1. Introduction
Identifying storms from remotely sensed imagery such as radar and satellite fields is a fundamental problem in automated weather analysis and guidance tools. Once storms are identified, they can be used to compute storm properties for classification. Trends in the properties of the identified entities can be presented to a decision maker.
A storm in weather imagery is a region of high intensity separated from other areas of high intensity. The “intensity” depends on the weather data in question: reflectivity from weather radar, flash density from cloud-to-ground lightning observations, or infrared temperature from weather satellite may all be used. High intensities may not always correspond to storms. For example, high radar reflectivity could be the result of anomalous propagation. Low infrared temperatures may be due to the earth’s surface being cold and not because of high-altitude clouds. In practice, storm identification will have to be preceded by a preprocessing step that corrects for such sensor-specific details. The intensity threshold chosen as indicative of a storm depends on the field being segmented and the type of storms to be examined. Thus, an application interested in multicell storms would use lower reflectivity thresholds than one interested only in individual convective cells.
Storm identification in remotely sensed images has proven resistant to general approaches. The problems associated with storm identification can been side-stepped in some cases by processing rectangular subgrids of images. For example, Tag et al. (2000) classify clouds by processing 16 × 16 pixel areas while Rinehart and Garvey (1978) estimate motion by maximizing the correlation between constant-size subgrids of successive images. However, such methods compromise the ultimate classification or motion estimates: incorporation of storm properties rather than just region properties ought to yield better classification results while maximizing the match between the moving entities rather than just subgrids ought to yield better motion estimates.
A simple approach to identifying storms is to threshold the image based on a physically reasonable value. For example, Augustine and Howard (1988) describe an automated identification method for mesoscale convective complexes (MCCs). Candidate MCCs were identified by thresholding the satellite infrared image at a temperature of 221 K. Contiguous pixels were combined through a process known in image processing as region growing (Jain 1989) and the areas, durations, and eccentricities of these regions were used to determine whether the region in question was an MCC.
A simple schematic of using a threshold is shown in Fig. 1a. For clarity, a one-dimensional intensity function is shown. All of the discussions based on the 1D functions extend quite naturally to images: the position in an image would be a pixel and the intensity function would be defined on the x–y plane. Also, for clarity, all diagrams and text assume that cells consist of high-intensity regions. For cases such as infrared satellite imagery, where cold cloud tops are of interest, one would look for minima, not maxima, but the principles involved remain the same.
Using a single, hard threshold is subject to one glaring problem. The threshold is global in nature and applies to the entire image. Mature storms may be more intense and cover a large area, but initiating cells may have only a few points above the threshold. Thus, based on a single threshold, it is not possible to distinguish between noisy points (some of which may happen to be above the chosen threshold) and initiating cells, which may have only a few points above the threshold. To identify initiating cells in a different location, it may be desirable to lower the threshold, but this has the undesirable side effect of greatly increasing the area of mature cells. This is not a problem when analyzing large-scale systems after the fact, but is a critical shortcoming when identifying cells over large areas for guidance purposes.
To avoid the problem of initiating cells having only a few pixels above a threshold, a method called hysteresis (Jain 1989) can be employed. The term hysteresis comes from the lag observed between the application of an electromagnetic field and its subsequent effect on a material. In an image processing context, the lagging effect is provided by the use of two thresholds: one to start the thresholding and the other (a lagging one) to stop it. Thus, two thresholds are maintained and a cell is defined as being made of contiguous pixels above the second (lower) threshold as long as it contains a certain area above the first (higher) threshold. This is shown in Fig. 1b. The hysteresis technique can be further improved by using a threshold that varies throughout the image, rather than a single, global threshold.
Both the above improvements to the basic thresholding technique have been incorporated in storm identification methods on weather radar images. For example, Crane (1979) defined the hysteresis level to be 3 dBZ below the threshold being used. Rosenfeld (1987) defined the threshold to be used as the nearest local maximum in the neighborhood of the pixel, provided that it was at least 2 dBZ stronger than every other adjacent maximum. Thresholding techniques involving simple heuristics perform well on individual cases, but require multiple changes in order to work in operational situations where a diverse set of storms may be expected throughout any given year (Rosenfeld 1987).
In the United States, the operational way to identify storms from radar images follows the work of Johnson et al. (1998). Their method involves the use of multiple thresholds and counting runs of values above a threshold along a radial. The technique accounts for noise by incorporating a heuristic of how many successive range gates are allowed to fall below the threshold before the run is terminated. The centroids of such runs are then used as a proxy for the storms and tracked on the basis of proximity to the expected position. While these heuristics work well for radar images from the Next-Generation Doppler Radar (NEXRAD) network for which they were devised, they are not easily transferable to other types of weather imagery, or even to data from radars that are not broadly comparable to the Weather Surveillance Radar-1988 Doppler (WSR-88D). In addition, the technique of Johnson et al. (1998) involves iterating through the dataset once for each threshold. Such performance is tolerable when using five thresholds on an image corresponding to a single radar scan [approximately 365 × 460 pixels; the problem that the Johnson et al. (1998) was intended to address], but it does not scale well when either the number of thresholds or the size of the image is increased.
