Directional Analysis and Filtering for Dust Storm detection in NOAA-AVHRR Imagery

Transcription

1 Directional Analysis and Filtering for Dust Storm detection in NOAA-AVHRR Imagery S. Janugani a,v.jayaram a,s.d.cabrera a,j.g.rosiles a,t.e.gill b,c,n.riverarivera b a Dept. of Electrical & Computer Engineering b Environmental Science & Engineering Program c Dept. of Geological Sciences The University of Texas at El Paso 0 W. University Ave., El Paso, TX ABSTRACT In this paper, we propose spatio-spectral processing techniques for the detection of dust storms and automatically finding its transport direction in 5-band NOAA-AVHRR imagery. Previous methods that use simple band math analysis have produced promising results but have drawbacks in producing consistent results when low signal to noise ratio (SNR) images are used. Moreover, in seeking to automate the dust storm detection, the presence of clouds in the vicinity of the dust storm creates a challenge in being able to distinguish these two types of image texture. This paper not only addresses the detection of the dust storm in the imagery, it also attempts to find the transport direction and the location of the sources of the dust storm. We propose a spatio-spectral processing approach with two components: visualization and automation. Both approaches are based on digital image processing techniques including directional analysis and filtering. The visualization technique is intended to enhance the image in order to locate the dust sources. The automation technique is proposed to detect the transport direction of the dust storm. These techniques can be used in a system to provide timely warnings of dust storms or hazard assessments for transportation, aviation, environmental safety, and public health. Keywords: filtering. Multispectral imagery, NOAA-AVHRR imagery, Dust storm, Band math analysis, Directional 1. INTRODUCTION Dust Storms are serious causes of physical, environmental and economic hazards. Due to this growing concern, detection of dust storms using multispectral imagery has become a topic of several recent studies. 1 6 The need to develop a real-time capability to detect, identify, quantify and track various dust storms is a well known problem to scientists and climate monitoring agencies. Detection of particulate effluents, the study of volcanology and environmental risk management are just a few of the many applications that demand an automated/in-situ processing system. The multispectral images from AVHRR, GOES, MODIS and other sensors have opened up new possibilities for detecting and monitoring atmospheric dust events and tracking the potential environmental risks. Although several traditional methods for detecting dust storms exist, they have difficulties in consistently detecting and distinguishing the dust storm features. Traditional methods include: band math analysis, 2 differencing of radiation temperature or the infrared split-window technique, single band thresholding and multi-band combinations. They have achieved good results in detecting dust storms, 5 however, they have limitations in detecting transport direction and in identifying the location of the dust storm sources. In this paper, we propose a spatio-spectral processing scheme to detect the presence of dust storms visually and with an automated technique for finding dust storm direction. The goal is to present seamless integration of advanced data analysis and modeling tools to scientists and climate monitoring personnel, advancing the stateof-the-practice in the utilization of satellite image data to various types of dust storm studies. The proposed methodology is a twofold approach. First, we present a scheme to visualize the directionality and point sources Further author information: (Send correspondence to S. D. Cabrera) Prof. S. D. Cabrera: sergioc@utep.edu, Telephone:

