighborhood

(a)

2

Fig. 2. (a) The 2 4 pixel grid with nonoverlapping 2 The 2 2 overlapping neighborhoods.

2

(b)

2 2 neighborhoods. (b)

B. Neighborhood Construction This section explains how UnECHO selects and analyzes the spatial neighborhoods. In the supervised version of ECHO, the spatial structure of the neighborhoods was obtained from the labeled data. In UnECHO, due to a lack of a priori information, the neighborhoods are arrangements of 2 2, 3 3 or 4 4 pixels, as shown in Fig. 1. Square neighborhoods were chosen in this implementation due to simplicity and performance considerations, not to restrictions in the theory. Neighborhood size has an impact over the execution speed of UnECHO. Larger neighborhoods take more time to analyze, but at the same time, the amount of neighborhoods in the image is reduced. Using the threshold computation method described in Section II-B, the amount of thresholds to compute increases along with the size of the neighborhoods. Hence, selecting larger neighborhoods results in slower execution speed. UnECHO uses nonoverlapping neighborhoods. The use of nonoverlapping neighborhoods is recommended to avoid introducing variations by the raster scanning and to improve performance. Another advantage of our method is that it avoids problems with convergences because it is not iterative due to its nonoverlapping approach. Fig. 2 shows a 2 4 grid with 2 2 neighborhoods circled using nonoverlapping neighborhoods (a) and overlapping neighborhoods (b). When neighborhoods overlap, the order in which the neighborhoods are analyzed might affect the classification results. Also notice the increased amount of neighborhoods to analyze. This affects directly the execution speed of the algorithm. In this nonoverlapping procedure, the shapes of the clusters and their borders do not affect the classification. Precisely those neighborhoods that are located in the borders are spectrally highly heterogeneous. As a consequence, they are not classified as one cluster. The classification of the pixels in these neighborhoods is according to the pixel-by-pixel clustering from the first stage.

As mentioned earlier, UnECHO will estimate from the same dataset, without input from the human analyst, the required required to evaluate the conditions of thresholds values heterogeneity and homogeneity of all the neighborhoods. In order to explain the computational process to estimate , we need to introduce the following notation. is the th neighborhood in This notation is as follows. the multidimensional data. The image to be analyzed will be divided in neighborhoods with the structures shown in Fig. 1 and explained in Section II-B. is the average distance of the pixels to the th cluster, according to (3). It represents the degree of spectral content heterogeneity that represents the degree of membership deficiency of the th neighborhood to the th cluster. is a vector that represents a particular classification distribution of the pixels in one neighborhood. This classification distribution is based on the results of the unsupervised classification applied in the first stage on a pixel-by-pixel -class (cluster) problem, is an -tuple: basis. For an , where is the number of pixels in a particular neighborhood that after the first-stage classification belongs to the th cluster. Examples of a composition structure from different neighborhoods are found in Fig. 3. These are three 3 3 neighborhoods with two clusters represented by the red and blue colors. In these cases, all of the neighborhoods have the composition structure Number of pixels that belong to the red cluster Number of pixels that belong to the blue cluster

means that the particular neighborhood has a composition in terms of the pixel-by-pixel classification that results from the first stage (i.e., C-means). In the examples of , where , for . Fig. 3, is the number of neighborhoods in the classification results from the first stage that has the particular composition . , , and are the only ones, If the neighborhoods in the resulted class map from the first stage clustering of the , then . imagery, with The steps to estimate the thresholds used in (2) are as follows. . The variable represents the Let cluster that provides the minimum spectral average distance (3). It has the minimum degree of heterogeneity, hence maximum degree of homogeneity, for the th neighborhood. Fig. 4 shows the representation of the spectral feature space neighborhood of a neighborhood in Fig. 3. Fig. 4(a) of the represents the Euclidean distances in (4) by dotted lines. These in , to the red represents distances between every pixels cluster spectral mean. Fig. 4(b) represents the Euclidean dis, to the blue tances in (4) of every pixel , that belongs to cluster spectral mean by a dotted line. According to the representations of Fig. 4, the dotted lines to the blue cluster mean location are, on the average, shorter than the dotted lines to the red cluster mean location. This implies that the average of the Euclidean distances to the mean of the blue cluster is smaller than

4


X

Fig. 4. (a) Representation of the neighborhood of Fig. 2(a) on a two-dimensional spectral feature space. The Euclidean distances from every pixel to the red cluster mean are represented using dotted lines. (b) neighborhood of Fig. 2(a) on a two-dimensional Representation of the spectral feature space. The Euclidean distances from every pixel to the blue cluster mean are represented using dotted lines.