a. Terminology
Image processing terms are explained when they are first used in this paper. For the reader’s convenience, quick definitions of the terms are provided below:
watershed transform—an efficient technique for finding regions of high intensity by testing all possible thresholds in one pass through the image;
basin—a region identified by the watershed algorithm by starting out at a local maximum;
saliency—criterion used to check whether a region should be retained; saliency is usually defined based on the depth of the basin but in this paper, it is modified to be a check on the area of the basin;
hysteresis—a second, lagging threshold that is usually defined as a decrement from the first threshold; in this paper, the hysteresis threshold is defined as the threshold at which the basin becomes salient; and
foothills—the foothills denote the region of support for a maximum; within this region of support, no new maxima may be found.
b. Organization
The rest of this paper describes a general-purpose technique for the identification of cells in different types of weather images and demonstrates its effects on a variety of radar and satellite images. The basis of the technique, the watershed transform, is described in section 2. The conceptual framework behind the extension of the watershed transform to enable finding storm cells in weather images is described in section 3. The actual implementation details of the algorithm, including pointers on parameter choices, are provided in section 4. The results of using this algorithm to identify storm cells in a variety of remotely sensed weather images are illustrated in section 5.
2. Watershed transform
The term watershed, in common geographical usage, refers to a drainage basin. The drainage basin collects all the water within the area covered by the basin; that is, water does not flow out of a basin. The watershed transform of Beucher and Lantuejoul (1979) and Beucher (1982) borrows its conceptual model from the idea of watersheds as drainage basins divided by ridges of land. The watershed transform is a region-based object identification approach that is now the method of choice in the image processing community for identifying objects in images (Roerdink and Meijster 2001). Unlike conventional image processing transformations, which rely on convolution and have frequency counterparts, the watershed transform is morphological and sequential; that is, it proceeds on a pixel-by-pixel basis based on the attributes of a pixel’s neighborhood. In keeping with the terminology in most numerical and optimization methods, the watershed algorithm is normally described as a minima-finding method. Here, we shall describe it as a maxima-finding method so that the relationship to finding high-intensity storm cells is straightforward.
Conceptually, the idea is to flood the image starting from the global maximum. The flooding level is slowly decreased so that flooding can proceed at lower and lower levels, a process called immersion. The entire area covered by water flowing from a single maximum is a termed a basin and the points where two basins meet form region boundaries (see Fig. 2a). A key advantage of the watershed approach is the lack of a prespecified threshold; in effect, all possible thresholds are attempted.
While it is certainly conceivable to implement the simple thresholding approach at all possible thresholds,1 simulating the immersion process would involve iterating through the entire image once for each threshold. The watershed approach was impractical for real-world applications, until Vincent and Soille (1991) described an efficient algorithm to implement it.
The method of Vincent and Soille (1991) implements the watershed transform efficiently by iterating through the image and forming a data structure that maps, to each digital image value, all the pixel coordinates that contain that value. The structure is then sorted in descending order of intensity values. Flooding the image at a threshold is simulated by simply processing all the pixels at a particular level. At each such pixel, all the directions in which the intensity is reduced are identified and the neighboring pixel in that direction is added to the basin. All the pixels in the basin are similarly analyzed until one reaches a local minimum, where there is no direction in which to descend (see Fig. 2b). The entire set of pixels accumulated starting from a single pixel is termed a basin; this corresponds to a cell.
In Vincent and Soille’s (1991) algorithm, all basins, at all flooding levels, are found in one single pass through the image. A naïve implementation of a multiple-thresholding scheme may yield the same result, but it is impractical to process large remotely sensed geospatial fields in real time by iterating through the entire image once for each threshold level.
In real-world images, the conventional watershed algorithm produces exceedingly small regions because there is a cell for each local maximum. There are two potential solutions to this problem, which in one form or another affects all cell identification methods. One approach is to smooth the image beforehand to reduce the number of spurious maxima. The second approach, introduced by Najman and Schmitt (1996), is to sequentially merge neighboring basins whose saliency is below some prespecified threshold. Najman and Schmitt (1996) defined the saliency as the difference between the maximum level and the level at the basin boundary (see Fig. 2b). Practical implementations of the watershed transform employ both of these solutions (Roerdink and Meijster 2001).
The watershed transform, however, does not work with weather imagery (see Fig. 3b). The image in this figure shows the result of applying the method of Najman and Schmitt (1996) with some appropriate smoothing and saliency checks on an infrared channel image from the Spinning Enhanced Visible and Infrared Imager (SEVIRI) on board Meteosat-8. As is readily apparent, the cells identified by the watershed transform look nothing like storm cells. Indeed, they look like watersheds.