2 of these dust storm events. The second approach provides an automated technique to compute the transport direction of the dust storm. In the visualization technique, image processing algorithms like the spectral-domain principal component analysis (PCA), minimum noise fraction (MNF) transforms 7 and the spatial-domain k- means unsupervised classification method are used as tools to help locate dust storms visually. Pointing out the dust sources is done based on image information, directional filtering in combination with edge detectors and spectral-domain classification techniques. Next, edge detectors like Sobel and Frei-Chen are applied to the selected filtered images for further enhancement of the streaks produced by the directional texture. False color composite images are created to visually enhance the directional streaks to be able to easily locate the dust sources. The second approach comprising the automation technique for finding direction of the dust storm involves performing the power spectrum analysis on bands 4 and 5. This step is performed because these wavelengths highlight the absorption and subsequent emission of thermal radiation by the silicate particles in the dust storms. The scheme involves block processing consisting of power spectrum analysis followed by binary thresholding and morphological enhancement. This local spectral analysis is first used to confirm the presence of high directionality information in certain regions of an image. Binary thresholding is performed on these blocks to enhance the directional texture. Further morphological enhancement is performed on these binary images to compute the dust storm s area and orientation. The similarity in orientation among neighboring blocks also gives confirming evidence of the presence of a dust storm in the area formed by putting together these blocks. 2. EARLIER WORK The motivation for this paper is derived from the experiments documented in Rivera and Gill s work. 4, 5 The methodology developed in that previous work, primarily uses band math analysis as the technique to detect dust storm areas. Dust sources are located by visual analysis and manual labeling. Other previous methods for dust storm detection include: differencing of radiation temperature, single band thresholding and multi-band combining. Some have achieved good results in detecting dust storms, however, these techniques have limitations in detecting direction of transport of the dust storm, and the visibility index. Above all, there is an absence of an automated detection scheme for in-situ processing systems. In a related work, hierarchial principal component analysis (HPCA) techniques 8 produced significant segmentation results. However, the technique is computationally challenging and expensive as they involve data fusion both spectrally and spatially from cameras placed at different angles. In most practical scenarios, it is not possible to get the raw data sets of the same scene (same geographical area at a particular time) from cameras placed at different angles. Koren et al. 9 proposed a dust source detection method using geo-referencing of the dust-storm image with the non-dust-storm image of the same location. The limitation of this method is that the technique is dependent on the availability of the image at the same location, with identical geometry and resolution. To avoid any false alarms, resulting from clouds or pollution, the image should also be captured at close to the same time. 3. SPATIO-SPECTRAL PROCESSING FOR DUST STORM DETECTION In this paper, we focus on developing an automated approach for locating dust sources and finding dust storm direction using NOAA-AVHRR imagery. Images collected with the AVHRR, as well as the MODIS sensor, are of particular interest in dust storm detection because of their relatively higher temporal resolutions (once per day) compared with LANDSAT-type systems (once every 16 days). The nighttime thermal imaging capability of AVHRR and the higher spectral resolution of MODIS can also assist in detecting dust emission and mapping the vulnerability of the landscape to wind erosion. 2 As mentioned earlier, our approach involves both spatial and spectral processing. A step-by-step view of our approach, involving visualization and automation, is shown in the form of a flowchart in Figure 1. The band math analysis, also used in this paper, produces results as shown in Figure 2. This difference between bands 4 and 5 is used as an initial step to detect the presence of the dust storm area. This is the first step in detecting the dust storm location as shown in the flowchart in Figure 1. The dark regions in the image indicate the presence of dust storm. The presence of clouds in the vicinity of the dust storm would make the problem more difficult since they have a very similar texture on the resulting imagery. In order to overcome

3 Figure 1. Block diagram for proposed spatio-spectral processing scheme of dust storm detection. the false alarm of misinterpreting the cloud pixels as part of the dust storm area, we can employ the classical unsupervised pattern recognition technique such as the k-means classification. 10 This step is performed to verify the dust storm presence and to match the results with those of band math analysis. The next step in the processing involves selecting a subimage (in this case a 512x512 pixel region) that includes the dust storm area. This is a cropping step from all bands of the original raw NOAA-AVHRR image data set. The k-means classification is performed on these bands to verify if it is able to segment the same dust storm region again. An illustration of this process can be seen in Figure 3. Here, the k-means classification is able to segment the same dust storm region. If band math analysis and k-means classification are able to segment a similar region, it indicates with high confidence the presence of the dust storm. Next, we proceed to locate the sources of the dust storms using the visualization technique and to find the direction of the dust storm with the automation technique. 4. LOCATING DUST SOURCES In the visualization approach, directional filtering is applied on the cropped individual image bands at different angles. Directional filters are significant in many image processing applications such as edge sharpening, feature enhancement, texture analysis and object recognition. In this paper we have employed steerable Gaussian filters and 2-D convolution filter masks at various angles. Steerable filters are a class of filters in which a filter of arbitrary orientation is synthesized as a linear combination of a set of basis filters. 11 In using this approach, the same filter is rotated to multiple angles to obtain a response for each angle.

