SSL quarterly progress report for quarter ending 31Mar 2016

Submitted 26 Apr 2016

3rd Quarterly Progress Report for CA NOO244-15-2-0005, Intraseasonal Tropical Cyclone Forecasting

  1. Overview

The progress made by Statistical Solutions LLC (SSL) as of the end of the third quarter (3/31/16) on research as delineated in the cooperative agreement (CA) N00244-15-2-0005 is evaluated as on schedule and on budget. At the end of the quarter, 39.5% of the total funding has been spent over 39.5% of the lifetime of the CA. Further details on funding and expenditures are available in SSL’s Interim Financial Report (SF-425). For the sake of both transparency and for achieving the widest possible dissemination of the public benefit resulting from this research, this progress report will be made publically available and posted on SSL’s website. Due to the public release of this form, the total government funding of the cooperative agreement is omitted, though all financial details will be found in the SF-425, which already has been sent to Dr. Tom Murphree of the Naval Postgraduate School (NPS), the Government Sponsor, and the Administrative Grants Office.

When comparing progress with the schedule and milestones agreed upon in the CA or as adjusted in discussion with Dr. Tom Murphree of NPS, progress made to the end of the quarter would indicate that the project is on schedule as all preparations necessary to begin a North Atlantic (NA) and eastern North Pacific ENP hurricane /tropical cyclone (TC) forecasting system model have been completed.

  1. North Atlantic Hurricane and Eastern North Pacific TC Forecasting

A. Data Mining

Upon the CA becoming effective, SSL, in consultation with Dr. Murphree of NPS, began laying the groundwork for researching North Atlantic (NA) hurricane forecasting, and more specifically mining the data necessary to determine if a statistical relationship between values of NA large scale environmental factors (LSEFs) and hurricane formation could be established To that end, SSL completed archiving the National Center for Environmental Prediction’s Climate Forecast System Reanalysis (CFSR) oceanic and atmospheric data from January 1st, 1979 through December of 2014 as reported in the second quarterly report.

The HURDAT2 data set, downloaded from the National Hurricane Center (updated 2014), has also been archived by SSL.

  1. Data, Data Handling, Calculation, and Process Validation

As reported in the previous quarterly report, SSL performed extensive data verification on the mined CFSR data sets. Comparing data for specific dates and locations with independent data from other sources was performed to check the integrity of the data that was archived and processed. Since then, in consultation with Dr. Murphree of NPS, SSL used NA CFSR LSEF data to force our pre-existing western North Pacific (WNP) statistical model (that used for research to include the experimental forecasting of WNP TC formations). Our assumption was that the WNP statistical model, while not trained (hence not ideal) for the NA, might possibly indicate on at least a rough scale, where the NA might be favorable for hurricane formation and where it might not be. Comparing the favorability of formation with actual HURDAT2 formation locations and times would also provide a level of validation that the integrity of the LSEF and derived data, as well as the overall forecasting process was intact following the archiving, handling, and calculation evolution. Furthermore, an alignment of CFSR LSEF forced statistical model favorability with actual formations would lend credence to the hypothesis that the LSEFs influence TC/Hurricane formation regardless of basin, perhaps identically so (ie the WNP model could be used without modification for formation forecasting in the NA or elsewhere).

  1. CFSR Forcing of WNP Statistical Model

In Figure 1 we observe that there are three (+) separate regions of elevated probability created in the NA by forcing the WNP statistical model with CFSR created values of the LSEFs at a daily, half degree resolution. Corresponding to the regions of highest formation probability are the actual HURDAT2 formation locations for hurricanes Humberto, Iris and Jerry, all of which formed on 22 Aug 1995, the date that the hindcast of Figure 1 is valid. Other case studies were performed, with similar results – the WNP model, forced with CFSR generated values of the LSEFs was quite skilled in indicating when and where NA hurricanes would form.

Fig 1. Hurricane formation probabilities generated by our WNP statistical model forced with CFSR LSEF data. The formation locations of hurricanes Humberto, Iris and Jerry are shown. Note the tight agreement between formation probability and formation.