It has been noted (Lakshmanan et al. 2003) before that the watershed algorithm performs poorly on weather images. While a clustering approach does identify cells in a satisfactory way (see Fig. 3c), it has the typical drawback of a clustering approach (Witten and Frank 2005) in that the final clusters are very sensitive to the choice of initial centers. This means that associating the clusters identified between time frames is not possible (see Fig. 4). The association problem can be side-stepped when estimating the motion of the clusters by finding the spatial dislocation at which the cluster’s pixel values in the current frame are most correlated with the previous frame (regardless of what the clustering result was for the previous frame). While the motion estimates derived with such a method proved to be very robust (Yang et al. 2006), the drawback of not associating clusters is that extracting trends of properties of the cluster across time is not possible.
3. Extending the watershed transform: Conceptual model
Why does the watershed transform not work on weather images? How can it be extended so that it can be used to identify storm cells in weather images?
One problem lies in the definition of the saliency of a basin, as the height difference between the maximum from which water flowed and the point at which the water level meets that from another basin (see Fig. 2a). This definition of saliency does not correspond to the typical understanding of a storm cell as evidenced in the work of Crane (1979) or Rosenfeld (1987). Instead, as seen in the definition of a mesoscale convective complex (Augustine and Howard 1988), a storm entity is defined as salient if it meets a certain size threshold. In other words, it is not the depth of the basin, but the areal extent of the basin that counts. Thus, the first change to the conceptual idea behind the watershed technique is to redefine the saliency to be the areal extent of the basin, rather than the height from which the water has flowed. This is shown in Fig. 5a. It should be noted that this saliency criterion leads to early stopping of the basin creation process, as in the first basin in Fig. 5b. Thus, the size threshold chosen for saliency can lead to cells that are smaller than would have been realized if the immersion process had been carried out completely.
The watershed technique appears, at first glance, to incorporate multiple thresholds that are local, and not global, because a basin is defined in terms of the level at which it starts to get flooded. Different basins do have different starting levels. However, a basin is stopped either when the water level meets that of another basin or the value falls below a threshold shown by the dashed line at the bottom of the intensity plot in Fig. 5a. This threshold is the one that corresponds to the threshold level of a simple thresholding scheme such as the one in Fig. 1a. So, instead of defining a global minimum threshold, it is necessary to define the minimum threshold as a hysteresis level. The extension of the watershed transform with a hysteresis threshold is shown in Fig. 5b.
By defining the minimum threshold to be a hysteresis level, rather than a global one, basin mixing can happen, where points that would have been within the basin defined by one maximum fall into a basin defined by a secondary maximum simply because the basin creation was stopped. To avoid this problem, it is necessary to define a region of support for each maximum (we shall call them foothills; see Fig. 5c). This region of support is larger than the basin itself but consists of points that should not be assigned to a different basin.
It is tempting to define the hysteresis level to correspond broadly to the idea espoused by Crane (1979), of defining a cell in terms of a local maximum and a second threshold defined in terms of being a certain intensity level [3 dBZ in the case of the radar reflectivity images used by Crane (1979)] below that local maximum. However, there is no guarantee that basins created using such a definition will meet the saliency check. Recall that we have modified the saliency definition to correspond not to the height of the basin, but to the area. Thus, the hysteresis level also has to be redefined as the level at which the size of the basin meets the saliency criteria. To avoid noise contamination from spurious peaks, it is necessary to place an upper limit on the depth as well. Defining the hysteresis level in terms of areal extent is shown in Fig. 5d.
4. Implementation of the extended watershed transform
The conceptual framework behind the extension of the watershed algorithm used in this paper is described in section 3. In this section, we will describe the stages of the algorithm as it is implemented and the parameters that the user needs to specify, and also provide pointers on how to choose those parameters.
The stages of the algorithm are shown in Fig. 6. First, the input image is smoothed to reduce the occurrence of spurious peaks. Then, the image is quantized and transformed to a data structure that maps an intensity level to all the pixels at that level. The data structure is sorted and centers (the points from which flooding will commence in the conceptual model) identified. Cells (basins in the conceptual model) are then identified, taking the hysteresis level and saliency checks into account.
a. Smoothing
The use of a saliency criterion that is a combination of depth and distance helps to avoid defining basins based on short or narrow maxima. However, smoothing the image before performing the watershed algorithm is still useful. This is because, as shown in the top row of Fig. 7, spurious peaks that are part of a valid basin do change the definition of that basin. To achieve stable cell identification, it is important to smooth the images beforehand. Second, smoothing reduces the number of centers to be considered when trying to determine which maximum a pixel would have been flooded by, thereby increasing the efficiency of the foothills stage.
Smoothing provides an opportunity to set the scale at which storms are identified. For example, by smoothing images over large neighborhoods, it is possible to look for mesoscale systems and avoid storm-scale entities. However, as will be discussed shortly, the scale is better set using the saliency criteria. Thus, care should be taken to ensure that the scale at which smoothing is carried out corresponds to one that will maintain basins of the size that would be expected on the basis of the saliency criteria.