4 Figure 2. Original band 4 (left), band 5 (center) and the bandmath (band 4 - band 5) result (right) for April 15, 3 event over southwestern North America. Figure 3. k-means classification of cropped 512x512 region of Band 1 (April 15, 3 event).

5 Consider the two-dimensional circularly symmetric Gaussian function G(x, y) = e (x2 +y 2). Let the nth derivative of a Gaussian in the x direction be Gn. Let f θ (x, y) representf(x, y) rotated through an angle θ about the origin. The first x derivative of a Gaussian is G 0 1 = x [e (x2 +y 2) ]= 2xe (x2 +y 2), and if rotated by 90 G 90 1 = y [e (x2 +y 2) ]= 2ye (x2 +y 2). Thus, the filter G 1 for any arbitrary orientation θ can be synthesized by a linear combination G θ 1 =cos(θ)g0 1 +sin(θ)g90 1. Here G 0 1 and G 90 1 are called the basis filters for G θ 1 and cos(θ), sin(θ) components are the interpolation functions for the basis filters. We can synthesize the image filtered at an arbitrary orientation using convolution by taking linear combinations of images filtered with G 0 1 and G Therefore, R1 0 = G0 1 I R1 90 = G 90 1 I R1 θ =cos(θ)r0 1 +sin(θ)r90 1. Figure 4 illustrates the results of directional filtering at different angles. The directional filtering was performed on the cropped version of band 4 for different angles: 20,30,40,,70,and90. Notice that in Figure 4, the texture of the dust storm is more prominent around 70. To concretize our findings, we have also performed directional filtering using direct convolution with another type of filters (those provided in the ENVI software) wherein a certain resulting output pixel value is the function of some weighted average of the brightness of the neighboring input pixels. The sum of the directional filter kernel/mask elements is zero for this family. Convolution of this filter mask with the input image results in a spatially filtered output image. These first derivative edge enhancement filters selectively bring out image features having a specific prominent direction. The outcome is a spatially high pass filtered image where the convolution with the directional filter mask contains areas with uniform pixel values represented by 0 and those with variable values that form a bright edge enhanced texture. It is found that computing directionality or orientation angle of the dust storm with both directional and convolution filters provides nearly identical results. It is observed that the 8 dust sources marked in yellow in Figure 5 match with the dust sources shown by Rivera. 5 Energy and entropy measurements are computed on the outputs of the directional filters to find the prominent direction of the dust storm. To enhance the streaks representing the directional components of the storm, we use edge detection operators like Sobel and Frei-Chen. To improve the visualization of the resulting images we use false coloring over composites from three out of the five bands. The display color assignment for any band is done in an arbitrary manner. In general, many environmental features are more readily discernible when satellite images are represented as false color composites. In our case, pseudo-coloring enhances the localization of the origin points of a dust storm. Figure 5 shows the located origin points of a dust storm on false color composite images of different band combinations are show in the Figure. It is observed that band combination 1, 4, 5 for colors Red, Green, and Blue is possibly the best one to locate the dust sources easily by visual interpretation.

6 Figure 4. Directional filtering of cropped band 4 for different angles 20, 30, 40,, 70, 90 (from left to right). Figure 5. (Left) False Color Composite image - Band 1, 4, 5 (R, G, B) with marked dust sources for April 15, 3 event. (Center) False Color Composite image - Band 1, 4, 3 (R, G, B) with marked dust sources for April 15, 3 event. The Figure on the right depicts the False Color Composite image - Band 1, 3, 4 (R, G, B) with marked dust sources for April 15, 3 event.