In addition to the positive results observed in the NA, we also observed positive results in the ENP. Note in Figure 1, a rather sizeable portion of the ENP is shown. In Figure 1, the CFSR LSEFs are zeroed out, and the SSL and NPS logos are inserted so as to not distract NA users with ENP activity, but in the original raw figures, all oceanic regions had CFSR generated values assigned to them at a half degree, daily resolution, from which WNP statistical model formation probabilities could be calculated. Thus, we observed that there were frequent occurrences of elevated formation probabilities, and that these elevated probabilities frequently matched up with actual TC formations according to the ENP HURDAT2 database. Figure 2 is an example of such an ENP CFSR forced nowcast of TC formation (Blanca).

Fig 2. ENP TC formation probabilities generated by our WNP statistical model forced with CFSR LSEF data. The formation location of Blanca is shown. Note the tight agreement between formation probability and formation just as we observed in Figure 1 with the NA.

This development led to discussions between Dr. Murphree and SSL as to whether pursuit of an ENP TC forecasting system (not an original part of the cooperative agreement SOW) was a worthwhile investment of time and resources, and what SOW research the ENP forecasting research would replace. Ultimately, what was decided was that ENP TC forecasting research held so much promise for benefit (meaning SSL/NPS would be engaged in experimental forecasting of the vast majority of Northern Hemisphere hurricanes/TCs and that the scientific community’s understanding of how TCs form may be strengthened as we compare/contrast the differences and similarities in the LSEFs at the point of formation for each of the three basins we would now cover) that we decided to pursue formation forecasting for the ENP. In order to initiate and maintain an ENP research effort without an increase in resources, something had to be put on the back seat, and that item was exploring relationships between TC formation and the MJO. SSL hopes to revisit this research at a later date

There is one final item of note from Figure 1 and to a lesser extent Figure 2. It is easy to overlook the regions of low, yet elevated probabilities covering much of the Gulf of Mexico and Western Caribbean, given the high probabilities at formation locations, but as will be shown shortly, these large regions of elevated formation probability are indicative of WNP model shortcomings when used in the NA and ENP.

  1. CFSv2 Forcing of WNP Statistical Model

Forcing the statistical model with CFSR data to produce 0-lead hindcasts is valuable from the standpoint of understanding the science of tropical cyclogenesis. However, because coastal dwellers, mariners, offshore oil rig personnel, GOs/NGOs, etc., all benefit from early warning, we also want to be able to forecast TC/hurricane formations. If forcing a statistical model with CFSR generated values of the LSEFs yields skilled nowcasts of formation, then forcing the statistical model with skilled forecasts of the LSEFs (as we do with our ongoing WNP TC forecasting research) should skillfully forecast TC/hurricane formations.

We began our exploration of NA hurricane forecasting by using 1-day lead CFSv2 forecasts of the LSEFs. Our experience from the WNP has shown that the 1-day lead forecasts are highly skilled, and the easiest to interpret, and thus were the logical choice for our initial exploration of NA hurricane forecasting.

Figure 3 is a 1-day lead hindcast valid for October 20th, 2012, the day hurricanes Sandy ( the hurricane located near 80° W ) and Tony formed. Two items are apparent from Figure 3: the first is that again we see that the contoured area in which Sandy formed (a contoured region is a region within which we believe TC/hurricane formation is possible) is unusually large for a short lead forecast. Secondly, while the contoured region within which Sandy formed is large, Sandy formed near the highest probability. In addition, Tony formed in the middle of a quite small region of elevated probabilities. The two formations match up well enough with the synoptic scale activity of the forecast that we again have at least modest validation that the data is calculated and plotted correctly.

Fig 3. NA hurricane formation probabilities generated by our WNP statistical model forced with CFSv2 LSEF data at a 1-day lead. The formation locations of Sandy (near 80° W) and Tony are shown. Note the good synoptic scale agreement between formation probability and formation.

Figure 4 is a current month monthly outlook, meaning it has a 1-month long valid period and was issued a day or two before the valid period started. This forecast correctly identifies the region where 4 out of five hurricanes form, where the lone miss is not off by much. Perhaps more importantly is the detail offered by this hindcast (spatial and temporal resolution) that is not available with seasonal forecasts, or at leads beyond what dynamical models are capable of producing with skill. Most importantly, though, is what Figure 4 tells us in comparison with Figure 5. The contoured region shown is created by use of forecasted LSEF data which is processed and used to force our WNP statistical model, creating (on a daily, 1° grid) the probability of hurricane formation for the given values of the LSEFs .

