1. Introduction
For the meteorologists of the 1970s and 1980s, manual inspection of satellite imagery and synoptic maps for features of interest was commonplace. Hours of work analyzing series of images by hand substantially advanced our knowledge of the earth’s meteorology (e.g., Streten 1973; McGuirk and Ulsh 1990). With the continued growth of volume of observational data available and computing advancements, automated feature detection in meteorological data became necessary and possible.
Many object-based methodologies, as they are commonly referred to, have been developed to identify weather features. Most widely applied examples include tropical and extratropical cyclone tracking (e.g., Murray and Simmonds 1991a; Benestad and Chen 2006), cloud tracking including mesoscale convective systems (Carvalho and Jones 2001), and remotely sensed precipitation systems (Skok et al. 2009).
Feature detection is the cornerstone of all these methodologies; a challenge that the field of computer vision is continually addressing. One of the primary goals of this field is to provide computers with humanlike visual processing abilities, an aim of all weather-feature-tracking algorithms. Hodges (1994) was perhaps the first to explicitly point this out and bring computer vision advances to bear. Scharenbroich et al. (2010) and Bain et al. (2011) are recent examples of applying the huge leaps in computer vision and machine learning techniques to improve tropical storm tracking and ITCZ detection.
These methodologies have enabled automated production of climatologies of the various meteorological features they have been built to detect (e.g., Murray and Simmonds 1991b; Hodges and Thorncroft 1997; Blamey and Reason 2012). Their use in model evaluation and intercomparison has provided invaluable insight into model performance (e.g., Hodges et al. 2011). Perhaps the most elegant example of an objective automated feature-tracking procedure is presented by Hewson and Titley (2010). Their methodology identifies frontal systems and follows their evolution from the earliest point in their life cycle by incorporating mathematical definitions of extratropical cyclones into a feature identification procedure. A primary application is in producing storm-track diagnostics from ensemble weather forecasts to aid human forecasters.
These object detection methodologies exploit key identifying features of the weather systems they are built to track; however, they rarely provide a full meteorological description of these systems. This deficiency limits their use in describing and quantifying the mechanisms that facilitate scale interactions between weather and climate. Therein lies the thesis of this paper; a tool that can describe the full meteorology of many synoptic systems will be useful for studies attempting to link weather to large-scale climate variations.
Over subtropical southern Africa, a substantial portion of wet season rainfall is attributed to tropical–extratropical (TE) cloud bands (Harrison 1984), known regionally as tropical temperature troughs (TTTs; Fig. 1). Hart et al. (2010) described three heavy rainfall producing TTTs and by comparison with previous studies, summarized their typical meteorology for the region. Here that study is extended further to investigate scale interactions between these synoptic systems and large-scale climate modes. To this end, a methodology is developed and presented in this paper.
A TE cloud band over southern Africa at 1200 UTC 19 Jan 2010 (Meteosat-9 channel 9; image courtesy of EUMETSAT).
Citation: Monthly Weather Review 140, 12; 10.1175/MWR-D-12-00127.1
The goal is to automate the synthesis of the meteorology of each cloud band, as described in gridded datasets such as reanalyses and general circulation model outputs. Key criteria for the success of the method include the following:
A robust and simple approach to flag potential cloud band events;
A flexible method to recognize regions of interest in multiple meteorological fields;
The ability to synthesize regions of interest into a comprehensive description of the meteorology.
2. Methodology
a. Data
The Liebman and Smith (1996) optimally interpolated, outgoing longwave radiation (OLR) is used for the period 1979–2011 as the main observational dataset. The daily mean OLR is the variable used to flag cloud band events. This is complemented by the National Centers for Environmental Prediction–Department of Energy Reanalysis II (NCEP2) data for the same period, providing a description of thermodynamic and momentum variables (Kanamitsu et al. 2002). Example systems are confined to 1979–99, a period covered by the Water Research Commission (WRC) daily station rainfall dataset containing 7665 stations within South Africa (Lynch 2003).
b. Building a metbot
Analysis of atmospheric data is primarily performed by humans visually analyzing plotted data; this is what the approach tries to automate. Human experts can summarize images and identify notable features easily, for example, low pressure cells represented in sea level contour maps, or clusters of cold cloud. In short the problem can be stated as follows: how do you enable a digital system to pick out meteorologically interesting features?
The analysis platform was Python: the Matplotlib with the Basemap toolkit provided plotting capabilities (Hunter 2007; Whitaker 2007), the NumPy module handled arrays (Oliphant 2006), the Python Imaging Library (PIL) module was used for image manipulation, and the Python wrapper for OpenCV performed more advanced image manipulation (Bradski 2000). Wrappers for CVblob C++ library (Liñán 2008), built on OpenCV, identified blobs in images and obtained details of their location, size, and orientation. These open-source computer vision libraries are developed by the robotics research and applications community as part of the Robot Operating System (ROS) project.