X

be classified totally to a single cluster, according to the rule in (2), it should be the blue cluster. Then . , for . This is the sum Let of all the degrees of heterogeneity for the neighborhoods that has the particular composition and has a larger value for the th cluster. In the examples of Figs. 3 and 4, let us assume the blue blue and red following: , , . Then blue . , for the neighborhoods that have the The threshold particular composition and have a larger degree of membership to the th cluster, is computed by (7)

Fig. 3. Examples of 3 algorithm.

2 3 neighboroods and its classification after first-stage

the average of the Euclidean distance to the red cluster mean. According to (3) and (4), . As a consequence, this neighborhood is spectrally more homogeneous for the blue cluster and more heterogeneous with a higher degree of membership deficiency for the red cluster. If this neighborhood could

According to (7), the threshold of neighborhoods with the composition structure that has the maximum degree of homogeneity for the th cluster is estimated as the average of all these possibilities in the image. Following our . Observe that previous example, the information of the structure is found in the results of the pixel-by-pixel clustering map obtained in the first stage. Several experiments were performed using different multispectral and hyperspectral datasets in order to observe the execution of the UnECHO classifier.


5

Fig. 6. (a) Result of C-means algorithm applied at first stage. (b) Result of UnECHO algorithm applied at the second stage. TABLE I C-MEANS CLASSIFICATION ACCURACY

TABLE II UNECHO CLASSIFICATION ACCURACY

Fig. 5. Segment of a HYDICE hyperspectral image analyzed.

III. EXPERIMENTS A. Experiment 1: HYDICE Imagery In this experiment, the hyperspectral data used is a segment of a HYDICE frame. HYDICE is a hyperspectral sensor that collects data in 210 bands with an IFOV of 3.5 m. This imagery posses a high spatial resolution. Fig. 5 shows the segment of the frame that was analyzed. The size of this HYDICE segment is 182 lines with 128 columns. This segment has two well-known landmarks: the roof of a supermarket, in white color, and the asphalt of a parking lot below the white structure. This information will be used for validation purposes. Fig. 6(a) shows the results of the C-means clustering used in the first stage of the algorithm and the results of the second stage in UnECHO [Fig. 6(b)] that integrates the spatial and the spectral information. The result of UnECHO classifier, shown in Fig. 6(b), possesses more details and is better defined in terms of the spatial structures such as buildings and parking lots. For example, UnECHO classified as one cluster, red color, the roof of a supermarket. Meanwhile in Fig. 6(a), pixel-by-pixel C-means clustering, it appears as two different and not well-defined clusters, brown and red. In Fig. 6(b), the asphalt in the whole image is mostly contained in one cluster, blue color, facilitating the process of identifying a parking lot and the roads in the class

map. In Fig. 6(a), the asphalt appears in at least two distinct clusters, blue and green, making more difficult the identification process after applying an unsupervised classification algorithm. Tables I and II have the results of a quantitative analysis on the classification accuracy of these landmarks, Roof (of supermarket) and Asphalt (in parking lot) that were available on a ground truth. These quantitative results were based on independent sets of testing samples. The number of testing samples used to validate the cluster Roof was 2184, and the numbers for the cluster Asphalt was 513. As can be seen from these results, UnECHO enhances the classification accuracy of C-means clustering by a total of 16%. UnECHO finds more objects with more detailed spatial structures that are meaningful to a human analyst and facilitates the identification process for spatial structures in an urban area, e.g., the orange and yellow clusters. Meanwhile, in pixel-by-pixel classification, due to the speckle appearance, there is a lack of objects that could have spatial meanings for a human analyst, like buildings, roads, and cars. In general, UnECHO presents a more clear class map, enhancing borders of regions with different shapes, without loosing the particular details of small objects like cars and narrow lines. B. Experiment 2: AVIRIS Data The second experiment was performed on a segment from an Airborne Visible Infrared Imaging Spectrometer (AVIRIS) image of the Kennedy Space Flight Center in Florida. It has an IFOV of 20 m. The frame has a total of 220 bands. The segment used in this experiment has 107 lines and 46 columns. This imagery combines rural zones with civil infrastructures. Fig. 7

6


Fig. 9.

Theoretical classification map and reference image.

Fig. 7. AVIRIS data. (a) Segment of the original image. (b) Result of C-means on the segment. (c) Results of UnECHO.

