NESDIS Atmospheric Motion Vector (AMV) Nested Tracking Algorithm: Exploring its Performance Jaime Daniels 1, Wayne Bresky 2, Steven Wanzong 3, Andrew Bailey 2, and Chris Velden 3 1 NOAA/NESDIS Office of Research and Applications, NOAA Science Center, Camp Springs, Maryland 20746, U.S.A. 2 I.M. Systems Group (IMSG), Rockville, Maryland 20852, U.S.A. 3 Cooperative Institute for Meteorological Satellite Studies (CIMSS), Space Science and Engineering Center (SSEC), University of Wisconsin Madison, Wisconsin 53706, U.S.A ABSTRACT A new Atmospheric Motion Vector (AMV) nested tracking algorithm has been developed for the Advanced Baseline Imager (ABI) to be flown on NOAA s future GOES-R satellite. The algorithm has been designed to capture the dominant motion in each target scene from a family of local motion vectors derived for each target scene. Capturing this dominant motion is achieved through use of a two-dimensional clustering algorithm that segregates local displacements into clusters. The dominant motion is taken to be the average of the local displacements of points belonging to the largest cluster. This approach prevents excessive averaging of motion that may be occurring at multiple levels or at different scales that can lead to a slow speed bias and a poor quality AMV. A representative height is assigned to the dominant motion vector through exclusive use of cloud heights from pixels belonging to the largest cluster. This algorithm has been demonstrated to significantly improve the slow speed bias typically observed in AMVs derived from satellite imagery. GOES-N/O/P, Meteosat SEVERI, and NPP/VIIRS imagery are serving as GOES-R ABI proxy data sources for the continued development, testing, and validation of the GOES-R AMV algorithms. This talk will focus on the performance of the nested tracking algorithm as supported by comparisons to a variety of reference/ground truth wind observations and case study analyses. This talk will also touch briefly on the performance and impact of the nested tracking AMVs within the National Centers for Environmental Prediction (NCEP) Global Forecast System (GFS). Finally, this talk will highlight some of the enhancements that have been made to the algorithm and discuss areas of future work. INTRODUCTION A new Atmospheric Motion Vector (AMV) nested tracking algorithm (Bresky et al, 2012, Daniels et al, 2010) has been developed for the Advanced Baseline Imager (ABI) to be flown on NOAA s future GOES-R satellite. The GOES-R satellite is scheduled to be launched sometime in the first quarter of 2016. In preparation for the launch of GOES-R, NOAA/NESDIS has applied the new GOES-R nested tracking algorithm to a variety of different sensors (GOES-N/O/P, Meteosat/SEVIRI, and NPP/VIIRS) that serve as GOES-R ABI proxy data sources. Doing this enables the continued development, testing, and validation of winds generated by this algorithm and provides important insight into the overall performance of the nested racking algorithm. Furthermore, the impact of proxy GOES-R winds generated by the nested tracking algorithm on NCEP s global forecasts can be assessed (Nebuda et al, 2014). In this paper we provide a brief review of the nested tracking algorithm, describe some of the new nested tracking output potentially relevant to NWP data assimilation, and show numerous examples of winds generated when the algorithm is applied to various ABI proxy imagery. The performance of winds generated by the nested tracking algorithm is determined via comparisons to radiosonde winds, winds generated by NOAA s heritage winds algorithm (Hayden C. and S. Nieman, 1996; Nieman et al, 1997), and EUMETSAT s operational winds algorithm (Borde et al, 2014). Additional 1
findings resulting from experiments involving changes to elements of the nested tracking algorithm are also provided. DESCRIPTION OF THE NESTED TRACKING ALGORITHM The new nested tracking algorithm (Bresky et al, 2012) was designed specifically to minimize the often observed slow speed bias of satellite winds which has been a long standing concern of the NWP community (Bormann et al, 2002). The nested tracking approach utilizes a small tracking window to minimize averaging and produce faster wind estimates, but does so without introducing the random tracking errors associated with a small window. The approach is able to identify the pixels having the greatest influence on the motion estimate. Using this information, together with the apriori knowledge of cloud height at each pixel (Heidinger, 2014; Heidinger, 2010; Heidinger and Pavolonis, 2009) enables the establishment of a direct link between the observed motion and the height of the tracer. The new tracking approach involves deriving a motion estimate for all possible 5x5 pixel subtarget