The image analysis process is straightforward:
Stretch raw data, for example, outgoing longwave radiation, onto the grayscale image range [0, 255].
Translate a chosen threshold for the raw variable into grayscale.
Plot the image and fetch it from buffer to obtain a visual interpolation of data.
Apply a threshold to this image array to produce a binary image and perform connected component labeling to identify blobs.
Run appropriate filter through blobs to keep those of interest to your application, the “metblobs.” In practice, filters were only used with rigor to flag cloud band blobs. They needed to have a certain latitudinal extent and be positively titled (see below). With all other features a minimum area value filtered out the very small blobs.
Retain information describing blob contour, centroid position, orientation, and area.
Repeat at each time step.
The implementation of the Chang et al. (2004) connected component labeling algorithm in Cvblob is the cornerstone of this methodology. It is a methodology for image segmentation. A short summary follows, however, a detailed explanation of the process in general can be found in Hodges (1994). Connected component labeling performs blob detection in a binary image by identifying adjacent pixels and labeling them as such. The process iterates until all pixels are labeled. Blobs thus emerge, and image moments are used to calculate their orientation, area, and centroid (Fig. 2). This provides an intuitive summary of an image with massive reduction in dimensionality of data. For example, interpolated OLR at relatively coarse 2.5° resolution may yield a 15 × 25 point latitude–longitude grid for a domain subset over southern Africa. These 375 data points may be reduced to 4 blobs of low OLR with position, area and orientation giving each blob 4 attributes; 16 meteorologically informative data points summarizing 375.
(top) Raw OLR and (bottom) the results of connected component labeling during the flagging process.
Citation: Monthly Weather Review 140, 12; 10.1175/MWR-D-12-00127.1
Thresholding the images presents difficulty when the choice of an appropriate value is not clear. The frequency distribution of OLR values is bimodal (Fig. 3) describing a peak for cold cloud and one for clear skies so threshold choice is obvious. For geopotential height this is not the case; however, the Laplacian 
Probability distributions of gridpoint values for (top) OLR and (bottom) ∇2Φ850. Triangles indicate threshold values (see section 2b).
Citation: Monthly Weather Review 140, 12; 10.1175/MWR-D-12-00127.1
The frequency distribution of 
The goal was to identify and describe tropical–extratropical cloud bands (Fig. 1). OLR was used to flag candidate events by ensuring blobs extended from 20° to 40°S as a contiguous band and were positively tilted (the bearing from their continental root poleward was in the interval [95°, 180°]). Figure 2 illustrates this flagging process. An area criterion was used to remove smaller features such as the yellow blob, thereby reducing the number of blobs tested for the other criteria. The large blob meets both the required latitudinal extent and positive tilt requirement as indicated by the angle of the green line crossing through the blob centroid. The tilt criterion often failed due to contamination low OLR values along the ITCZ, producing blobs that extended eastward in the tropics. This produced negatively tilted angles. This problem was largely solved by only applying the positive tilt criterion, in a smaller domain (23°–33°S), which only considers the subtropical portion of the band, while retaining the larger domain to ensure TE connection. Input NCEP2 data are reduced by only selecting for time periods 48 h on either side of these flagged OLR features.
The meteorology was described by calculating metblobs for geopotential height at 850, 500, and 250 mb (∇2Φlev), u and υ wind at 850 and 250 mb, and moisture transports at 850, 700, and 500 mb. Jet regions in both the wind and moisture flux fields were obtained by applying the Laplacian to the vector magnitude (∇2|〈u, υ〉lev|). Again, somewhat arbitrarily, the amplitudes A = 1.0 and A = 0.5 were deemed appropriate to keep a useful level of detail in both the wind and moisture flux, respectively.
Meteorologists synthesize image features from different levels of different variables into coherent understandings of synoptic weather systems. Approximating this ability was the second key goal of building a metbot. It essentially involved further abstraction of the detected blobs on different levels. Tracks of blobs are built and then associated with tracks of the cloud band blobs. These associations are represented in events.
Blob feature synthesis:
Each set of metblobs is tracked by checking if there is any blob within a predefined radius in the subsequent image and matching to the closest if more than one exists.
Tracks of the reference blobs, OLR in this case, are checked for blobs that were flagged as cloud bands. If they do contain flags, they are kept as the base track, thus instantiating an event.
Each event is associated with tracks of other metblobs, which need to overlap in time within a predefined radius of the base track for a predefined period.
