Filename
stringlengths 22
64
| Paragraph
stringlengths 8
5.57k
|
|---|---|
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | The objective of this article is to provide the met-ocean practitioner with a careful assessment of the strength of evidence in support of climate change effects on HS from CMIP5- and CMIP6-derived wave output data. The layout of the article is as follows. Section 2 introduces the GCM data, and Section 3 outlines the non-stationary EV methods used. We focus on analysis of POT for reasons of increased statistical efficiency, but also consider inferences based on analysis of BM. Section 4 provides a few illustrative plots of the GCM data. It should be noted that the quantity of data available for analysis (across different GCMs, forcing scenarios, locations and EV modelling) is huge, and we cannot hope to illustrate the data and the analysis adequately in one article. For this reason, the authors have prepared a supplementary material (SM) document (Ewans and Jonathan 2022) to accompany the main text here. The SM document provides a large number of figures and descriptions, referenced throughout this article using the prefix “SM”; thus, supporting figures for Section 4 are provided in SM4. As emphasised above, a quantitative assessment of the inherent variability of GCM output corresponding to no climate forcing is an important basis to frame out assessment of climate change under different forcing scenarios. This analysis is reported in Section 5 for CMIP6 piControl output corresponding to a time period of 700 years, incorporating EV analysis using models assumed either stationary or non-stationary in time. Section 6 then estimates the strength of evidence for changes in return value in time for the GCM output, primarily under RCP4.5 and RCP8.5 climate scenarios, but also for the extended set of CMIP6 scenarios. Supporting descriptions for the statistical modelling in Section 3 are provided in the appendices. Note that for brevity, the term “GCM” is used generically to refer to the source of wave output for all CMIP5 and CMIP6 models (whereas of course the wave output for CMIP5 is from a WAVEWATCH-III wind-wave model, forced by winds from a GCM). For the purposes of the current work, we also treat the forcing scenarios “RCP4.5” (for CMIP5 models) and “SSP245” (CMIP6) are comparable, since they assume the same extent (4.5 Wm−1) of radiative forcing achieved in 2100. We treat scenarios ”RCP8.5” and “SSP585” as comparable since they both achieve 8.5 Wm−1 forcing by 2100. Hence for brevity, we sometimes refer to the “RCP4.5” or “RCP8.5” scenario for all GCMs, on the understanding that when applied to a CMIP6 model, we refer to the corresponding SSP scenario. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Figure 1: Locations of EoM (left) and SoA (right) grid points used. Dots indicate CMIP5, and circles CMIP6. Grey lines are bathymetric contours at 100m and 1,000m water depth. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | The CMIP5 data considered in this study are drawn from those used by Meucci et al. (2020). They consist of WAVEWATCH-III model output, with wind forcing from seven GCMs: ACCESS1.0, BCC-CSM1.1, GFDL-CM3, HadGEM2-ES, INMCM4, MIROC5, and MRI-CGCM3. Wave output including HS are produced at a spatial resolution of 1o degrees, at 6-hourly intervals (Hemer et al. 2012). Data include a historical 27-year period (1979–2005), a mid- 21st century 20-year period (2026-2045) projection, and an end-21st century 20-year period (2081-2100) projection. Output for two different RCPs are provided: RCP4.5, intermediate emissions scenario, and RCP8.5, high-emissions scenario, for each projection period. Meucci et al. (2020) found that the global distribution of mean and 99th percentile HS values from these seven models compared favourably against those from the much larger analysis of output from 83 models by Morim et al. (2019), providing confidence that the subset of seven models is representative of the larger ensemble. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | These is general concern regarding the extent to which the GCMs considered adequate characterise tropical cyclone (TC) events. Meucci et al. (2020) report that the BCC-CSM1.1, MIROC5, and MRI-CGCM3 models capture at least 80% of TCs, as well as winds with speeds more than 30 m/s, when compared with observations made by Shimura et al. (2017) for the western North Pacific region. But in general, the seven-model ensemble was found to only have limited accuracy in representing TC events and tended to underestimate extremes in TC regions. Therefore, whereas some information on TCs can be inferred from the 7-member output considered here, care should be taken in interpreting changes to estimated return values for HS in TC regions. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | In this work, we choose to focus on two interesting geographic regions for detailed analysis, the selection of which is informed by Figure 2 of Meucci et al. (2020), a global map showing changes in 100-year HS between historical and end-21st century periods. We are particularly interested in considering regions within which a strong climate change effect on 100-year HS might be anticipated. Accordingly, we choose data for a region EoM where Meucci et al. (2020) report a significant decrease of more that 10% in 100-year HS, and a region SoA showing a significant increase of more than 10% in 100-year HS. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | In each region we isolate HS data for meridional and zonal transects of grid points with a common centre location, as shown in Figure 1. For each GCM, we then perform a non-stationary EV analysis independently per grid point and per GCM. This procedure allows us to assess spatial trends in time changes of extreme HS per GCM, as well as compare different GCMs. Grey lines in Figure 1 correspond to contours of water depth at 100m and 1,000m. All locations SoA correspond to water depths greater than 1,000 metres. This is generally true EoM, although some of the more eastward locations along the zonal transect are in the more shallow water of the Nazareth Bank. Figure SM2.1 (in Supplementary Material, Ewans and Jonathan 2022) illustrates time-series of peaks over threshold (POT, top row) and annual maxima (AM, bottom row) of HS from the ACCESS CMIP5 GCM for forcing scenario RCP4.5, at the centre location of EoM (left) and SoA (right) neighbourhoods. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | CMIP6 experiments are considered (see Table 2 of Song et al. 2020) : a 700-year pre-industrial period (piControl: nominal years 301-1000), a 165-year historical period (1850-2014), three 86-year future scenarios (SSP125: RCP 2.5, SSP245: RCP 4.5, and SSP585: RCP 8.5, all for the years 2015-2100), and two 150-year climate sensitive experiments (1pctCO2: 1% increase in atmospheric CO2 per year, abrupt-4xCO2: 4% immediate and abrupt increase of CO2 from the piControl value, both for nominal years 301-450). The HS values from the FIO-ESM v2.0 model were available at 3-hourly time steps over a grid varying from 0.27o latitude at the equator to 0.54o latitude in the far north, and 1.125o longitude. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | The spatial resolution for EoM varies between 0.27o to 0.36o in degrees latitude, and 1.25o degrees longitude. Hence, the latitude resolution is in the order of the 0.25o suggested by Timmermans et al. (2017) to be needed to resolve TCs, but the longitude resolution is not. The spatial resolution for SoA is 0.53o in degrees latitude and 1.25o degrees longitude. Figure SM2.2 illustrates the corresponding data to Figure SM2.1, for the CMIP6 model output. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | For POT analysis, storm events are identified by isolating continuous intervals of HS between consecutive up- and down-crossings of a pre-specified threshold level. The threshold level is specified so that more than 20 but not exceeding 25 peaks are selected per annum, for each data set. For analysis of BM, events (corresponding to blocks of length e.g. one or five years, i.e AM or 5-year maxima 5YM respectively) are isolated in the obvious way. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | in year t, where ηS and ηE are the parameter values at the start year (e.g. 1979) and end year (e.g. 2100) of P years of data to be estimated. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | t=1 are empirical annual counts of threshold exceedances, and ρt is also parameterised as in Equation 1. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | when ξt (cid:54)= 0 and µt − σt log[−(1/ρt) log p] otherwise. Note that since all model parameters change in time, then so does the value of Qt. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Parameter estimation is performed sequentially using Bayesian inference. First we perform quantile regression, generating a sample { (cid:98)ψS k=1 of size nI from the joint posterior distribution of EV threshold parameters. We then use the non-stationary threshold corresponding to posterior mean parameter estimates from the quantile regression, (a) in Poisson regression, to generate a sample {(cid:98)ρS k=1 from the joint posterior of ρ; and (b) in GP regression k }nI to sample {(cid:98)σS k=1 from the joint posterior of GP parameters, where nI > 10000. These sets of posterior parameters are used to estimate the distribution of T -year return value, and in particular to compare the estimates Q1 (for the first year) with QP (for the last year) with T = 100. Appendix B provides a discussion on the merits of adopting the posterior mean of quantile regression parameters as the basis for subsequent inference. Figures SM3.1-2 illustrate typical non-stationary EV fits to CMIP5 and CMIP6 samples of POT. Figure SM3.3 gives posterior densities of 100-year POT HS for the start year (1979) and end year (2100) from an EV model for CMIP5 POT output. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | when ξt (cid:54)= 0 and µt−σt log(− log p) otherwise. Parameter estimation is performed using Bayesian inference as described in Appendix A, yielding a sample of nI joint posterior estimates {(cid:98)µS k=1 where nI > 10, 000. These can be used to estimate the empirical distribution of quantities of interest, such as Q1 and QP , and compare them. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | For the stationary EV analysis considered in Section 5, all model GP and GEV parameters η are assumed not to vary with time. Note that if X (Section 3.1) and Xb (Section 3.2) correspond to peaks over threshold and BM of the same physical process, then we expect relationships between the sets of corresponding GEV and GP parameters. We do not consider these relationships in this work. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Figure 2: Comparison of the ordered 40 largest values (T40) of POT HS for the “Historical” (1979–2005), “Middle” (2026-2045) and “End” (2081-2100) time periods, for all CMIP5 GCMs, and the FIO-ESM CMIP6 GCM, for forcing scenario RCP4.5 (and SSP245, top row) and RCP8.5 (and SSP585, bottom row) at the centre location EoM. Left-hand panels give scatter plots for the Middle period on the Historical period; right-hand panels give the corresponding plots for the End period on the Historical period. The line y = x is added for guidance. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | using annual maxima (AM) of HS, and in general the features observed are consistent with those described here for POT. Figures SM4.3-34 provide more extensive comparisons along all transect locations for the CMIP5 ACCESS and CMIP6 GCMs; corresponding figures for other CMIP5 GCMs are available from the authors on request. Figures 2-3 provide a reasonable summary of the trends observed. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Figure 3: Comparison of the ordered 40 largest values of POT HS for the Historical, Middle and End time periods, for all CMIP5 GCMs, and the FIO-ESM CMIP6 GCM, for forcing scenario RCP4.5 (top row) and RCP8.5 (bottom row) at the centre location SoA. Left-hand panels give scatter plots for the Middle period on the Historical period; right-hand panels give the corresponding plots for the End period on the Historical period.The line y = x is added for guidance. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | with Historical. A bootstrapping scheme is employed to estimate the uncertainties in the rate of POT the figures, implemented as follows: first (a) an EV threshold for the analysis is set equal to the 80%ile of the distribution of POT HS over all years of data for the Historical period, then (b) the numbers of exceedances of this level are estimated empirically for each of the Historical, Middle and End periods, and the ratios of rates calculated. Finally (c) we repeat steps (a) and (b) using bootstrap resamples of the full Historical, Middle and End period to estimate uncertainties. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | For EoM (Figure 4) the value of mean and median ratio of rates is around unity or below it, but again there is considerable uncertainty in the ratio of rates within GCM as well as between GCMs. The clearest trend appears for the End:Historical comparison for RCP8.5 forcing scenario. For SoA (Figure 5) the ratio of rates is generally around unity or larger than it, indicating an increase in the rate of occurrence of large POT HS; again this trend is most pronounced for the End:Historical comparison with RCP8.5 forcing. Each panel of Figures 4 and 5 also shows a dashed line and accompanying values, corresponding to the estimated probability that the annual rate of occurrence In general, for EoM this probability is near zero (although there are notable of extreme POT HS has increased. Figures SM4.35-39 exceptions); for SoA, the estimated probability is near unity (again with a few exceptions). provide further illustrations, supporting the general trends observed in Figures 4 and 5: the tail of the distribution of HS becomes shorter with time for EoM, and longer for SoA, under both RCP scenarios. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Figure 4: Box-whisker plots for ratio of annual rates of POT HS at centre location EoM, for all CMIP5 and CMIP6 GCMs. Panels show ratios for the the Middle relative to Historical periods (left) and the End relative to Historical periods (right), for RCP4.5 (and SSP245, top) and RCP8.5 (and SSP585, bottom) forcing scenarios, using a POT threshold corresponding to the 80%ile of the data for the historical period. Each box-whisker indicates the mean (red crosses), median, 25%, 75% quartiles (blue lines), the smallest and largest values not exceeding 1.25 × inter-quartile range from the median (black), and outliers (blue dots). Also shown (dashed black line, and values per GCM) is the estimated probability that the ratio of annual rate exceeds unity. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Figure 5: Box-whisker plots for ratio of annual rates of POT HS at centre location SoA, for all CMIP5 and CMIP6 GCMs. Panels show ratios for the the Middle relative to Historical periods (left) and the End relative to Historical periods (right), for RCP4.5 (top) and RCP8.5 (bottom) forcing scenarios, using a POT threshold corresponding to the 80%ile of the data for the historical period. Each box-whisker indicates the mean (red crosses), the median, 25%, 75% quartiles (blue lines), the smallest and largest values not exceeding 1.25 × inter-quartile range from the median (black), and outliers (blue dots). Also shown (dashed black line, and values per GCM) is the estimated probability that the ratio of annual rate exceeds unity. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Natural climate variability, including the action of inter-annual and longer-term atmospheric oscillations, results in non-stationary temporal effects in met-ocean databases spanning several decades. Nevertheless, the climate might still be considered stationary in the long term, and it is reasonable to expect that the pre-industrial climate would be stationary (e.g. Eyring et al. 2015), representing a period several hundred years prior to the industrial revolution. Thus, pre-industrial data sets offer the possibility to study the effects of natural climate variability on estimates of extremes. In this respect, Bao et al. (2020) examine the time evolution of net radiative flux parameters from the piControl run and conclude that the FIO-ESM model reaches the pre-industrial equilibrium state, after a 300-year spin up period. Consequently, the piControl data set might be expected to represent a stationary climate and we use the data set to asses the natural variability of the wave climate at a location. This allows us to assess the inherent variability in wave height extremes when estimated from a typical length of data. In particular, we would like to understand the natural variability, in a stationary climate, of estimates of the 100-year HS values and their changes, for data lengths corresponding to those typically available for analysis. Our approach to this is in two parts: first to assess the variability associated with the 100-year HS using a stationary extremal analysis, and then using non-stationary EV models. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | We also estimate how the 100-year HS might vary due to the more-or-less arbitrary choices made by the practitioner in making that estimate. For example, we examine the serial variation of estimates for 100-year HS over the 700 years of output, based on EV analysis of a continuous subsample of data corresponding to a specified time interval (henceforth referred to as a “segment length” for clarity, and to distinguish this interval from the “block length” for analysis of BM). We consider the effect of choice of EV threshold for POT analysis, and the choice of block length for a BM analysis on the resulting estimated return value. In particular, we demonstrate that the HS data for EoM represents a mixed population of storm types, demanding the use of higher EV thresholds (POT) and block lengths (BM) for reliable estimation of 100-year HS. We also quantify the zonal and meridional variation of return value for EoM and SoA. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | The panels of Figure 6 show the evolution of estimates for 100-year POT HS based on GP modelling of data corresponding to an interval of time of specific “segment length”, for the centre location EoM. The posterior distribution of return value from the Bayesian inference is summarised in terms of its median value (blue line) and a central 95% credible interval (blue band). The corresponding posterior median and central 95% credible interval, estimated using all 700 years of data, is also shown (orange line and band) in each panel as a benchmark, together with the underlying In this Figure, EV threshold NEP4 is used admitting only the most extreme values to the sample (black dots). analysis. The corresponding figures for NEP1-3 are provided as Figures SM5.1-3. The choice of segment lengths to examine was dictated by the lengths of GCM output available for the various climate scenarios to be considered in Section 6: 86 years for the three future scenarios (SSP1-25, SSP245, and SSP585, respectively), 150 years for the two climate sensitive experiments (1pctCO2 and abrupt-4xCO2) simulations, and 165 years for the historical data set. For comparison, we also include return value estimates based on periods increasing in decadal increments from 20 to 100 years corresponding to the lengths of hindcast data sets that might typically be available for estimating Hs100. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Figure 6 indicates that the variability in posterior median return value in time is largest for the smallest segment lengths in general; changes in posterior median return value of 5m over a relatively short period of time are present, even for segment lengths as large as 50 years. The 95% credible interval for return value is long-tailed with positive skew, often exceeding 25m even though the median value is approximately 12m. It appears that the return value estimates for all segment lengths are, taking a long-term temporal view, approximately unbiased relative to the posterior median return value estimated using the full piControl data. The corresponding behaviour for SoA in Figure 7 (for NEP4, and Figures SM5.5-7 for NEP1-3) is generally similar. Again, for NEP4, estimates of return value are approximately unbiased (over the 700 years) for all segment lengths considered, but the local temporal variability in posterior median return value, and the 95% credible interval for return value, is again huge for small