Any smoothing filter may be used, but different smoothing filters provide better results for different types of images. For example, a rank-based filter2 is better for images with speckled noise while an oriented filter is better for processing entities on the scale of squall lines (Lakshmanan 2004). A small-scale Gaussian filter is a good general-purpose filter because of its superior space–frequency properties (Jain 1989) and can be used when there are no other overriding considerations.
b. Quantization
Large values of δ can be used to reduce the dynamic range in images. This is useful if the instrument tends to yield noisy data, but where the spatial resolution is too poor for spatial smoothing to be employed for noise reduction.
c. Transformation
The quantized values are now placed into a data structure that enables the efficient simulation of the immersion process:
d. Finding centers
The next step is to find all of the candidate local maxima. The list of centers can be a subset of the list of pixels; it is not necessary to perform the full algorithm at pixels that are proximate to a local maximum because these pixels will either be part of that maximum’s basin or its foothills. Thus, the subset of pixels that needs to be considered as maxima can be obtained by iterating through the pixels in reverse order of intensity and removing from the list neighbors of those pixels that are candidate centers. However, the algorithm may be implemented, at the cost of some efficiency, without doing this prepruning by merely initializing centers to a copy of the pixels data structure.3
e. Immersion simulation
The output of the algorithm is a labeled image, where each pixel’s value is the basin that it belongs to. Pixels that are not part of any basin (storm cell) will be assigned a basin number of zero. First, all of the pixels in the labeled image, B (for basin), are initialized to a negative number. Then, the immersion simulation of the extended watershed algorithm is carried out by iterating through the centers’ data structure (see procedure 2 below).
Note that hlevel is slowly decreased (with increasing depth) so that the basin is grown only until it meets the saliency check. Capturing a center involves growing the basin starting at the center and capturing all points above the hlevel that would be immersed with water flowing from the center (see procedure 3 below).
Reserving all of the foothills also involves region growing starting at the center and capturing all points that are below the hlevel and to which the current center is the closest of all centers. The implementation is very similar to procedure 3 except that in addition to the threshold check on value, there needs to be another check; this time on the distance between the pixel xy and each center. Note that this distance needs to be geodesic (i.e., in the space of the image and including only those pixels that are greater than a). This is because the foothills of a basin should include only those pixels to which the basin is connected. As a fast approximation, the Euclidean distance may be used, and the geodesic distance verified only for points that are within that Euclidean distance.
f. Affecting the scale
The size of the cells detected by the extended watershed technique can be affected in two ways. Short and narrow peaks can be smoothed away by spatial filtering of the image prior to performing the immersion simulation. Thus, for example, smoothing the image with a filter with large spatial support (a large σ value) will lead to rounded peaks and larger basins. The direct means of affecting the scale is to set the saliency threshold appropriately. The saliency in the technique of this paper is the minimum spatial extent of a basin before it is regarded as a valid storm cell. By increasing the saliency, we can choose retain only those storms that have the required spatial scale (see Fig. 8). Note also that the early stopping provided by the saliency criterion causes the technique to behave more like a centroid-finding technique at small scales (Fig. 8d) and as a complete storm identification technique at larger scales (Fig. 8c).
How cells smaller than the saliency threshold should be treated depends on whether smaller, but contiguous, cells should be combined into larger lines and fronts. Recall that the hysteresis level is defined in terms of the maximum depth at which the saliency threshold for an area can be met. By decreasing the hysteresis level, smaller storms can be combined to create larger entities. Thus, to detect squall lines (where it is to be expected that the area between small cells has low intensity values), the saliency area threshold would be set high but the hysteresis level would be lowered. This is the algorithm presented in this paper. To avoid such merging of smaller cells into larger entities, the hysteresis level could be set to a prespecified decrement from the threshold level, and not allowed to be lowered below that. Preliminary analysis of doing this was not encouraging; it appears that basin merging is required to match human perceptions of what a valid storm cell is (see Fig. 9c).
It is possible to implement hierarchical storm detection by saving all the basins that meet the saliency criteria at each hysteresis level. Basins at a high hysteresis level are all contained within basins at any lower hysteresis level.
Note that while scale can be affected either by smoothing the image or by setting the saliency threshold, methods such as wavelet analysis cannot be used. Wavelet analysis is a multiresolution methodthe technique introduced in this paper needs to operate at the original resolution of the image.
g. Associating and tracking storms
Often, the main purpose of identifying storms in radar or satellite images is to track them across time. By tracking storms across time, it is possible to retrieve trends in storm properties and use these in nowcasting. While such tracking is not a focus of this paper, it is important to demonstrate that tracking is possible when storm cells are identified using the method described in this paper.
The major problem to be solved when tracking storms across time is to determine, for each storm in the previous image, the storm in the current image with which it is associated. This is a difficult problem because storms change with time. They may also split or merge with other storms. The method of carrying over properties of storms when storms split or merge also needs to be determined.
One heuristic for association, used by Johnson et al. (1998), is to project the location of every storm’s centroid based on its past movement and to then associate every storm in the previous image with the storm in the new image whose centroid is closest to that projected centroid. Another heuristic, used by Morel et al. (1997), is to associate storms based on the degree of overlapping pixels. Neither heuristic works well if the storms move or change significantly between images such as if the temporal gap between images is large or if the spatial resolution of the images is poor.