Finally, an array of each event describing its base track and associated tracks is formatted. Thus, each event describes the evolution of a suite of meteorological features associated with a flagged cloud band at some point in their shared life spans.
3. Results
For the datasets described above, the metbot identified 821 cloud band events over the period 1979–2011 in the domain 0°–60°S, 0°– 80°E. Figure 4 displays a climatology of cloud band positions produced by calculating the annual average for how many times a grid point falls within any flagged cloud band’s contour. This figure displays the two preferred locations for tropical–extratropical interaction in the southwest Indian Ocean, a feature identifiable in principal component analysis of daily OLR (Todd and Washington 1999), and indeed noted by early manual satellite imagery analysis (Harangozo and Harrison 1983; Harrison 1984).
Seasonal gridpoint frequency for cloud bands.
Citation: Monthly Weather Review 140, 12; 10.1175/MWR-D-12-00127.1
The potential usefulness of this database, however, lies in the meteorological detail it retains in of each of those events. Obtained results are documented here by projecting the database back into the raw data field from which it was built. This process is achieved simply by using the metblob contours to create arrays of data masks for the meteorological variables of each event.
Hart et al. (2010) documented two heavy rainfall events that occurred in close succession over South Africa during 1–8 January 1998. This wet spell is used as an example to illustrate the success and limitations of the metbot to highlight synoptic features. Figure 5 presents selected times for this event. This figure also serves to illustrate how time is represented relative to event start dates.
Synthesis of meteorology for cloud band event 1–8 Jan 1998: 250-mb jets (black arrows), 850-mb moisture jets (red arrows), 250-mb troughs (gray dashed line), 850-mb depressions (magenta dashed line), OLR (shaded grayscale), and WRC rainfall within (jet color map dots) and outside (cool color map dots) the OLR contour. (Dates are displayed below each panel, see text for panel title explanation.)
Citation: Monthly Weather Review 140, 12; 10.1175/MWR-D-12-00127.1
The general format of the time header for each panel in Fig. 5—basetrack number:timecode—is explained here. Since the time step for OLR, the reference variable used to flag cloud bands, is daily, each flagged time step is denoted Dn_hr with n being the number of the flagged day for a given base track. The time stamp of the NCEP2 variables is contained in hr. Hence Figs. 5c,d, respectively, show meteorology at 0000 UTC on the first day an event is flagged and the second day it is flagged, in this case 2 and 7 January 1998. Note the 4-day separation between flags; this can occur since one body of low OLR values was tracked through this time as indicated by the basetrack number. Figures 5a,b display NCEP2 variables 36 and 24 h, respectively, before flagging of a cloud band in Fig. 5c. Each panel describes the suite of meteorological features that the metbot related to the base track.
Although ∇2|〈u, υ〉250| was used to detect jet regions in the 250-mb wind field, the raw wind field is displayed in these panels (black arrows). The same follows for 850-mb moisture fluxes (red arrows). Blob contours of ∇2Φ250 (light gray) and ∇2Φ850 (magenta) are plotted to indicate locations of detected depressions. Raw OLR values within the metblobs contour are shaded. Station rainfall under the OLR footprint is displayed with the jet color map (top-right color bar), wet (>5 mm) stations outside the OLR contour are shaded with the cool color map (bottom-right color bar).
Hart et al. (2010) discussed the meteorology of this event in detail, so this discussion only highlights features that the metbot emphasized for this event. It should be noted that the system was not tuned to this event and thus some shortcomings are revealed that suggest important caveats.
Thirty-six hours (Fig. 5a) before the event was flagged, a northeast moisture conveyor was present across the central subcontinent. A weak upper-tropospheric jet was positioned over Namibia. Enhanced upper-level winds extended up the east coast of Mozambique, upstream of an upper-tropospheric trough situated near Madagascar. The signature low-level warm conveyor and upper tropospheric jet of a midlatitude cyclone was present south of southern Africa. Twelve hours later (Fig. 5b), heavy rain was falling under a band of low OLR extending southeastward off the continent, into the midlatitude cyclone. Cyclonic low-level moisture movement was centered around a depression over Botswana. The Namibian jet had gained a more poleward orientation and the Madagascan upper-tropospheric trough had developed deep convection along its leading edge.
Figure 5b indicates the first problem encountered with the flagging process. This day should have been flagged as a cloud band, however, it failed the positive tilt criterion within the tilt subdomain (23°–33°C, see section 2b). This occurred since it was broad (Fig. 5b) and the blob extended northwest toward Madagascar. These cases of failed tilt criterions were rare and since at some point in a system’s life cycle it should pass the criterion, this problem is not crucial.