TABLE III NUMBER OF TESTING SAMPLES

Fig. 8. Spectral reflectance, used as class means, of four different types of corals. Reflectance versus wavelength.

shows the segment of the original image, the results of -means, and the results of UnECHO. A feature of interest in this image is a road that crosses the image segment from top to bottom. Observe that UnECHO classifier is able to detect the road with more detail than C-means. It is more capable of uncovering spatial structures with high informational value for a human analyst, like civil infrastructure. C. Experiment 3: Synthetic Data This experiment was conducted using synthetic data generated with Hydrolight v.4.2. The data were created using the spectral responses of several targets, at a wavelength range from 400–700 nm in steps of 10 nm. The effects of Case 2 waters, scattering and absorption, were added to this dataset using this data generator. These are waters in which the total absorption is determined by the absorption of chlorophyll and related pigments, dissolved organic matter, and inorganic particles [8]. These Case 2 waters have optical properties that present difficulties to the identification process of objects, such as corals, under the surface of coastal waters. The resulting spectra generated from Hydrolight are shown in Fig. 8. Observe that these spectra are very similar. Each spectrum was assigned

Fig. 10.

MLC.

to an area of the reference image shown in Fig. 9. This reference image, of 150 lines and 150 columns, will also be used as a ground truth when determining the classification accuracy. Noise with an independent identically distributed distribution, Gaussian with zero mean, and a standard deviation of 0.0025 was added to every pixel in the dataset in order to create high variability among the pixels. In this particular experiment, the second stage of UnECHO was applied to a supervised classification result. The class map of a supervised classification result was used as the input of the second stage in UnECHO. This experiment has the objective of testing the performance of UnECHO in enhancing a supervised classification. Table III shows the number of testing samples per supervised class used to validate the algorithm. Figs. 10–13 present the classification results and accuracies for supervised classification and for different types of neighborhoods applied to UnECHO. Fig. 10 shows the results for maximum-likelihood classification (MLC). Notice the speckled appearance of the result due to the noise added. Figs. 11–13 show the results obtained by using UnECHO postprocessing


7

TABLE IV ML CLASSIFICATION ACCURACY

Fig. 11. MLC with the enhancement of UnECHO classifier, 2 neighborhood.

TABLE V UNECHO CLASSIFICATION ACCURACY 2

2 2 NEIGHBORHOOD

TABLE VI UNECHO CLASSIFICATION ACCURACY 3

2 3 NEIGHBORHOOD

TABLE VII UNECHO CLASSIFICATION ACCURACY 4

2 4 NEIGHBORHOOD

22

Tables IV–VII present the classification accuracy of different classifiers on labeled testing samples. As can bee seen from these results, UnECHO enhances the classification accuracy of a supervised classifier, ML, by 8.29% in a 2 2 neighborhood, 16.6% in a 3 3 neighborhood, and by 19.36% in a 4 4 IV. CONCLUSION Fig. 12. MLC with the enhancement of UnECHO classifier, 3 neighborhood.

Fig. 13. MLC with the enhancement of UnECHO classifier, 4 neighborhood.

23

24

with neighborhoods of 2 2, 3 3, and 4 4 windows respectively. Observe how the total classification accuracy improves using the 2 2, the 3 3 and the 4 4 window in relation to MLC alone.

In this paper, we presented UnECHO, a mechanism of integrating spatial and spectral information in an unsupervised classifier. This method was developed for unsupervised enhancement of pixels spectral homogeneity in a local neighborhood. It uses the concept of neighborhoods and degrees of homogeneity similar to ECHO. UnECHO is divided in two stages. In the first stage, it uses the spectral information on a pixel-by-pixel basis. This first stage can be a result from either a clustering algorithm such as C-means or a supervised classification such as ML classifier. In the second stage, it uses both spectral information from the original image and the content of the first stage clustering map. One of its main advantages is that it simplifies the retrieval process of spatial structures. The experiments with synthetic and real hyperspectral data showed results that are more meaningful to the human analyst than traditional pixel-by-pixel clustering. Quantitative analyses confirm the results. ACKNOWLEDGMENT The authors will like to acknowledge the NASA Kennedy Space Center for providing the AVIRIS hyperspectral data and the Topographic Engineering Center of the U.S. Army Corps of Engineering Work for providing the HYDICE hyperspectral data.