Fig 4. Current month monthly outlook of NA hurricane formation probabilities generated by our WNP statistical model forced with CFSv2 LSEF data. The formation locations of Oscar, Patty, Rafael, Sandy (near 80° W) and Tony are shown.

Figure 5 is a plot of all NA hurricane formations (Jun-Nov, 1979-2015). This plot indicates that there isn’t much of the tropical NA where hurricanes will not form. Yet Figure 4 correctly forecasts little activity in the Gulf of Mexico, a region by climatology is very active. It also correctly forecasts that the bulk of the cyclogenesis activity would occur West of roughly 40° W, again in contrast to what one might think about potential formation locations after looking at climatology.

Fig 5. NA hurricane formations, June-November, 1979-2015. Locations from the HURDAT2 data-base.

Figure 6 is a 2-month lead monthly outlook, meaning it has a 1 month long valid period and was issued two months before the valid period starts. This forecast, similar to the forecast of Figure 4, correctly identifies the region where 4 out of five hurricanes form. Likewise, in addition to the spatial and temporal resolution inherent in these monthly outlooks is the fact that the forecasted region captures 80% of the total number of formations while offering a substantial reduction in the total area considered possible for hurricane formation as indicated by climatology or for that matter, seasonal forecast.

Fig 6. Two month lead monthly outlook of NA hurricane formation probabilities generated by our WNP statistical model forced with CFSv2 LSEF data. The formation locations of Oscar, Patty, Rafael, Sandy (near 80° W) and Tony are shown.

While present but not necessarily obvious from Figures 4 and 6, other forecasts that were explored indicate that the WNP statistical model greatly overpredicts hurricane formation in the Gulf of Mexico/Caribbean. Thus, while Dr. Murphree and SSL were happy to get the validating plots that we did, it also became obvious that if we want to issue experimental forecasts for the NA or ENP, models explicit for each basin must be developed and used. To that end, the construction of matrices of data for each basin have been prepared, and once fully populated, will be used for basin specific TC/hurricane model building. It is anticipated that model building will

be complete in time for the beginning of the 2016 TC/hurricane season.

  1. Public Benefit

Key attributes of projects funded by cooperative agreements are that the results of the research are to provide a public benefit, and that the results are to be made publically available so that the public, scientific community, and others may take advantage of the research results. Without doubt, improved understanding of TC/hurricane formation and the influence of the environment on those formations is of great benefit to the public. To make public the research effort, and the ongoing results of that research, SSL, with Dr. Murphree as a co-author, have already published a peer-reviewed conference (2015 Cincinnati-Dayton INFORMS technical symposium) paper titled: A Dynamical-Statistical Approach to Forecasting Tropical Cyclogenesis in the Western North Pacific. This paper, as well as a corresponding presentation made by SSL at the same conference, outlined our TC research to date, our results, and the way ahead. The paper is available at the Cincinnati-Dayton Chapter INFORMS website Both paper and presentation are also available on the SSL website at

Additionally, all experimental forecast products, including keys, legends, and backround materials are posted on the SSL website at: Finally, for purposes of both transparency and public awareness, this quarterly progress report, past and future quarterly reports, and the final technical report for this cooperative agreement will also be made available to the public on SSL’s website.

IV. Way Forward

For the next quarter, SSL will continue with the statistical model building process for both the NA and the ENP. The intent is to be able to issue formation forecasts by 1 June 2016. Part of this process will include an extensive hindcast effort to calibrate the scale of the contours used. This requires striking a delicate balance of achieving as many successes as possible, while minimizing the amount of contoured area per forecast (ideally 10% or less). The remainder of the quarter will be used to issue experimental forecasts and gain insight from real world forecasters, continue the calibration process, gain experience as can only be had by issuing real time, real world forecasts, and begin the TC intensity research effort as delineated in the SOW.

Last, Dr. Murphree and SSL have been discussing potential new journal articles, as the forecasting effort for the NA and ENP has given us insights regarding tropical cyclogenesis worth sharing with the scientific community and the public at large.