A somewhat better approach is that of Dixon and Wiener (1993) where storm association is treated as an optimization problem. Long-lived, larger, and more intense storms should be retained as much as possible. In addition, storms should be associated with the closest storm (as measured by the intercentroid distance) and with the storm with which there is considerable overlap. The cost function used could be a weighted sum of all of these utility functions, but the choice of the appropriate weights would have been heuristic and not easily transferred between different types of weather images. For simplicity, we used a greedy optimization approach that maximized the utilities in the above-mentioned order. Once associated, the storm movement was estimated using a constant-acceleration Kalman filter model (Kalman 1960; Brown and Hwang 1997).
5. Results
The storm cell identification method introduced in this paper was applied to several types of remotely sensed images.
a. NEXRAD
Radar reflectivity data from the NEXRAD network are combined in real time to create a 3D mosaicked dataset that covers the coterminous United States (CONUS) at a resolution of approximately 1 km × 1 km × 1 km (Lakshmanan et al. 2006). The effect of applying the technique of this paper to the reflectivity composite derived from that 3D mosaic is demonstrated in Fig. 9.
The parameter values used are shown in Table 1. The saliency threshold should be chosen depending on the scale at which storm identification is desired and whether or not multicell clusters are permissible. The radar data are quality controlled (Lakshmanan et al. 2007) as part of the data fusion process. In that stage, artifacts due to anomalous propagation and sun strobes are removed. Thus, a relatively light hand in terms of smoothing suffices for this dataset.
The technique takes 13 μs of CPU time (0.6 s of clock time) to process a 3000 × 6000 pixel image using a 1-GHz Advanced Micro Devices (AMD) Opteron processor.
b. CASA
The Collaborative Adaptive Sensing of the Atmosphere (CASA) network presently consists of four small, low-cost, low-power X-band radars in southwestern Oklahoma. Data from these radars are combined in real time and the cells identified using the technique of this paper (see Fig. 10) are used as input to a meteorological command and control unit that adaptively scans the atmosphere (Brotzge et al. 2005).
The CASA data are characterized by extremely high resolution with the range gates being 0.05 km × 1.8°. The pixel size in the merged radar grid is chosen to be 0.1 km × 0.1 km, similar in scale to the average radar gate sizes. However, the CASA data are subject to significant ground clutter problems because the CASA network (unlike the NEXRAD one) was designed to scan the lower atmosphere (0–3 km) and because thorough quality control has not been applied to the CASA reflectivity data. Thus, an aggressive rank filter was used in the smoothing stage and a high saliency threshold of 2.5 km2 (corresponding to 250 pixels) was applied (see Table 1).
c. GOES
The extended watershed algorithm described in this paper was applied to 1-min interval scans of the 11-μm infrared channel from the Geostationary Operational Environmental Satellite (GOES-12, eastern sector) on 12 October 2001 (see Fig. 11). The spatial resolution of the data was approximately 4 km × 4 km. Compared to the radar data, the spatial resolution of the satellite data is significantly poorer. Thus, storm cells are extracted at somewhat larger scales than is possible using radar data. Even a saliency threshold of 160 km2 (see Table 1) corresponds to a mere 10 pixels in the image. Also, no smoothing filter was applied because the infrared data are already quite clean and smoothing would further degrade the effective spatial resolution of the data.
d. SEVIRI
The algorithm was also applied to infrared brightness count data from the Meteosat-8 Spinning Enhanced Visible and Infrared Imager (SEVIRI) satellite. The SEVIRI satellite has imaging capabilities similar to the Advanced Baseline Imager (ABI) that will be deployed on board the GOES-R.
The infrared (11 μm) channel data used were from 5 January 2005 and were collected at 15-min intervals. The spatial resolution was approximately 2.7 km × 2.7 km. Because of the higher resolution than is available from GOES, it was possible to employ a Gaussian filter at the smoothing stage. As with the GOES data, the saliency thresholds had to be higher so as to obtain the storm cores (see Fig. 12). The dynamic range of the SEVIRI data is much higher than that of GOES. However, because of the relatively poor spatial resolution (as compared to, say, NEXRAD), the high variance cannot be addressed by spatial smoothing alone. Therefore, a higher δ was used (see Table 1). The impact of not using a higher δ is shown in Fig. 12c; note the large numbers of spurious cells that still do meet the saliency criteria.
e. Summary and conclusions
The underlying ideas in storm cell identification have not changed since the seminal papers of Crane (1979) and Rosenfeld (1987). Because these methods are too simplistic to work in practice, operational storm cell identification methods such as those of Johnson et al. (1998), Dixon and Wiener (1993), or Augustine and Howard (1988) have relied on a set of heuristics. The heuristics of each of the above techniques are tuned to the particular dataset they were applied to and cannot be easily applied to other types of meteorological data. Furthermore, these methods rely on multiple passes through the image data and therefore do not scale to high resolution or large datasets as produced by CASA, mosaicked NEXRAD, or SEVIRI. This paper thus pointed out the need for a sophisticated, efficient technique that would work on a variety of weather imagery formats.