7 5. TRANSPORT DIRECTION OF DUST STORM In the automation process, we perform block processing on subimages that are cropped regions (512x512 pixels) of bands 4 and 5. Block processing 12 means dividing an image into blocks and processing them individually in an independent fashion to provide spatial localization. The smallest block size used is 128x128 based on keeping a minimum amount of frequency domain resolution in the Discrete Fourier Transform (DFT) representation of each block. A local power spectral density analysis in each of these blocks is used to find the prominent direction of the texture in that image block. This analysis can also be used to define a candidate dust storm region consisting of multiple blocks. The presence of a prominent direction in the texture of the candidate dust storm region can also be used to verify its presence as an automated detection scheme. Binary thresholding based on Otsu s 13 method is performed on the power spectrum representations for all blocks in order to define the orientation of the spectral information. Otsu s method chooses a threshold to minimize the intra-class variance of the two classes of pixels that are converted to 0 or 1 in forming the output binary image. Morphological enhancements are then applied to these binary images to better define the bright regions corresponding to large spectral energy. The size and prominent orientation of each largest bright region for each block are then calculated. These two numbers are the features that can be used to define dust storm blocks (large areas) with their corresponding orientations. The results of these experiments are presented in the following section. 6. EXPERIMENTS The current on-orbit operational satellites maintained by NOAA are NOAA-18, 17, 16, 15, 14 and Other sensors which are of particular interest in dust storm detection and monitoring are the Moderate Resolution Imaging Spectrometer (MODIS) and a sensor onboard the Geostationary Operating Environmental Satellite (GOES). MODIS has thirty-six spectral bands and GOES has five spectral bands. Imagery from GOES are the most suitable to track the time evolution of active and short lived dust storms because of their high temporal resolution. 15 Also,there are three different MODIS sensors - Terra, Aqua and Aura - each on a different satellite covering an area at different times of the day. So, as many as three MODIS images are available per day for the same geographical region. We experimented with three data sets for dust storms in the Chihuahuan desert region of southwest North America. Most of our analysis was done using NOAA-AVHRR multispectral data of two dust events which have recently been the subject of other investigations. 4, 5 These events occurred on April 15th, 3 (20:23 UTC), shown in Figure 6, and on December 15th 3 (19:51 UTC), shown in Figure 9. A third, non-dust-storm image dataset from April 1st, 3 is shown in Figure 12 and was used for analyzing susceptibility to false dust storm detections. Next we show the results of the automation process which starts with power spectrum analysis on blocks of the cropped dust storm region. The power spectrum analysis of April 15, 3 is shown in Figure 6. The blocks which are marked in red indicate the presence of strong dust storm evidence. After binary thresholding using Otsu s method, 13 the image closing operation is then performed (see Figure 7) in order to ease the extraction of the area and orientation information for the binary objects in the image, see the results in Figure 8. The results obtained from the automation technique for December 15, 3 event are as shown in the next set of Figures, see Figures 9, 10, and 11. To analyze false alarms, we study two data sets which have no dust storm evidence in them. The first one is the April 1, 3 image data. The second one is a cropped non-dust storm region taken from the original size December 15, 3 raw image data. In this paper, only the April 1, 3 image data is analyzed to investigate the performance of the proposed technique on non-dust storm imagery. We seek to analyze false alarms that could occur in an automated method when processing imagery during the absence of dust storms. Band math analysis is done as a first step. The marked dark region in Figure 12 is not a dust storm but is the result of this analysis. Therefore, this suspected 512x512 region is cropped for false alarm analysis. The suspected region from the April 1st, 3 image is then subjected to k-means classification based segmentation. The resulting region does not agree well with the region found by band math analysis, see Figure 13. Hence, we need not go further into the next steps of visualization and automation since we do not have the proper verification of the presence of a dust storm.

8 Figure 6. (Left) Cropped 512x512 band 4 of April 15, 3 event. (Right) Power Spectrum Analysis of April 15, 3 event. Figure 7. (Left) Binary thresholding of the power-spectrum of April 15, 3 event. (Right) Morphological enhancement on the binary thresholded image of April 15, 3 event.

9 Figure 8. Area and orientation computations of the binary objects in the morphologically enhanced image. Figure 9. (Left) Band 4 of December 15, 3 event. (Right) Power Spectrum Analysis of December 15, 3 event.

10 Figure 10. (Left) Binary thresholding of the power-spectrum of December 15, 3 event. (Right) Morphological enhancement on the binary thresholded image of December 15, 3 event. Figure 11. Area and orientation computations of the binary objects for December 15, 3 event.