regions nested within the larger target window. Small 5x5 sub-targets are better suited than larger target windows for tracking the small-scale motion representative of the instantaneous wind. The new approach produces a field of local motion vectors associated with each target window. One example is shown in Figure 1. Smaller 5x5 box nested within a larger 19x19 target scene 19x19 Target Scene Figure 1. An example of the local motion field derived with nested tracking. The white vectors show the local motion derived with a 5x5 box centered on the pixel location. The average local motion vector is shown in green and the vector derived by tracking the entire scene is shown in red. Note that local motion vectors are not generated near the boundary where a full 5x5 box does not exist. The DBSCAN cluster analysis (Ester et al., 1996) algorithm is applied to the field of local motion vector displacements in order to extract coherent motion clusters within the target window. The dominant motion in the target scene is determined from the largest motion cluster. The presence of smaller, coherent motion clusters indicate motions that differ from the dominant motion either in scale and/or because it comes from a different level in the atmosphere. Sometimes a smaller coherent motion cluster represents some motion that has nothing to do with the instantaneous wind (ie., the movement of a frontal boundary). Some of the nested tracking algorithm diagnostic output is new and may lend itself useful for quality controlling nested tracking winds in the Numerical Weather Prediction (NWP) data assimilation process (Nebuda et al, 2014). Table 1 includes a list of some of the diagnostics currently outputted by the nested tracking algorithm. 2
Table 1: Current Diagnostic Output from the Nested Tracking Algorithm Standard deviation of displacements (in pixels) in largest cluster Standard deviation of displacements divided by the magnitude of the average displacement Size of the largest cluster Median, Minimum, and Maximum cloud-top pressure (hpa) in the largest cluster Median, Minimum, and Maximum cloud-top temperature (K) in the largest cluster Median, Minimum, and Maximum cloud optical depth in the largest cluster Dominant cloud phase within the target scene Dominant cloud type within the target scene Standard deviation of cloud top pressure values (hpa) within the target scene APPLICATION TO AVAILABLE GOES-R ABI PROXY DATA The nested tracking algorithm is being routinely applied to a variety of different sensors (GOES- 13/15, Meteosat-10/SEVIRI, and Soumi-NPP/VIIRS) that all serve as GOES-R ABI proxy data sources. Use of these proxy data sources provides us with a key opportunity to explore the performance of the nested tracking algorithm and to make and test improvements to it before the launch of GOES-R. The performance of the algorithm is discussed in the following section. Examples of winds generated from the proxy data are shown in Figures 2 and 3. More details related to the application of the nested tracking algorithm on Soumi-NPP/VIIRS are described in Key et al, 2014. Figure 2. Nested tracking winds derived from Meteosat-10/SEVIRI FD from 0.60um, 3.9um, 6.2um, and 10.8um imagery. 3
Figure 3. Nested tracking winds derived from GOES15 10.7um imagery (upper left), GOES-13 10.7um imagery (upper right), simulated GOES-R ABI 11.2um imagery (lower left) and Suomi-NPP/VIIRS 10.7um imagery (lower right). PERFORMANCE OF NESTED TRACKING AMVs To assess the performance of the nested tracking algorithm, retrieved AMVs were compared to radiosonde winds. Table 2 shows the nested tracking Meteosat-10/SEVIRI AMV/rawinsonde collocation statistics and NCEP GFS forecast winds/rawinsonde collocation statistics sorted by pressure layer (high, mid, low) and Quality Indicator (QI > 80) for the period March 20 May 31, 2014. The AMVs were generated over the entire full disk domain using successive images separated by a 15 minute time interval. The comparison metrics indicate that the quality of satellite AMVs closely mirrors the quality of the short-term (6-12 hr) NCEP GFS forecast winds. This is especially true at the upper and lower levels of the troposphere. The magnitudes of the speed bias associated with the satellite AMVs are generally quite good at each of the layers. This is an encouraging result. To get a better sense of the wind speed differences between the AMVs and collocated rawinsondes, histograms of wind speed differences were constructed. Figure 4 shows these histograms for mid and upper layers of the atmosphere. Both histograms show well defined Gaussian distributions that are not very broad and centered close to zero. High Level (100 400 hpa) Mean Vector Difference (m/s) Standard Deviation (m/s) Speed Bias (m/s) Abs. Directional Difference (Deg) Mean Speed (m/s) Sample Size Mid Level (400 700 hpa) Mean Vector Difference (m/s) Standard Deviation (m/s) Satellite AMV 4.84 3.49-0.72 12.67 15.39 23400 NCEP GFS Forecast 4.76 3.05-0.71 15.29 15.41 23400 4.62 3.39 3.78 2.73 4