The event was flagged as a cloud band on the 2 January 1998 (Fig. 5c) as the Madagascan cloud fell below the OLR threshold. Rainfall had eased over South Africa, however, intensification of the low-level depression and associated moisture flow was evident by the expansion of the magenta blob contour and moisture flux vectors. This intensification continues through the day in time codes D1_6 to D1_18. While not shown, this illustrates that the metbot retains information at the time resolution of the NCEP2 data. Subdaily variability in meteorology is especially important for the tropical moisture fluxes.
By this stage, an upper-tropospheric jet had developed along the axis of the cloud band, strengthening poleward ahead of an upper-tropospheric trough associated with the extratropical cyclone. Deep convection continued over the southwest Indian Ocean with the metbot still noting an OLR blob that it then tracked “backward” onto the continent to flag a second day in this event. The continental low-level depression was still present with poleward moisture fluxes contributing to a widespread daily rainfall total of ~10 mm over much of South Africa. Upper-tropospheric jet regions south of the continent were present, while the Madagascan upper-tropospheric trough remained strong, with a jet zone still present on its western flank. An extended moisture conveyor was collocated with low OLR values downstream of this depression.
This second day that was flagged raises a concern about the integrity of the blob-tracking algorithm which is, admittedly, crude. From prior knowledge of cloud band dynamics, there is an expectation that a second event should have been created, however, due to the coarse temporal resolution of the data, and rapid deformation of patterns in the OLR field, the metbot tracked this blob back on to the continent. Applying a more rigorous tracking algorithm may address this issue; the method needs to be less sensitive to large variations in blob centroids that occur with drastic blob deformation as more cloud meets the OLR threshold. The maximum spatial correlation tracking technique (MASCOTTE) could present a solution (Carvalho and Jones 2001).
Using data with a higher temporal resolution, such as the 8-times daily Cloud Archive User Services (CLAUS) dataset used in the three-dimensional object identification method in Dias et al. (2012), would also mitigate these tracking caveats. However, extensive time coverage of the daily OLR data used here is necessary for further work on interannual variability that uses this methodology.
Attempting to better understand rainfall variability underpins why the methodology described here was developed. Figure 5 indicates how rainfall can be precisely attributed to the cloud band presence. While significant rain falls outside the deep convective footprint of low OLR (Fig. 5d) the core event rainfall does lie under this footprint. Delineating stations through masks (as in Fig. 5) will prove to be useful in calculating rainfall metrics. However, for long-term variability studies, metrics summarizing this information may be more appropriate. Table 1 presents summary statistics for each day of the cloud blob track. Flagged cloud band days are boldfaced. Within the OLR footprint the mean and maximum rainfall is calculated along with percentage wet stations (>1 mm) and percentage heavy rainfall stations (>50 mm). OutMean, OutMax, and OutWet summarize the same values outside the footprint.
Rainfall summary statistics for event 1–8 Jan 1998 (mean/outmean, max/outmax values are in mm day−1, wet/outwet and heavy values are frequencies). Flagged cloud bands are in boldface.
Wetness percentage outside the cloud contour is substantial toward the end of this event (Table 1, 7 January 1998), as indicated in Figs. 5b–d. However, during the first period of heavy rain, almost all of it is within the contour; 1 January 1998 clearly had the heaviest falls with a daily mean station rainfall of 23 mm and maximum of 386 mm. This table of rainfall information is built into each event object, allowing intercomparison of system intensities across all events.
The following question now arises: are any of these features in Fig. 5 common across events? The contours defining the edge of each metblob are exploited to calculate the frequency of each grid point falling within the contour of a chosen meteorological feature. Figure 6 presents these values for all events that produced rainfall in the WRC dataset within the period [−24, +24 h] and had a centroid west of 45°E. This longitudinal delineation is appropriate since Fig. 4 suggests two preferred locations of cloud band formation. All variables discussed in Fig. 5 are presented. The reader is reminded that the jet regions for 850-mb moisture flux and 250-mb winds, qv_850 and v_250, respectively, are calculated using the ∇2|〈u, υ〉lev|metblobs contours; hence, the same image applies for qu_850 and u_250 variables even though only the υ components are presented here.
Frequency (n = 268 events) that a grid point falls within a metblob’s contour for features presented in Fig. 5.
Citation: Monthly Weather Review 140, 12; 10.1175/MWR-D-12-00127.1
By choice of OLR blobs for the basetrack, its gridpoint frequencies (Fig. 6; olr_0) are the highest of all variables. This also illustrates the sharp east–west gradient in cloud band frequency across southern Africa, closely resembling the annual rainfall gradient (not shown).