11 Figure 12. Bandmath (Band subtraction) image of April 1, 3 event. Figure 13. (Left) Original 512x512 image of April 1, 3. (Right) k-means classified 512x512 image of April 1, 3.

12 7. CONCLUSION AND FUTURE WORK The spatio-spectral processing technique proposed in this research is one possible approach to directional analysis of dust storm evidence in satellite imagery. Future work may targeted towards more quantitative and qualitative results demonstrating directionality information of the transport of dust storms. Further analysis on the enhanced directional streaks resulting from visualization technique could be done to see how accurately they correspond to actual dust plumes created at the sources of dust. In addition, it is of interest to see if their shape closely follows the path of the dust transport over longer distances. Performance of the technique for monitoring volcanic ash could also be considered. The behavior and properties of volcanic ash particles determine the significant bands that can be used for further analysis. Future work also includes investigation of more advanced multi-resolution directional filtering approaches based on Wavelets, and Bamberger Pyramids. Other possible future work includes the assessment of this processing technique on the GOES and MODIS images, which are more frequently available so that it could be used operationally to automatically detect dust storms. These processing techniques may also be evaluated on the images with no dust storm evidence and during different times of day to analyze false alarms. False alarms that might result from the textures of clouds, smoke, fire and air pollution may also be studied. Further investigation on supervised and unsupervised classification techniques is also suggested. REFERENCES [1] J.A. Lee, T.E. Gill, K. M. M. A. A. P., Land use/land cover and point sources of the 15 December 3 dust storm in southwestern North America, Geomorphology 105, (9). [2] Ackerman, S. A., Remote sensing aerosols using satellite infrared observations, Journal of Geophysical Research 102, (1997). [3]H.El-Askary,M.Kafatos,X.L.T.E.-G., Introducingnewapproachesforduststormsdetectionusing remote sensing technology, Proc. of IEEE International Geoscience and Remote Sensing Symposium. [4] N. I. Rivera Rivera, T.E. Gill, K. G. J. H. M. B. R. F., Wind modeling of chihuahuan desert dust outbreaks, Atmospheric Environment 43, (9). [5] Rivera, N. I. R., [Detection and characterization of dust source areas in the Chihuahuan desert, Southwestern North America], Masters Thesis, Environmental Science and Engineering, University of Texas at El Paso (December 6). [6] N. Khazenie, T. L., Identification of aerosol features such as smoke and dust, in NOAA AVHRR data using spatial textures, Proc. of IEEE International Geoscience and Remote Sensing Symposium (IGARSS) (1992). [7] A. A. Green, M. Berman, P. S. and Craig, M. D., A transformation for ordering multispectral data in terms of image quality with implications for noise removal, IEEE Transcations on Geoscience and Remote Sensing 26, (1988). [8] A. Agarwal, H. El-Askary, T. E.-G. M. K. J. L.-M., Hierarchial PCA techniques for fusing spatial and spectral observations with application to MISR and monitoring dust storms, IEEE Geoscience and Remote Sensing Letters 4, (October 7). [9] I. Koren, J.H. Joseph, P. I., Detection of dust plumes and their sources in northeastern libya, Can. J. Remote Sensing 29, No. 6, (3). [10] R. O. Duda, P. E. Hart, D. G. S., [Pattern Classification], Wiley Inter-Science, Hoboken, NJ, second ed. (1). [11] W.T. Freeman, E. A., The design and use of steerable filters, IEEE Transactions on Pattern Analysis and Machine Intelligence 13 (1991). [12] Schowengerdt, R. A., [Remote Sensing Models & Methods for Image Processing], Academic Press, Burlington, MA, seventh ed. (1997). [13] Otsu, N., A threshold selection method from gray-level histograms, IEEE Transactions on Systems, Man, and Cybernetics 9, No. 1, (1979). [14] ( [15] P. S. Chavez, D. J. Mackinnon, R. L. R. M. G. V., Monitoring dust storms and mapping landscape vulnerability to wind erosion using satellite and ground-based digital images, Aridlands News Letter 51.