The watershed transform of Vincent and Soille (1991) and the improvements made by Najman and Schmitt (1996) were introduced as the most popular object identification technique. This popularity stems from two sources: 1) the watershed algorithm tests all possible thresholds, removing the need to define a threshold to identify objects, and 2) the algorithm can be implemented in a single pass through the image. For several reasons, pointed out in detail in the paper, the formulation of the watershed transform does not apply well to the problem of finding storm cells. Hence, we adapted and extended the watershed algorithm to the storm cell problem. This involved changing the saliency criterion from the one proposed by Najman and Schmitt (1996); the introduction of hysteresis into the watershed algorithm, making the hysteresis level dependent on the saliency; and finally the use of a concept that we called foothills to avoid basin contamination.
The paper also provided the implementation details of the extended watershed algorithm. Like the conventional watershed transform, the method described in this paper relies on a single pass through the image to achieve cell identification at all possible thresholds within some practical quantization bounds. The implementation yields a robust, fast algorithm that is capable of processing an image with 18 million pixels in a fraction of a second on a conventional workstation.
The algorithm was demonstrated on a variety of radar and satellite imagery sources. Good choices for the parameters for the enhanced watershed algorithm were also discussed. In particular, we demonstrated its performance on high-resolution, large-scale radar reflectivity fields (mosaicked NEXRAD); extremely high-resolution, but very noisy, radar reflectivity fields (mosaicked CASA); a coarse-resolution infrared temperature satellite channel (GOES); and a moderate-resolution, global infrared temperature satellite channel (SEVIRI). Parameter choices for each of these cases were described and it was shown that the extended watershed technique successfully identified storms in all of these fields.
This paper presents a formal, efficient way of finding regions where a remotely sensed field’s values are greater (or lesser) than a locally varying threshold. If the colder infrared temperatures correspond to the ground temperature, or higher dBZ values in radar reflectivity correspond to anomalous propagation, the identified entities will not be storms. In practice, the technique outlined in this paper should be used to identify storm candidates, and the overall context used to determine whether the identified candidates are indeed storms.
Acknowledgments
Funding for this research was provided under NOAA–OU Cooperative Agreement NA17RJ1227, Engineering Research Centers Program (NSF 0313747). We acknowledge Dr. Robert Kuligowski of NESDIS and Dr. John Moses of NASA for their assistance with the satellite data.
The extended watershed algorithm described in this paper has been implemented within the Warning Decision Support System-Integrated Information (WDSS-II) as the w2localmax algorithm. It is available for download online (www.wdssii.org).
REFERENCES
Augustine, J., and Howard K. , 1988: Mesoscale convective complexes over the United States during 1985. Mon. Wea. Rev., 116 , 685–701.
Beucher, S., 1982: Watersheds of functions and picture segmentation. Proc. Int. Conf. on Acoustics, Speech and Signal Processing, Paris, France, IEEE, 1928–1931.
Beucher, S., and Lantuejoul C. , 1979: The use of watersheds in contour detection. Proc. Int. Workshop Image Processing, Real-Time Edge and Motion Detection/Estimation, Rennes, France, IRISA, 2.1–2.12.
Brotzge, J., Westbrook D. , Brewster K. , Hondl K. , and Zink M. , 2005: The meteorological command and control structure of a dynamic, collaborative, automated radar network. Preprints, 21st Int. Conf. on Interactive Information Processing Systems (IIPS) for Meteorology, Oceanography, and Hydrology, San Diego, CA, Amer. Meteor. Soc., 19.15. [Available online at http://ams.confex.com/ams/pdfpapers/85177.pdf.].
Brown, R., and Hwang P. , 1997: Introduction to Random Signals and Applied Kalman Filtering. John Wiley and Sons, 484 pp.
Crane, R., 1979: Automatic cell detection and tracking. IEEE Trans. Geosci. Electron., 17 , 250–262.
Dixon, M., and Wiener G. , 1993: TITAN: Thunderstorm identification, tracking, analysis and nowcasting—A radar-based methodology. J. Atmos. Oceanic Technol., 10 , 785–797.
Jain, A., 1989: Fundamentals of Digital Image Processing. Prentice Hall, 569 pp.
Johnson, J., MacKeen P. , Witt A. , Mitchell E. , Stumpf G. , Eilts M. , and Thomas K. , 1998: The Storm Cell Identification and Tracking Algorithm: An enhanced WSR-88D algorithm. Wea. Forecasting, 13 , 263–276.
Kalman, R., 1960: A new approach to linear filtering and prediction problems. Trans. ASME, J. Basic Eng., 82 , 35–45.
Lakshmanan, V., 2004: A separable filter for directional smoothing. IEEE Geosci. Remote Sens. Lett., 1 , 192–195.
Lakshmanan, V., Rabin R. , and DeBrunner V. , 2003: Multiscale storm identification and forecast. J. Atmos. Res., 67 , 367–380.