Upper-tropospheric depressions (hgt_250) are detected over the southwestern coast in up to 20% of days accounted for by the methodology. An upper-tropospheric jet (v_250) frequency maxima of 0.25 occurs equatorward of this region as expected. The slightly higher jet frequency may suggest the threshold for depression detection needs to be relaxed slightly as upper-tropospheric troughs and jets should be closely linked. Nevertheless, occurrence of upper-tropospheric troughs upstream of cloud band development is widely noted and expected.
The association of the central subcontinent low-level depression with cloud band development is supported by the frequency field for hgt_850, agreeing with previous studies (e.g., Cook 2000), indicating a quasi-stationary low pressure system, the Angola low. The presence of a depression on the southwestern tip of Madagascar during up to 15% of event days suggests a synoptic feature that has received little attention.
The moisture flux frequency field (Fig. 6; qv_850) exhibits a number of features, despite the coarse resolution of the data. First, moisture jets over the East African coast are common. Second, a broad region of strong moisture flux exists across Botswana and Zimbabwe, a feature present in the case studies of Hart et al. (2010). Third, moisture jets over the Agulhas Current do occur in association with cloud bands. Finally, midlatitude cyclones and their strong moisture fluxes frequently occur at the poleward end of cloud band systems.
Using frequency thresholds as a filter to retain the most common features, Fig. 7 presents a composite event for continental rainfall-producing cloud band systems. Positions of the cloud band centroids (blue dots) for all events used in the composite illustrate the zonal distribution of these systems. Plotting the composite in coordinates relative to these centroids has been tested. This process did sharpen the band of OLR and tighten up the vector fields; however, beyond this there was little qualitative difference to the panels presented here.
Composite event evolution for all events that produce continental rainfall in WRC station dataset: 250-mb troughs (shaded blue), 850-mb depressions (shaded orange), OLR (gray shading), with 250-mb jets (black arrows) and 850-mb moisture jets (red arrows), and OLR blob centroid position (blob).
Citation: Monthly Weather Review 140, 12; 10.1175/MWR-D-12-00127.1
The accepted generalization for the TTTs is apparent from Fig. 7c. An upper-tropospheric trough lies off the west coast, with enhanced upper-level flow on its leading edge over southern Africa. Northeast moisture fluxes across the western subcontinent help sustain deep convection in a band of cloud that terminates in a midlatitude cyclone southeast of the continent. The continental low-level northeasterlies appear, in part at least, due to the presence of the Angola low. This emphasizes a key role the Angola low may play in supplying moisture to TTTs (Cook et al. 2004). Its relation to summer rainfall variability (Rouault et al. 2003; Reason et al. 2006) and the potential importance in modulating the regional response to ENSO (Reason and Jagadheesha 2005) has already been explored.
Lack of coherent structures in Figs. 7a,d indicate that large event-to-event differences exist in the early and decaying stages of TTTs life cycles. Post-TTT intensification of tropical convection is indicated by the lower OLR values across Zimbabwe and Zambia in Fig. 7d.
Without more appropriate diagnostics, it is unwise to say too much about baroclinic wave growth in the region from this figure. Nevertheless, it is clear that wave growth does occur as suggested by the equatorward extension of the upper-tropospheric trough from −24 to 0 h. This growth is expected by theory because of both poleward advection of warm moist air ahead of the trough and strong convective heating that is likely, within the band of cloud.
4. Discussion and conclusions
In this section, the metbot is put into context in three research domains: South African rainfall variability, subtropical convergence zones, and object-based methods for atmospheric science.
These results present an alternative method to the generalization of TTTs. Previous TTT composites have been based on daily precipitation principal component (PC) extremes (Todd and Washington 1999) or partitioning of days by cluster analysis (Fauchereau et al. 2009; Crétat et al. 2010). These composites were presented as circulation anomalies, whereas the raw circulation field is presented here. The findings support earlier generalizations and give weight to the common synoptic features highlighted in Hart et al. (2010), based on analysis of available case studies. As expected, however, there is substantial noise in the meteorology of individual event as shown by the low frequencies in Fig. 6. Conducting the analysis with a subset of events that caused extreme rainfall may reduce this noise and reveal higher frequencies of some of these features, if they manifest more clearly during intense events.
We, however, argue that explicitly retaining these event-to-event differences is the strength of this methodology over the previous methods. The literature has long noted the infrequency of continental TTTs (e.g., Washington and Todd 1999). Thus, individual events could substantially modify season total rainfall, especially if they exhibited unusual persistence or occurred farther west than expected in climatology. For example, note the few number of cloud band centroids over the semiarid western South Africa in Fig. 7c. These represent rare events in a water-scarce region. The lack of dependence of the metbot on PC or cluster analysis, which favors recurrence, is its strength. It allows exploration of the interannual variability in extremes that it can explicitly identify. The ability to associate rainfall summary statistics with individual events further expands its use in this regard. The time code value n (Dn_hr, as explained in section 3) also provides a simple measure of persistence of cloud band conditions, an important property when considering intraseasonal wet spell variability and/or contribution to total season precipitation.