Speed Bias (m/s) -0.19-0.52 Abs. Directional Difference (Deg) 13.40 12.29 Mean Speed (m/s) 13.12 12.79 Sample Size 10573 10573 Low Level (> 700 hpa) Mean Vector Difference (m/s) 2.94 2.71 Standard Deviation (m/s) 1.93 1.76 Speed Bias (m/s) -0.04 0.01 Abs. Directional Difference (Deg) 13.71 13.00 Mean Speed (m/s) 8.63 8.68 Sample Size 9349 9349 Table 2. Nested tracking Meteosat-10/SEVIRI AMV/rawinsonde collocation statistics (column 2) and NCEP GFS/rawinsonde collocation statistics (column 3) for March 20 June 2, 2014. Figure 4. Histograms of speed bias differences (AMV rawinsondes) for Meteosat-10/SEVIRI (10.8um) nested tracking AMVs in the upper (left) and mid (right) troposphere. In order to get a sense of the quality of the AMV height assignments, a Level-of-Best Fit assessment of the AMVs was performed using rawinsonde vertical wind profiles. Figure 5 shows the level of bestfit profiles of RMSE and absolute speed bias at 200 mb, 300mb, 500mb, and 700mb for nested tracking winds derived from Meteosat-10/SEVIRI 10.8um imagery for the period March 22, 2014 June 2, 2014. In the ideal case, the level of best fit (ie. as indicated by a minimum in RMSE or speed bias curve) would coincide exactly with the pressure level assigned to the AMVs. Inspection of Figure 5 indicates that minimums in the level-of-best-fit (RMSE and Bias) curves match the AMV heights exactly at all, but one of the levels. At 200mb, the minimums in the RMSE and speed bias profile curves occur about 25mb below the assigned AMV pressure level. This result suggests that the heights associated with these AMVs are slightly too high up in the atmosphere. 5
Figure 5. Level of best-fit profiles of RMSE (green) and absolute speed bias (blue) at 200 mb, 300mb, 500mb, and 700mb nested tracking winds derived from Meteosat-10/SEVIRI 10.8um imagery for the period March 22, 2014 June 2, 2014. While not discussed or shown in this paper, there is a considerable amount of work (Nebuda et al, 2014) being done to evaluate the quality and impact of the nested tracking AMVs in NCEP s Global Forecast System. A short study was done to provide information about the overall quality of the nested tracking AMVs relative to the overall quality of the AMVs derived from the heritage algorithm used operationally at NESDIS today. In this study, we generated AMVs from the new nested tracking algorithm and from the NESDIS heritage winds algorithm using Meteosat-8/SEVIRI 10.8um imagery over two different months (August 2006 and February 2007). We created a homogeneous collocation dataset between the these two sets of AMVs and radiosonde winds and generated comparison statistics at 100mb layers from near the surface to the top of the troposphere. Figure 6 shows the vertical profiles of these comparison statistics (RMSE and speed bias). Immediately evident is the reduction in the speed bias throughout the entire profile. The only exception is in the upper most layer where the magnitude of the speed bias is somewhat less for the AMVs derived from the heritage winds algorithm. It is very likely that this is a result of the application of a speed bias correction by the heritage winds algorithm. The RMSE curves for the nested tracking AMVs and the heritage AMVs are comparable with the nested tracking AMVs showing lower RMSEs below 700mb and the heritage AMVs showing lower RMSE at and around 400mb. It is noted that we are working to better understand and hopefully resolve the undesirable characteristic of the speed bias profile associated with the nested tracking AMVs between 200-450 mb. We suspect this is a height assignment related issue and are working closely with the GOES-R cloud application team to address and correct this. 6
Figure 6. RMSE and speed bias profiles (AMVs radiosondes) for AMVs generated from Meteosat-8 10.8um imagery for August 2006 and February 2007 using the new nested tracking algorithm (blue curves) and the heritage NESDIS AMV algorithm used operational today at NESDIS (red curves) Collaborative work between NESDIS and EUMETSAT in 2013 was done to learn and share information about each other s AMV retrieval algorithm. As part of this collaboration, a study was performed to assess the quality of nested tracking AMVs relative to the quality of AMVs derived from EUMETSAT s Cross Correlation Contribution (CCC) method. To do this the CCC algorithm was integrated in the NESDIS nested tracking software. Both the nested tracking AMVs and the CCC AMVs used the same cloud heights to assign heights to the AMVs. AMVs were derived from