Lakshmanan, V., Smith T. , Hondl K. , Stumpf G. J. , and Witt A. , 2006: A real-time, three-dimensional, rapidly updating, heterogeneous radar merger technique for reflectivity, velocity, and derived products. Wea. Forecasting, 21 , 802–823.
Lakshmanan, V., Fritz A. , Smith T. , Hondl K. , and Stumpf G. J. , 2007: An automated technique to quality control radar reflectivity data. J. Appl. Meteor. Climatol., 46 , 288–305.
Morel, C., Orain F. , and Senesi S. , 1997: Automated detection and characterization of MCS using the meteosat infrared channel. Proc. Meteor. Satellite Data Users Conf., Brussels, Belgium, EUMETSAT, 213–220.
Najman, L., and Schmitt M. , 1996: Geodesic saliency of watershed contours and hierarchical segmentation. IEEE Trans. Pattern Anal. Mach. Intell., 18 , 1163–1173.
Rinehart, R., and Garvey E. , 1978: Three-dimensional storm motion detection by conventional weather radar. Nature, 273 , 287–289.
Roerdink, J., and Meijster A. , 2001: The watershed transform: Definitions, algorithms and parallelization strategies. Fund. Inform., 41 , 187–228.
Rosenfeld, D., 1987: Objective method for analysis and tracking of convective cells as seen by radar. J. Atmos. Oceanic Technol., 4 , 422–434.
Tag, P., Bankert R. , and Brody L. , 2000: An AVHRR multiple cloud-type classification package. J. Appl. Meteor., 39 , 125–134.
Vincent, L., and Soille P. , 1991: Watersheds in digital spaces: An efficient algorithm based on immersion simulations. IEEE Trans. Pattern Anal. Mach. Intell., 13 (6) 583–598.
Witten, I., and Frank E. , 2005: Data Mining. Elsevier, 524 pp.
Yang, H., Zhang J. , Langston C. , and Wang S. , 2006: Synchronization of multiple radar observations in 3D radar mosaic. Preprints, 12th Conf. on Aviation, Range and Aerospace Meteorology, Atlanta, GA, Amer. Meteor. Soc., P1.10. [Available online at http://ams.confex.com/ams/pdfpapers/104386.pdf.].
Thresholding with hysteresis. (a) The basic technique involving the use of a global threshold works only for mature cells that have many points above the threshold. The second cell from the left appears to be very small and can be easily assumed to be noise. (b) Employing two thresholds, a high one to identify potential storms and a lower hysteresis threshold within which regions can continue to grow and can capture growing cells while remaining noise tolerant. Thus, the second cell is now captured, but the third is rightly ignored.
Citation: Journal of Atmospheric and Oceanic Technology 26, 3; 10.1175/2008JTECHA1153.1
The watershed transform. (a) Conceptually, the idea is to flood the area one level at a time, starting from the global maximum and to label all of the area covered by water from the same maximum as being a basin. (b) Starting at a pixel at the flooding level, neighbors in the directions of decreasing intensity levels are added to the basin.
Citation: Journal of Atmospheric and Oceanic Technology 26, 3; 10.1175/2008JTECHA1153.1
The watershed algorithm does poorly when applied to weather images. (a) Infrared channel 9 (11 μm) image from the SEVIRI satellite. The image is over the North Sea on 5 Jan 2005 and depicts an area of approximately 600 km × 400 km. (b) Cells identified by the watershed algorithm. The gray shades have no significance beyond providing a way to distinguish between different cells. (c) The multiscale clustering approach of Lakshmanan et al. (2003) does well, but has the disadvantage of it not being possible to extract the trends of these clusters. This is because clustering is sensitive to the initial starting point. See also Fig. 4.
Citation: Journal of Atmospheric and Oceanic Technology 26, 3; 10.1175/2008JTECHA1153.1
Clustering techniques identify valid cells, but are very sensitive to starting points. (a) GOES band-4 data from 27 Jul 2005. The area shown covers an area of about 600 km × 400 km centered over AR. (b) As in (a), but for data from 5 min later. Note that little has changed in the two panels. (c) Cells identified at the most detailed scale at the first time step. (d) As in (c), but for cells identified 5 min later. Note the drastic differences in the two panels; several of the cells identified in the first time frame are not identified at the second time frame, even though the input images are nearly the same. It is not possible to track cells across time if the identification process is that sensitive.
Citation: Journal of Atmospheric and Oceanic Technology 26, 3; 10.1175/2008JTECHA1153.1
Extensions to the basic watershed algorithm so that it can be used to detect storm cells in weather images. (a) The saliency needs to be defined in terms of areal extent, rather than the depth of the basin. (b) The minimum threshold of the basins should not be global, but should instead be a hysteresis level (hlevel in this panel). (c) While the basin is defined in terms of hysteresis level and saliency, a larger region of support, termed foothills, needs to be defined. (d) The hysteresis level needs to be defined in terms of the areal extent subject to a maximum depth.