By nature of its geometry, the cloud band flagging process has potential use for two well-documented atmospheric features: subtropical convergence zones and tropical plumes. A detailed description of the synoptic variability of the south Indian Ocean convergence zone is beyond the scope of this study, however, this methodology could be well suited to such an investigation. Its application to other convergence zones may help reveal nuances that PC analysis smoothes away; it could be interesting, for example, to repeat a study such as Matthews (2012) with this approach perhaps using the 3-hourly CLAUS dataset (Hodges et al. 2000). The metbot provides an avenue to explicitly capture the synoptic pulses fundamental to the South Pacific convergence zone dynamics (Matthews 2012). Statistics of these pulses could be built up to offer a more detailed description of these dynamics than is feasible from PC analysis; in particular, the role different meteorological components play could be quantified with more rigor. This could be complementary to studies such as Widlanksy et al. (2011).
Applying this methodology to tropical plumes in general may require a different choice for the flagging variable. Plumes over the North Atlantic and North Pacific often have only mid- to high-level cloud in the tropics with deep cloud and heavy rainfall found near their termination points over land (Knippertz 2007). Hart et al. (2010) suggested that TTTs fit the theoretical framework for tropical plumes but highlighted this key difference. The lack of deep cloud along much of the axis would modify the necessary OLR thresholds; indeed preliminary work (not shown) attempting to apply this metbot in the North Pacific and Atlantic revealed this problem. Jiang and Deng (2011) applied an adaptive zonal and meridional thresholding technique to column integrated water vapor to identify atmospheric river features in the National Aeronautics and Space Administration (NASA) Modern-Era Retrospective Analysis for Research and Applications (MERRA) dataset. Thus, robust methods to flag atmospheric rivers exist. However, they are unable to provide the level of meteorological detail that case studies inherently do (e.g., Ralph et al. 2011). A proposal is that a similar metbot, based on the Jiang and Deng (2011) methodology to flag systems, could go some way to bridge the gap to case studies and provide a powerful tool to evaluate atmospheric river dynamics in a multiscale framework and across many datasets.
Herein lies a key point; while the case for (semi) objective identification methods is made regularly, it is very difficult to achieve the level of detail of a case study since that is not their goal. However, these details are important when attempting to develop a mechanistic understanding of multiscale interactions.
The metbot methodology, while much cruder than many of the feature-tracking methodologies referenced here, has the potential to be the tool to do this for the subtropics over southern Africa. The adaptability of the metbot is demonstrated in Hart et al. (2012, manuscript submitted to Climate Dyn.), where midtropospheric cutoff low tracks produced by Favre et al. (2012) are associated with cloud band events. Coding a relation building tool such as this metbot, around the many, well-established, feature-tracking methods, is likely to help bridge this object–case study gap in other regions that are dominated by different weather systems.
To conclude, the three criteria for a successful metbot, as outlined in the introduction, have been met. Blob detection provided an effective region-of-interest locator and applying two simple criteria to blob properties enabled cloud band flagging. A simple set of blob association rules were used to track blobs and build a data object that contained a detailed meteorology of each cloud band event. Some examples of the use of this set of event objects have been demonstrated and the authors conclude that the methodology captures essential features of the evolution and rainfall of TTTs over southern Africa.
Acknowledgments
This work was only possible thanks to developers of opencv (http://opencv.willowgarage.com) and elsewhere and the developer of cvblob whose code is made freely available, with Python wrappers, as part of the Robotics Operating System project. Interpolated OLR and NCEP2 data were obtained from NOAA/OAR/ESRL PSD, Boulder, Colorado (http://www.cdc.noaa.gov). The first author gratefully acknowledges funding through SANAP and a D & E Potter Foundation Ph.D. Fellowship. Juliana Dias and an anonymous reviewer are thanked for their helpful suggestions.
APPENDIX
Software Details
The software platform underlying this work is as follows:
Operating System: Linux
Software Libraries: OpenCV (http://opencv.willowgarage.com/), cvblob (http://code.google.com/p/cvblob/)
Software platform: Python
Python Modules: NumPy, ScientificPython (Hinsen 2007), Matplotlib, Basemap, PyClimate
PIL, python wrappers to cvblob and OpenCV
REFERENCES
Bain, C. L., J. De Paz, J. Kramer, G. Magnusdottir, P. Smyth, H. Stern, and C.-C. Wang, 2011: Detecting the ITCZ in instantaneous satellite data using spatiotemporal statistical modeling: ITCZ climatology in the east Pacific. J. Climate, 24, 216–230.