each algorithm from Meteosat-8 10.8um imagery for February 2007. A homogeneous collocated dataset between these two sets of AMVs and radiosonde winds was then constructed. From this dataset, we generated comparison statistics at 100mb layers from near the surface to the top of the troposphere. Figure 7 shows the vertical profiles of these comparison statistics (RMSE and speed bias). For this one month period, it can be seen that the magnitudes of the RMSEs and speed biases for the nested tracking AMVs are lower throughout the troposphere. The results are far from definitive as the study was done for only one month. Continuing collaboration and work is planned that will enable additional comparisons and testing to be done to improve both algorithms. 7
Figure 7. RMSE and speed bias profiles (AMVs radiosondes) for AMVs generated from Meteosat-8 10.8um imagery for August 2006 and February 2007 using the new nested tracking algorithm (blue curves) and the EUMETSAT CCC AMV algorithm (red curves). SUMMARY OF ONGOING ACTIVITIES AND FUTURE PLANS The new nested tracking algorithm developed at NOAA/NESDIS for the future GOES-R ABI was designed to minimize the slow speed bias commonly observed with satellite-derived AMVs at mid and upper levels of the atmosphere. The algorithm continues to being routinely applied to several GOES-R ABI proxy data sources that include GOES-13/15, Meteosat-10/SEVIRI, Soumi- NPP/VIIRS, and simulated GOES-R ABI. The GOES-R winds application team has taken advantage of these data sources to develop and test further enhancements to the nested tracking algorithm. AMV validation results to date indicate that the algorithm is performing very well. As hoped, marked reductions are observed in the magnitude of the wind speed bias and RMSE throughout the troposphere. These reductions represent an improvement over AMVs generated by the heritage AMV algorithm used in NESDIS operations today. As a result, not only will the nested tracking algorithm be implemented for GOES-R, but it will also replace the heritage AMV algorithm being used operationally for GOES-E, GOES-W, Soumi-NPP/VIIRS, NOAA/AVHRR, Metop/AVHRR, Aqua/MODS, and Terra/MODIS today at NESDIS. A considerable amount of work is being done to evaluate nested tracking AMVs in NCEP s Global Forecast System. NESDIS plans to make these AMVs available to International NWP centers so they too can assess the quality the nested tracking winds and their impact on forecast quality. REFERENCES Borde, R., M. Doutriaux-Boucher, G. Dew, and M. Carranza, 2014: A Direct Link between Feature Tracking and Height Assignment of Operational EUMETSAT Atmospheric Motion Vectors, Journal of Atmospheric and Oceanic Technology. Bormann, N., G. Kelly, J.-N. Thépaut, 2002: Characterising and correcting speed biases in Atmospheric Motion Vectors within the ECMWF system. Proc. Sixth Intl. Winds Workshop, Madison, WI, EUMETSAT, 113-120. 8
Bresky, W., J. Daniels, A. Bailey, and S. Wanzong, 2012: New Methods Towards Minimizing the Slow Speed Bias Associated With Atmospheric Motion Vectors (AMVs). J. Appl. Meteor. Climatol., 51, 2137-2151 Daniels, J., 2010: GOES-R Advanced Baseline Imager (ABI) Algorithm Theoretical Basis Document For Derived Motion Winds, GOES-R Program Office, 96 pp, www.goes-r.gov. Ester, M., H.-P. Kriegel, J. Sander and X. Xu, 1996: A Density-Based Algorithm for Discovering Clusters in Large Spatial Databases with Noise. Proc. Second Intl. Conf.on Knowledge Discovery and Data Mining (KDD-96), Portland, OR, 226-231. Hayden, C.M. and S.J. Nieman, 1996: A primer for tuning the automated quality control system and for verifying satellite-measured drift winds. NOAA Tech. Memo. NESDIS 43, 27 pp. Heidinger, A. K., 2014: NOAA GOES-R AWG Cloud Height Algorithm (ACHA). Proc. Twelfth Intl. Winds Workshop, Copenhagen, Denmark, EUMETSAT. Heidinger, A., 2010: GOES-R Advanced Baseline Imager (ABI) Algorithm Theoretical Basis Document For Cloud Height, GOES-R Program Office, 77 pp, www.goes-r.gov. Heidinger, A. K. and Pavolonis, M. J., 2009: Gazing at cirrus clouds for 25 years through a split window, Part 1: Methodology, J. Appl. Meteorol. Clim., 48, 1100 1116, 2009. Key J., J. Daniels, S. Wanzong, A. Bailey, H. Qi, W. Bresky, D. Santek, C.Velden, and W. Wolf, 2014: VIIRS Polar Winds Status and Use. Proc. Twelfth Intl. Winds Workshop, Copenhagen, Denmark, EUMETSAT. Nubuda, S., J. Jung, D. Santek, J. Daniels, and W. Bresky, 2014: Assimilation of GOES-R Atmospheric Motion Vectors in the NCEP Global Forecast System. Proc. Twelfth Intl. Winds Workshop, Copenhagen, Denmark, EUMETSAT. Nieman, S.J., W.P. Menzel, C.M. Hayden, D. Gray, S.T. Wanzong, C.S. Velden, and J. Daniels, 1997: Fully automated cloud-drift winds in NESDIS operations. Bull. Amer. Meteor. Soc., 78, 1121-1133. 9