8


REFERENCES [1] T. Yamasaki and D. Gingras, “Image classification using spectral and spatial information based on MRF models,” IEEE Trans. Image Process., vol. 4, no. 9, pp. 1333–1339, Sep. 1995. [2] Y. He and A. Kundu, “2D shape classification using hidden markov model,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 13, no. 11, pp. 1172–1184, Nov. 1991. [3] G. J. McLachlan, Discriminant Analysis and Statistical Pattern Recognition, 1st ed. New York: Wiley, 1992, pp. 413–443. [4] D. A. Landgrebe, “The development of a spectral-spatial classifier for earth observational data,” Pattern Recognit., vol. 12, pp. 165–175, 1980. [5] J. A. Richards, Remote Sensing Digital Image Analysis, an Introduction, 2nd ed. New York: Springler-Verlag, 1993. [6] A. Baraldi, P. Blonda, F. Parmiggiani, and G. Satalino, “Contextual clustering for image segmentation,” Opt. Eng., vol. 39, no. 4, pp. 1–17, Apr. 2000. [7] L. O. Jimenez and J. Rivera, “On the integration of spatial and spectral information in unsupervised classification for multispectral and hyperspectral data,” presented at the SPIE Conf., Florence, Italy, Sep. 1999. [8] C. D. Mobley and C. D. Mobley, Light and Water, Radiative Transfer in Natural Waters. Orlando, FL: Academic, 1994.

Luis O. Jiménez received the B.S.E.E. degree from the University of Puerto Rico at Mayagüez, (UPRM), Mayagüez, in 1989, the M.S.E.E. degree from the University of Maryland, College Park, in 1991, and the Ph.D. degree from Purdue University, West Lafayette, IN, in 1996. He is currently a Professor of electrical and computer engineering at UPRM and the Director of the Laboratory of Applied Remote Sensing and Image Processing. He is also the Director of the Center for Subsurface Sensing and Imaging Systems, which is the UPRM component of a National Science Foundation Engineering Research Center. The objective of this particular center is to revolutionize our ability to detect and image biomedical and environmental–civil objects or conditions that are underground, underwater, or embedded in the human body. His research has been in the area of hyperspectral image analysis, remote sensing, pattern recognition, and image processing. Dr. Jimenez is member of the IEEE Geoscience and Remote Sensing Society, the IEEE System, Man, and Cybernetics Society, and the SPIE Society. He is also member of the Tau Beta Pi and Phi Kappa Phi honor societies.

Jorge L. Rivera-Medina received the B.S.C.E. and M.SC.E. degrees from the University of Puerto Rico at Mayagüez, Mayagüez, in 2000 and 2003, respectively. He is currently a Development Engineer at Eastman Kodak Company, Rochester, NY, since 2001. His research interests are in the are as of pattern recognition, remote sensing, image processing, and video compression.

Eladio Rodríguez-Díaz received the B.S.E.E. and M.S.E.E. degrees from the University of Puerto Rico at Mayagüez (UPRM), Mayagüez, in 2000 and 2003, respectively. He is currently pursuing the degree of Ph.D. in electrical engineering at Boston University, Boston, MA. His research work was done at the Laboratory of Applied Remote Sensing and Image Processing of the Electrical and Computer Engineering Department, UPRM, and with the Center for Subsurface Sensing and Imaging Systems, the UPRM component of the Engineering Research Center. His research work has been in the area of remote sensing, pattern recognition, inverse problems, and image processing. Mr. Rodríguez-Díaz is a member of the Tau Beta Pi and Golden Key national honor societies.

Emmanuel Arzuaga-Cruz (M’02) received the B.S.Cp.E. and M.S.Cp.E. degrees from the University of Puerto Rico at Mayagüez (UPRM), Mayagüez, 2000 and 2003, respectively. He is currently working in the Department of Electrical and Computer Engineering of UPRM as a Software Developer for the Laboratory for Applied Remote Sensing and Image Processing. His work is related with the development of remote sensing and pattern recognition software for the Center of Subsurface Sensing and Imaging Systems and the Tropical Center for Earth and Space Studies funded by the National Aeronautics and Space Administration. Mr. Arzuaga-Cruz is a member of the IEEE Computer Society.

Mabel D. Ramírez-Vélez (M’02) received the B.S.E.E. and M.S.E.E. degrees from the University of Puerto Rico at Mayagüez (UPRM), Mayagüez, in 2001 and 2003, respectively. She is currently working as an Educational Module Developer for the Center of Subsurface Sensing and Imaging Systems’ High Tech Tools and Toys Laboratory in the Electrical Engineering Department, UPRM. Her research interests include, estimation, remote sensing, pattern recognition, image processing, and digital signal processing.