Citation: Journal of Atmospheric and Oceanic Technology 26, 3; 10.1175/2008JTECHA1153.1
Stages in the implementation of the cell identification algorithm. The dashed box is the stage at which immersion is simulated while the transformation stage is responsible for the efficiency of the method. The immersion stage consists of two iterations through a data structure: one through all the scales and the other through all the levels.
Citation: Journal of Atmospheric and Oceanic Technology 26, 3; 10.1175/2008JTECHA1153.1
(a) The saliency criteria prevent basins being formed because of narrow or short peaks such as the second peak in the figure. However, spurious peaks like the third that occur in a valid basin cause the basin boundaries to be poorly defined. (b) Smoothing the image before performing the enhanced watershed algorithm makes the basins cleaner. (c) The quantization function allows the values integers to take advantage of the efficiency of the watershed transform. (d) Quantization can be used to transform the image values, so as to look for cells defined by colder values in fields such as satellite infrared temperature.
Citation: Journal of Atmospheric and Oceanic Technology 26, 3; 10.1175/2008JTECHA1153.1
Affecting the scale at which storms are identified using the technique of this paper through smoothing and/or saliency. (a) Input is a multiradar reflectivity composite image covering an area of approximately 300 km × 300 km over the Texas Panhandle on 8 May 2007. (b) No smoothing; saliency of 400 km2. (c) Smoothing using a Gaussian filter of σ = 3 km; saliency of 400 km2. (d) Smoothing via a Gaussian filter of σ = 3 km; saliency of 50 km2.
Citation: Journal of Atmospheric and Oceanic Technology 26, 3; 10.1175/2008JTECHA1153.1
The extended watershed algorithm of this paper applied to a multiradar reflectivity composite image. The image covers an area of approximately 300 km × 300 km over the Texas Panhandle and comes from 8 May 2007. (a) Radar reflectivity. (b) Cells identified at a saliency of 100 km2. The positions of the cells as they are tracked are superimposed on the image. (c) Cells identified at a saliency of 400 km2. Note that smaller cells have been combined into larger entities to meet the saliency criterion.
Citation: Journal of Atmospheric and Oceanic Technology 26, 3; 10.1175/2008JTECHA1153.1
The extended watershed algorithm of this paper applied to a reflectivity composite image from the CASA network. The image covers an area of approximately 120 km × 120 km over Oklahoma and comes from 8 May 2007. (a) Radar reflectivity. The rings denote the ranges of the four CASA radars used in creating this mosaic. (b) Cells identified in white, with their foothills in gray. Note that even though the data are noisy, the aggressive smoothing and saliency settings in Table 1 help reduce the impact of the noise. (c) Identified cells being used in the meteorological command-and-control unit of the CASA system.
Citation: Journal of Atmospheric and Oceanic Technology 26, 3; 10.1175/2008JTECHA1153.1
The extended watershed algorithm of this paper applied to a GOES-12 11-μm infrared image. The image covers an area of approximately 4000 km × 2500 km over the Gulf of Mexico and comes from 12 Oct 2001. (a) Infrared channel. (b) Cells identified at a saliency of 160 km2. The positions of the cells as they are tracked are superimposed on the image. A clustering technique (see Fig. 4) is not consistent enough to provide such tracks. (c) Cells identified at a saliency of 500 km2.
Citation: Journal of Atmospheric and Oceanic Technology 26, 3; 10.1175/2008JTECHA1153.1
The extended watershed algorithm of this paper applied to a SEVIRI infrared image. The image covers an area of approximately 1000 km × 750 km over northern Europe and comes from 5 Jan 2005. (a) Infrared 11-μm channel. (b) Cells identified at a saliency of 160 km2 and δ = −20 K. The cells have a gray scale following an intensity scale based on the minimum IR temperature within the cell. (c) Impact of choosing δ = −1: note the large number of spurious cells.
Citation: Journal of Atmospheric and Oceanic Technology 26, 3; 10.1175/2008JTECHA1153.1
Parameter values used for the different datasets. The a, b, and δ reflect the physical characteristics of the fields being processed. For example, the δs for the infrared datasets are negative to enable the algorithm to search for cold tops (lower temperatures). A large δ value can be used to address noise in images with poor spatial resolution, as seen in the SEVIRI parameters. The saliency threshold reflects the spatial resolution of the data. Thus, it can be extremely small for CASA but needs to be quite high for GOES. The smoothing parameters reflect the noisiness of the data. Thus, the noisy CASA dataset was subjected to aggressive smoothing using a rank filter while the clean GOES dataset was not smoothed at all.
When applied to digital images, there is only a finite number of candidate thresholds. These correspond to the actual data values to be found in the image.
A rank-based filter is a filter that sets a pixel’s value based on the sorted values of its neighbors. The most commonly used rank-based filter is the median filter, where the output is the 50th percentile value within a specific neighborhood. Percentile values greater than 50% lead to dilation of storm entities, and values less than 50% lead to erosion of storm entities.
In a two-dimensional image, there are eight neighboring pixels around each local maximum, so prepruning will lead to at least an eightfold increase in the efficiency of the immersion simulation stage.