Benestad, R. E., and D. Chen, 2006: The use of a calculus-based cyclone identification method for generating storm statistics. Tellus, 58A, 473–486.
Blamey, R. C., and C. J. C. Reason, 2012: Mesoscale convective complexes over southern Africa. J. Climate, 25, 753–766.
Bradski, G., cited 2000: The OpenCV Library. Dr. Dobb’s: The World of Software Development. [Available online at http://www.drdobbs.com/open-source/the-opencv-library/184404319.]
Carvalho, L. M. V., and C. Jones, 2001: A satellite method to identify structural properties of mesoscale convection systems based on the maximum spatial correlation technique (MASCOTTE). J. Appl. Meteor., 40, 1683–1701.
Chang, F., C.-J. Chen, and C.-J. Lu, 2004: A linear-time component-labeling algorithm using contour tracing technique. Comput. Vis. Image Underst., 93 (2), 206–220.
Cook, C., C. J. C. Reason, and B. C. Hewitson, 2004: Wet and dry spells within particularly wet and dry summers in the South African summer rainfall region. Climate Res., 26, 17–31.
Cook, K. H., 2000: The South Indian convergence zone and interannual rainfall variability over southern Africa. J. Climate, 13, 3789–3804.
Crétat, J., Y. Richard, B. Pohl, M. Rouault, C. J. C. Reason, and N. Fauchereau, 2010: Recurrent daily rainfall patterns over South Africa and associated dynamics during the core of the austral summer. Int. J. Climatol., 32, 261–273, doi:10.1002/joc.2266.
Dias, J., S. N. Tulich, and G. N. Kiladis, 2012: An object-based approach to assessing the organization of tropical convection. J. Atmos. Sci.,69, 2488–2504.
Fauchereau, N., B. Pohl, C. Reason, M. Rouault, and Y. Richard, 2009: Recurrent daily OLR patterns in the Southern Africa/Southwest Indian Ocean region: Implications for South African rainfall and teleconnections. Climate Dyn., 32, 575–591.
Favre, A., B. Hewitson, M. Tadross, C. Lennard, and R. Cerezo-Mota, 2012: Relationships between cut-off lows and the semi-annual and southern oscillations. Climate Dyn., 38, 1473–1487, doi:10.1007/s00382-011-1030-4.
Harangozo, S. A., and M. S. J. Harrison, 1983: On the use of synoptic data in indicating the presence of cloud bands over southern Africa. S. Afr. J. Sci., 79 (10), 413–414.
Harrison, M. S. J., 1984: A generalized classification of South African rain-bearing synoptic systems. J. Climatol., 4, 547–560.
Hart, N., C. J. C. Reason, and N. Fauchereau, 2010: Tropical–extratropical interactions over southern Africa: Three cases of heavy summer season rainfall. Mon. Wea. Rev., 138, 2608–2623.
Hart, N., C. J. C. Reason, and N. Fauchereau, 2012: Cloud bands over southern Africa: Seasonality, contribution to rainfall variability, and modulation by the MJO. Climate Dyn., in press.
Hewson, T., and H. A. Titley, 2010: Objective identification, typing and tracking of the complete life-cycles of cyclonic features at high spatial resolution. Meteor. Appl., 17, 355–381.
Hinsen, K., cited 2007: ScientificPython. [Available online at http://dirac.cnrs-orleans.fr/ScientificPython/.]
Hodges, K. I., 1994: A general method for tracking analysis and its application to meteorological data. Mon. Wea. Rev., 122, 2573–2586.
Hodges, K. I., and C. D. Thorncroft, 1997: Distribution and statistics of African mesoscale convective weather systems based on ISCCP Meteosat imagery. Mon. Wea. Rev., 125, 2821–2837.
Hodges, K. I., D. W. Chappell, G. J. Robinson, and G. Yang, 2000: An improved algorithm for generating global window brightness temperatures from multiple satellite infrared imagery. J. Atmos. Oceanic Technol., 17, 1296–1312.
Hodges, K. I., R. W. Lee, and L. Bengtsson, 2011: A comparison of extratropical cyclones in recent reanalyses ERA-Interim, NASA MERRA, NCEP CFSR, AND JRA-25. J. Climate, 24, 4888–4906.
Holton, J. R., 2004: An Introduction to Dynamic Meteorology. 4th ed. International Geophysical Series, Vol. 88, Academic Press, 535 pp.
Hunter, J. D., 2007: Matplotlib: A 2D graphics environment. Comput. Sci. Eng., 9 (3), 90–95.
Jiang, T., and Y. Deng, 2011: Downstream modulation of North Pacific atmospheric river activity by East Asian cold surges. Geophys. Res. Lett.,38, L20807, doi:10.1029/2011GL049462.
Kanamitsu, M., W. Ebisuzaki, J. Woollen, S. K. Yang, J. Hnilo, M. Fiorino, and G. L. Potter, 2002: NCEP-DOE AMIP II reanalysis (R-2). Bull. Amer. Meteor. Soc., 83, 1631–1643.
Knippertz, P., 2007: Tropical-extratropical interactions related to upper-level troughs at low latitudes. Dyn. Atmos. Oceans, 43, 36–62.
Liebman, B., and C. A. Smith, 1996: Description of complete (interpolated) outgoing longwave radiation dataset. Bull. Amer. Meteor. Soc., 77, 1275–1277.
Liñán, C. C., cited 2008: cvblob. [Available online at http://cvblob.googlecode.com.]
Lynch, S., 2003: Development of a RASTER database of annual, monthly and daily rainfall for Southern Africa. Rep. 1156/1/03, WRC, 78 pp.
Matthews, A. J., 2012: A multiscale framework for the origin and variability of the South Pacific convergence zone. Quart. J. Roy. Meteor. Soc., 138, 1165–1178, doi:10.1002/qj.1870.
McGuirk, J. P., and D. J. Ulsh, 1990: Evolution of tropical plumes in VAS water vapor imagery. Mon. Wea. Rev., 118, 1758–1766.
Murray, R. J., and I. Simmonds, 1991a: A numerical scheme for tracking cyclone centres from digital data. Part 1: Development and operation of the scheme. Aust. Meteor. Mag., 39, 155–166.
Murray, R. J., and I. Simmonds, 1991b: A numerical scheme for tracking cyclone centres from digital data. Part 2: Application to January and July general circulation simulations. Aust. Meteor. Mag., 39, 167–180.
Oliphant, T. E., cited 2006: Guide to NumPy. [Available online at http://www.tramy.us/.]
Ralph, F. M., P. J. Neiman, G. N. Kiladis, and K. Weickmann, 2011: A multiscale observational case study of a Pacific atmospheric river exhibiting tropical–extratropical connection and a mesoscale frontal wave. Mon. Wea. Rev., 139, 1169–1189.
Reason, C. J. C., and D. Jagadheesha, 2005: A model investigation of recent ENSO impacts over southern Africa. Meteor. Atmos. Phys., 89, 181–205.
Reason, C. J. C., W. Landman, and W. Tennant, 2006: Seasonal to decadal prediction of southern African climate and its links with variability of the Atlantic Ocean. Bull. Amer. Meteor. Soc., 87, 941–955.
Rouault, M., P. Florenchie, N. Fauchereau, and C. J. C. Reason, 2003: South East tropical Atlantic warm events and southern African rainfall. Geophys. Res. Lett.,30, 8009, doi:10.1029/2002GL014840.
Sáenz, J., J. Zubillaga, and J. Fernández, 2002: Geophysical data analysis using python. Comput. Geosci., 28 (4), 457–465.
Scharenbroich, L., G. Magnusdottir, P. Smyth, H. Stern, and C.-C. Wang, 2010: A Bayesian framework for storm tracking using a hidden-state representation. Mon. Wea. Rev., 138, 2132–2148.
Skok, G., J. Tribbia, J. Rakovec, and B. Brown, 2009: Object-based analysis of satellite-derived precipitation systems over the low- and midlatitude Pacific Ocean. Mon. Wea. Rev., 137, 3196–3128.
Streten, N., 1973: Some characteristics of satellite-observed bands of persistent cloudiness over the Southern Hemisphere. Mon. Wea. Rev., 101, 486–495.
Todd, M., and R. Washington, 1999: Circulation anomalies associated with tropical-temperate troughs in Southern Africa and the south west Indian Ocean. Climate Dyn., 15, 937–951.
Washington, R., and M. Todd, 1999: Tropical-temperate links in Southern Africa and Southwest Indian Ocean satellite-derived daily rainfall. Int. J. Climatol., 19, 1601–1616.
Whitaker, J. S., cited 2007: Matplotlib Basemap Toolkit documentation. [Available online at http://matplotlib.github.com/basemap/.]
Widlanksy, M. J., P. J. Webster, and C. D. Hoyos, 2011: On the location and orientation of the South Pacific Convergence Zone. Climate Dyn., 36, 561–578.