segment lengths. Figures SM5.5- 7 and Figure 7 illustrate that increasing the EV threshold has little effect on return value estimate for the centre location SoA, whereas the posterior distribution of return value, and its temporal variability, are smaller for lower non-exceedance probabilities. This suggests that perhaps NEP1 would be an appropriate choice of EV threshold SoA. The same cannot be said from inspection of Figures SM5.1-3 and Figure 6 for the centre location EoM. Here, for NEP1, NEP3 and NEP4, the posterior median return value is around 12m; but for NEP2, the value is nearer 16m (with wide posterior distributions even for estimation using all 700 years of data). |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Figure 6: Estimates of 100-year HS in time for the CMIP6 piControl data, using EV threshold NEP4. Each panel gives return value estimates from stationary EV analysis using an interval of data of “segment length” specified in panel title, starting at the year indicated on the x-axis, at the centre location EoM. Within each panel, the posterior median estimate is given as a blue line, and its 95% credible interval (CI) as a blue band. Also shown (in orange, and common to all panels) are the corresponding posterior median and 95% CI using the full 700 years of data for EV analysis. In each panel, the sample data is represented as black dots. The panel title also gives the value of non-exceedance probability τ corresponding to threshold level NEP4. The 95% credible intervals for shorter segment lengths have been truncated to ease comparison. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | empirical distribution at around 11m. Fits using NEP1 and NEP2 are poor; the fact that NEP1 provides a reasonable estimate for return value is fortuitous. For comparison, Figures SM5.10 shows the corresponding fits for SoA, all of which appear reasonable, with the possible exception of NEP1. The effect of a mixed population of storms EoM is also observed in the analysis of BM; Figures SM5.11-12 show the posterior median GEV tail fits based on annual maxima (AM), 5-year maxima (5YM for brevity), 10-year maxima and 20 year maxima data. The fit for AM is very poor, but once block length is extended to 5 or more years, an excellent tail fit is obtained. For the SoA region, an excellent tail fit for all block lengths is obtained. We conclude that, if an analysis of BM is desired EoM, block lengths of at least 5 years are necessary to avoid fitting to a mixed population of storms. Physically, we interpret this as the occurrence of tropical cyclone events in the CMIP6 output EoM with a rate of less than one per annum, but generally at least one every five years. Using 5YM data, the effect of segment length on the stationary GEV analysis of piControl output for centre locations EoM and SoA is illustrated in Figure SM5.13. Because we require at least 20 observations to estimate the GEV model, we insist on segment lengths of at least 100 years. The resulting estimates are reassuringly in excellent agreement with those found using POT analysis with NEP4. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Figure 8 summarises the meridional and zonal variation of 100-year POT HS for EoM (top) and SoA (bottom) estimated using stationary GP analysis with threshold level NEP4, in terms of 95% credible intervals for different choices of segment length from 20 years (dark blue) to 165 years (dark red). The calculation procedure is as follows. For each combination of starting time and segment length, we calculate the posterior median return value from the Bayesian inference. We then calculate the 95% interval over all steps for the segment length, and refer to this interval as the “natural 95% interval”, providing an indicative range for our best estimate of return value for a random start year. There is evidence for increasing return value with distance from the equator as might be expected. There are no strong zonal trends. Supporting plots of meridional and zonal trends for different EV thresholds and segment lengths are given in Figures SM5.14-17. The corresponding zonal and meridional trends using AM and 5YM data are also illustrated in Figures SM5.18-19. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Figure 7: Estimates of 100-year HS in time for the CMIP6 piControl data, using EV threshold NEP4. Each panel gives return value estimates from stationary EV analysis using an interval of data of “segment length” specified in panel title, starting at the year indicated on the x-axis, at the centre location SoA. Within each panel, the posterior median estimate is given as a blue line, and its 95% credible interval (CI) as a blue band. Also shown (in orange, and common to all panels) are the corresponding posterior median and 95% CI using the full 700 years of data for EV analysis. In each panel, the sample data is represented as black dots. The panel title also gives the value of non-exceedance probability τ corresponding to threshold level NEP4. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | for the return value using the full sample, and (b) the empirical 95% uncertainty interval of posterior median return values for the segment length given on the x-axis. This empirical estimate of uncertainty is obtained by reading off the 2.5 and 97.5 percentiles of the set of median return values over all starting years for a given segment length. With respect to the right-hand abscissa, the solid, dashed and dotted orange lines indicate the percentage of starting years for which (a) the estimated posterior median return value for given segment length lies outside the 95% credible interval estimated using the full sample, (b) the estimated posterior median return value using the full sample lies outside the 95% credible interval estimated using the given segment length, and (c) there is no overlap at all between 95% credible intervals estimated from the full sample and from a given segment length. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | For both EoM and SoA, the width of the 95% credible interval is very large (around 5m) for 20-year segments, reducing with increasing segment length as might be expected to around 1m. For small segment lengths, the chance that the estimated return value lies outside the 95% credible interval based on the full sample analysis is around 70%, reducing with increasing segment length to around 30% for a segment length of 165 years. Conversely, the chance that the (posterior median) return value estimated using the full sample lies outside the 95% credible interval estimated for a particular segment length is small, around the expected 5% level. Supporting plots are provided in Figures SM5.20-22. We might surmise from the left panel of Figure 9 that there might still be some influence of a mixed distribution of storms increasing observed uncertainties in 100-year HS EoM. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Figure 8: Estimates of meridional (left) and zonal (right) variation of 100-year POT HS for EoM (top) and SoA (bottom) for the CMIP6 piControl data, using EV threshold NEP4. 13 pairs of ordered coloured lines in each panel give 95% credible intervals for the return value from stationary EV analysis using intervals of data of specific length (in years, as in Figures 6 and 7). The calculation procedure is outlined in the text. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Figure 9: Effect of segment length and threshold on estimate of 100-year return value for POT HS at centre location EoM. Panels represent estimates for EV threshold levels NEP3-4. In each panel, with respect to the left-hand y-axis, the dark and light orange bands refer respectively to (a) the 95% credible interval for the return value using the full sample, and (b) the empirical 95% uncertainty interval of posterior median return values for the segment length given on the x-axis. With respect to the right-hand y-axis, solid, dashed and dotted orange lines indicate the percentage of starting years for which (a) the estimated posterior median return value for given segment length lies outside the 95% credible interval estimated using the full sample, (b) the estimated posterior median return value using the full sample lies outside the 95% credible interval estimated using the given segment length, and (c) there is no overlap at all between 95% credible intervals estimated from the full sample and from a given segment length. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | estimates from non-stationary EV models for all available unique segments of a given length in the piControl output. The choice of segment lengths considered again reflects the lengths of CMIP5 and CMIP6 output available for different climate scenarios (see Section 2). The segment length 122 years is a special case, corresponding to the total time span of CMIP5 output. For this segment length, to mimic the time-structure of the CMIP5 data, we subsample the 122 years of piControl output into three appropriately-spaced intervals of lengths 27, 20 and 20 years; therefore the actual number of years of data available for analysis in this case is 67. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Figure 10: Effect of segment length and threshold on estimate of 100-year return value for POT HS at centre location South of Australia. Panels represent estimates for EV threshold levels NEP1-4. For details, see Figure 9. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Figure 11: Kernel density estimates of the median difference in 100-year POT HS (over the length of the segment) estimated using an arbitrary segment of CMIP6 piControl data of given length (in years, see panel legend, from a total run length of 700 years) and a non-stationary GP model with EV threshold level NEP4. Estimates centre location EoM (left) and SoA (right). |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | centre locations EoM (left) and SoA (right). Densities are approximately symmetric, and centred around zero. The density EoM is somewhat broader than SoA. At either location, were we to choose segments of data of length between 86 and 165 years for analysis at random, and use them to estimate the change in median 100-year POT HS over the segment length, we would not be particularly surprised to observe variation of ± 3m regardless of the segment length. Figures SM5.24-31 and SM5.32-39, provide supporting visualisation of the time evolution of the posterior distributions of 100-year POT HS at the start and end years. Figures SM5.40-47 illustrate density estimates for the median difference in 100-year POT HS for all individual combinations of location and segment length. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | of GCM and forcing scenario. The fitted model provides an estimate for the joint distribution of return value for all years of interest in the periods 1979-2100 (CMIP5) and 2015-2100 (CIMP6) and beyond. Derived quantities of interest, e.g. the probability that the return value will increase over some period of time, or the expected size of the change in the return value over time, can then be estimated. Finally, we can average our summary statistics over GCMs to obtain “best overall estimates” for the quantities of interest under a given forcing scenario. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | We estimate the non-stationary GP model introduced in Section 3, for POT of all CMIP5 and CMIP6 output for all scenarios. For each combination of GCM and climate scenario, the MCMC analysis provides a sample from the joint posterior distribution of model parameters. We are then able to calculate the corresponding distribution of other quantities of interest, such as the 100-year POT HS at any time. In particular, we are able estimate the probability that the value of 100-year POT HS for the final year (referred to as “Hs100End”, for the year 2100) is greater than the corresponding value in the first year (referred to as “Hs100Start”, taken here to be the first year of CMIP5 data, 1979). Note that this requires extrapolation in time under the fitted EV model for CMIP6 data. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Figures 12 and 13 summarise our findings for RCP4.5 (and SSP245) and RCP8.5 (and SSP585) for EoM and SOA. Panels show the meridional (left) and zonal (right) trends observed for RCP4.5 (top) and RCP8.5 (bottom). Thin coloured lines represent 100-year POT HS for individual CMIP5 and CMIP6 GCMs, with different line styles for each threshold level NEP1-4 (with only NEP3 and NEP4 for EoM). The thick black lines provide the global mean (solid) and the global median (dashed) over all GCMs and NEPs. There is huge variability between estimated probabilities for different GCMs, and large meridional and zonal variability. However, the global mean and median show more consistency. For EoM, global means and median probabilities tend to be less than 0.5 (suggesting a decrease in return value), whereas for SoA the global probability estimates tend to be larger than 0.5. For SoA, we also see that global estimates of probability for RCP8.5 are larger than for RCP4.5. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Figure 12: Summary of RCP trends for 100-year POT HS and all GCMs EoM. Panels show the meridional (left) and zonal trends observed for RCP4.5 (and SSP245, top) and RCP8.5 (and SSP5-85, bottom) output. Thin coloured lines represent return value using EV threshold NEP3 (dash-dot) and NEP4 (dot) for individual CMIP5 and CMIP6 GCMs, and the thick black lines provide the global mean (solid) and the global median (dashed) over all GCMs and NEPs. The colour scheme for CMIP5 GCMs is given in Figures 2 and 5; in addition, thin black lines are used to represent the CMIP6 values. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Figure 13: Summary of RCP trends for 100-year POT HS and all GCMs SoA. Panels show the meridional (left) and zonal trends observed for RCP4.5 (top) and RCP8.5 (bottom) output. Thin coloured lines represent return value using EV threshold NEP1 (solid), NEP2 (dashed), NEP3 (dash-dot) and NEP4 (dot) for individual CMIP5 and CMIP6 GCMs, and the thick black lines provide the global mean (solid) and the global median (dashed) over all GCMs and NEPs. The colour scheme for CMIP5 GCMs is given in Figures 2 and 5; in addition, thin black lines are used to represent the CMIP6 values. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | exceptions. The opposite is true SoA. Figures SM6.9-16 show broadly similar trends for non-stationary EV analysis of AM. Figures SM6.25-28 illustrate results for the extended set of CMIP6 scenarios, based on non-stationary EV analysis for POT. Again, there is general consistency in estimates of the probability of change in return value for appropriate choice of EV threshold level. However, the large variation with longitude for a given forcing scenario in particular seems difficult to explain on physical grounds. SoA, it is interesting that the 1% CO2 and abrupt 4 × CO2 scenarios tend to be associated with significant reductions in return value. There is some evidence here again for erratic behaviour of EoM estimates using NEP3 (e.g. Figure SM6.25-26 for abrupt 4 × CO2), possibly suggesting the presence of a mixed population of storms under certain scenarios only at this threshold level. Figures SM6.29-32 show corresponding plots for analysis of CMIP6 AM; unfortunately, again because of sample length, analysis of 5YM is not feasible. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | We summarise our findings regarding the values of 100-year POT HS at the start (1979) and end (2100) of the observation period in Figure 14. The figure summarises our beliefs about the return value at the start and end time points in terms of the mean (over all GCMs) of the posterior median return value (per GCM) meridionally and zonally, for both EoM and SoA regions and each of the RCP4.5 (and SSP245) and RCP8.5 (and SSP585) scenarios. For both regions, there is evidence that return value increases with increasing latitude, but there is little zonal variation. For the EoM region, there is general evidence for both RCP scenarios that return value reduces in time by approximately 0.5m. SoA, there is general evidence for an increase in return value over the period of approximately 0.5m for RCP4.5 and 1.0m for RCP8.5. Supporting Figures SM6.17-24 give estimates for the change in 100-year POT HS for individual GCMs, and estimate the significance of those changes, under the RCP4.5 and RCP8.5 scenarios. Results again show large variability across GCMs, and little evidence of significant changes in return value. However, there is a tendency for generally reducing trends EoM, and generally neutral or increasing trends SoA. Figures SM6.33-44 give a detailed breakdown of estimates for the meridional and zonal variation of return value for individual GCMs and the start and end years of the study. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Figure 14: Summary of inferences for 100-year POT HS return value from all CMIP5 and CMIP6 GCMs. Panels show the mean (over all GCMs) of the posterior median return value at the start time (blue) and the end time (orange), for RCP4.5 (and SSP245, solid) and RCP8.5 (and SSP585, dashed) for EoM (top) and SoA (bottom) meridionally (left) and zonally (right). |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Estimation of return values from relatively small samples of data is problematic (e.g. Serinaldi 2015, Jonathan et al. 2021). In their basic analysis step for CMIP5 output, Meucci et al. (2020) use a stationary EV analysis of data for a fixed time period of approximately 20 years. They then compare return value estimates, and report that, regardless of which individual GCM is considered, only rarely is a statistically significant change (at 95% level) in 100-year HS at a location estimated. For this reason, those authors then aggregate standardised data from different GCMs to increase sample size for analysis (but again, for 20-year time periods only). Using the resulting composite data, significant changes in 100-year HS are found. In the current analysis, we choose to estimate independent non-stationary EV models for the output of each GCM, using all available data (for CMIP5 or CMIP6) in one analysis, and then estimate the statistical characteristics of return value, and time differences in return value per GCM. Finally, we combine inferences about return values and return value differences over GCMs. We believe our approach has advantages, for the following reason: Meucci et al. (2020) combines samples of data drawn from different GCMs, likely to have different statistical properties, together and assumes that extremes of the composite sample can be considered drawn from a homogeneous tail. Although the standardisation step used by Meucci et al. (2020), equating the first and second sample moments of the distributions from each GCM is widely used, the resulting composite sample corresponds to a mixture of different distributions with different tail characteristics in general. Estimating a common tail using such data usually leads to bias, as shown in Jonathan et al. (2008). The approach used in the current work treats each GCM as an independent source of estimates for 100-year HS and its uncertainty using a statistical model appropriate for all relevant data. We then come to a final view by aggregating 100-year HS estimates over all GCMs. Despite the differences in approach, the overall conclusions of this work are broadly in agreement with those of Meucci et al. (2020). |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | region, despite large variations in historical storm tracks. Vanem (2017) reports that spatial smoothing via regional frequency analysis can be a useful tool to estimate climate change more efficiently for storms in the North Atlantic. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | In the current work, we observe considerable variability in predictions of changes in storm severity at any of a number of locations east of Madagascar (EoM) and south of Australia (SoA) from a set of 7 CMIP5 and one CMIP6 GCM. Examination of predicted storm severity along meridional and zonal transects within each region also reveals large differences in return value estimated at a given year, and change in return value over a number of years. It is reasonable therefore to be rather suspicious that specific inferences from the work (e.g. for specific GCMs, locations, years or choices of modelling hyper-parameters) may be biased to a considerable extent, compared to more general inferences (e.g. those summarising over GCMs, locations, years and modelling hyper-parameters). |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | We adopt non-stationary extreme value (EV) models for estimation of return values, within which all model param- eters are allowed to vary linearly with time. We consider the restriction to linear trends appropriate, since we expect that the effect of climate change on the tail of the distribution of HS to be relatively small; hence the data for analysis is unlikely to provide evidence in favour of more complex models. We estimate the models using Bayesian inference, using simple MCMC algorithms coded in MATLAB, available at Jonathan (2021). We see that careful specification of the EV model, and choice of EV threshold (for POT analysis) or block length (for BM analysis) is critical. EoM, we find that relatively large thresholds (POT, or block lengths for BM) are necessary to avoid bias in return values causes by a mixed population of extratropical storm and occasional tropical cyclones. SoA, we find that threshold (or block length) selection is less critical. Results using non-stationary GP modelling of POT are generally in very good agreement with those from non-stationary GEV modelling of BM, provided that thresholds and block lengths are set sensibly. Indeed, the estimated return value (and change in return value) appears less sensitive to threshold (or block length) choice than to the choice of GCM, or the arbitrary choice of a location with a geographic neighbourhood, again provided that the range of thresholds (or block lengths) admitted is physically plausible to avoid issues with fitting to heterogeneous tails. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Non-stationary EV analysis of the assumed steady-state 700-year piControl data suggests that estimates as large as ±3m for the change in 100-year return value of POT HS in N years, made from samples corresponding to N years of observation for N ∈ {86, 122, 150, 165}, are not unusual. We propose a simple time-randomisation scheme to assess the significance of observed changes in 100-year HS, and use the 700-year long piControl output from the CMIP6 GCM to explore the sensitivity of estimates of return value to the “segment” length of data available as the basis for estimation, and the arbitrary time during the 700-year period at which the data are selected. For a segment length of 20 years for EV analysis, the 95% credible interval for 100-year HS is around 5m wide, reducing to around 3m for a segment length of 100 years. For a 20-year segment length, there is a probability of around 0.7 that the estimated 100-year return value lies outside the 95% credible interval for the return value using a segment length of 700 years. Large meridional and zonal uncertainties in return value estimates for small segment lengths in particular suggests that joint spatial modelling may be advantageous, or more pragmatically that making best estimates for a neighbourhood by averaging return values estimates, made independently for locations within the neighbourhood, may be useful (McSweeney and Jones 2013). Despite huge between-GCM and large within-neighbourhood variability found, averaging predictions over climate models and space suggests that a lowering of storm severity is likelier than not EoM under RCP4.5 (and SSP245) and RCP8.5 (and SSP585) scenarios, whereas the opposite is true SoA. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | The authors would like to thank Simon Brown and Rob Shooter (UK Metoffice), Alberto Meucci and Ian Young (U. Melbourne, Australia) and Zhenya Song (First Institute of Oceanography, Qingdao, China) for discussions. The supplementary material (SM) for the current article is available online at Ewans and Jonathan (2022). Software for the non-stationary extreme value analysis is provided at Jonathan (2021). |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Inference for the GEV, GP, quantile and Poisson regression models described in Section 3 is performed using Markov chain Monte Carlo (MCMC, see e.g. Gamerman and Lopes 2006) following the method of Roberts and Rosenthal (2009). In this approach, all the parameters θ of the model are jointly updated for a sequence of nB + nI MCMC iterations. At each iteration, a new set of parameter values is proposed, and accepted according to the Metropolis- Hastings acceptance criterion based on (a) the sample likelihood evaluated at the current and candidate states, and (b) the values of the prior densities for parameters at the current and candidate states. Following a certain number nB of so-called burn-in iterations, the Markov chain is judged to have converged, so that the subsequent nI iterations provide a valid sample from the joint posterior distribution of parameters. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Diffuse prior distributions were specified as ξ ∼ U (−0.5, 0.2); σ ∼ U (0, 10); µ, ψ ∼ U (−10, 10). Likelihoods for the models are available from the distributions given in the main text. An appropriate starting solution θ1 for the MCMC inference was obtained by random sampling from the prior distributions of parameters, ensuring a valid likelihood. For the first nS < nB iterations, candidate parameter values θc k ∼ N (0, 0.12I) (cid:1) + βN (cid:0)θk−1, 0.12/4(cid:1), where following Roberts and Rosenthal (2009). Thereafter θc β = 0.05, Σk is the empirical variance-covariance matrix of parameters from the past k iterations, and θk−1 is the current value of parameters. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Throughout, a candidate state is accepted using the standard Metropolis-Hasting acceptance criterion. Since prior distributions for parameters are uniform, and proposals symmetric, this criterion is effectively just a likelihood ratio. That is, we accept the candidate state with probability min(1, L(θc)/L(θ)), where L(θ) and L(θc) are the likelihoods evaluated at the current and candidate states respectively, with candidates lying outside their prior domains rejected. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | As discussed in Section 3.1, in this work the posterior mean quantile regression thresholds (cid:98)ψS and (cid:98)ψE (corresponding to the start and end years of data) are used for subsequent estimation of EV and threshold exceedance rate models. Posterior mean (cid:98)ψS and (cid:98)ψE are also used for simulation under the model, and estimation of distributions for T -year return values. Since estimation of EV models is generally considerably more uncertain that EV threshold, our neglect of threshold uncertainty was not considered an important deficiency. To investigate this concern further, we estimated the distribution of 100-year return value for the start and end years using fitted models, under different procedures for propagation of uncertainty in EV threshold. Results are illustrated in Figure SMB.1 for a sample corresponding to an interval of P = 86 years of PI Control data offshore Madagascar. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Three possible approaches to propagating threshold uncertainty are illustrated, and correspond to different choices regarding the non-stationary threshold estimate used for simulation (see Equation 1, with η = ψ) based on ψS and ψE. These choices use (a) independent draws of (cid:98)ψS and (cid:98)ψE from their posterior marginal distributions, (b) a common draw of ( (cid:98)ψS, (cid:98)ψE) from its posterior joint distribution, and (c) the posterior mean values of (cid:98)ψS and (cid:98)ψE. The figure shows that variation in the distribution of the 100-year return value from choices (a)-(c) is small, for the start and end years, and for the different threshold non-exceedance probability values τ given in the four panels. Similar results are observed for other locations and time periods. For this reason, we consider it reasonable to use the posterior mean (cid:98)ψS and (cid:98)ψE for estimation of return values. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Chavez-Demoulin, V., Davison, A., 2005. Generalized additive modelling of sample extremes. J. Roy. Statist. Soc. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Eyring, V., Bony, S., Meehl, G.A., Senior, C.A., Stevens, B., Stouffer, R.J., Taylor, K.E., 2015. Overview of the Coupled Model Intercomparison Project Phase 6 (CMIP6) experimental design and organization. Geosci. Model Dev. 9, 1937–1958. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Lobeto, H., Menendez, M., Losada, I.J., 2021. Future behavior of wind wave extremes due to climate change. Sci. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Meucci, A., Young, I.R., Hemer, M., Trenham, C., Watterson, I.G., 2022. 140 Years of Global Ocean Wind-wave Climate Derived from CMIP6 ACCESS-CM2 and EC-Earth3 GCMs. Global Trends, Regional Changes, and Future Projections. J. Clim. doi:10.1175/JCLI-D-21-0929.1. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Morim, J., Hemer, M., Wang, X.L., Cartwright, N., Trenham, C., Semedo, A., Young, I., Bricheno, L., Camus, P., Casas-Prat, M., Erikson, L., Mentaschi, L., Mori, N., Shimura, T., Timmermans, B., Aarnes, O., Breivik, Ø., Behrens, A., Dobrynin, M., Menendez, M., Staneva, J., Wehner, M., Wolf, J., Kamranzad, B., Webb, A., Stopa, J., Andutta, F., 2019. Robustness and uncertainties in global multivariate wind-wave climate projections. Nat. Clim. Change 9, 711–718. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Morim, J., Trenham, C., Hemer, M., Wang, X.L., Mori, N., Casas-Prat, M., Semedo, A., Shimura, T., Timmermans, B., Camus, P., Bricheno, L., Mentaschi, L., Dobrynin, M., Feng, Y., Erikson, L., 2020. A global ensemble of ocean wave climate projections from cmip5-driven models. Sci. Rep. 7, 105. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Parding, K.M., Dobler, A., McSweeney, C.F., Landgren, O.A., Benestad, R., Erlandsen, H.B., Mezghani, A., Gregow, H., Raty, O., Viktor, E., El Zohbi, J., Christensen, O.B., Loukos, H., 2020. GCMeval - An interactive tool for evaluation and selection of climate model ensembles. Clim. Serv 18, 100167. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Roberts, G.O., Rosenthal, J.S., 2009. Examples of adaptive MCMC. J. Comp. Graph. Stat. 18, 349–367. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Sansom, P.G., Stephenson, D.B., Ferro, C.A.T., Zappa, G., Shaffrey, L., 2013. Simple uncertainty frameworks for selecting weighting schemes and interpreting multimodel ensemble climate change experiments. J. Clim 26, 4017 – 4037. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | IPCC, 2021. Summary for Policymakers. In: Climate Change 2021: The Physical Science Basis. Contribution of Working Group I to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change. URL: www.ipcc.ch/report/ar6/wg1/downloads/report/IPCC_AR6_WGI_SPM.pdf. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Vanem, E., 2017. A regional extreme value analysis of ocean waves in a changing climate. Ocean Eng. 144, 277–295. |
Processed_Uncertainties_in_estimating_the_effect_of_climate_.txt | Vehtari, A., Ojanen, J., 2012. A survey of Bayesian predictive methods for model assessment, selection and comparison. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | ABSTRACT Unmanned Aerial Vehicles (UAVs) have moved beyond a platform for hobbyists to enable environmental monitoring, journalism, film industry, search and rescue, package delivery, and entertainment. This paper describes 3D displays using swarms of flying light specks, FLSs. An FLS is a small (hundreds of micrometers in size) UAV with one or more light sources to generate different colors and textures with adjustable brightness. A synchronized swarm of FLSs renders an illumination in a pre-specified 3D volume, an FLS display. An FLS display provides true depth, enabling a user to perceive a scene more completely by analyzing its illumination from different angles. An FLS display may either be non-immersive or immersive. Both will support 3D acoustics. Non-immersive FLS displays may be the size of a 1980’s computer monitor, enabling a surgical team to observe and control micro robots performing heart surgery inside a patient’s body. Immersive FLS displays may be the size of a room, enabling users to interact with objects, e.g., a rock, a teapot. An object with behavior will be constructed using FLS-matters. FLS- matter will enable a user to touch and manipulate an object, e.g., a user may pick up a teapot or throw a rock. An immersive and interactive FLS display will approximate Star Trek’s Holodeck. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | application of Section 1.1. Advantages of FLS displays compared with today’s flat screens will be multi-fold. First, their rendered luminations will capture reality more accurately. A lumination will be rendered by controlling position of FLSs in a 3D volume that has depth. Second, their 3D luminations provide depth to the naked eye. Moreover, a user looking at an FLS display from different angles may observe different content. Third, FLS displays may be immer- sive by occupying an arbitrary space, transforming a wall, a ceiling, a room, a concert hall into a 3D depth display. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | With immersive displays, we introduce FLS-matter, a swarm of FLSs that will realize a geometric shape in 2D (e.g., square) and 3D (e.g., a pyramid). Its lumination will scale to render its shape in different sizes. FLS-matter is the building block of an object with behavior. As detailed in Section 4, it will enable a user to interact with objects rendered as luminations. An interactive immersive FLS display will manipulate both the user’s depth perception and spatial awareness, approximating Star Trek’s Holodeck. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | A successful realization of the research ideas presented in this pa- per will provide fundamental insights into implementing a Holodeck using swarms of FLSs. A Holodeck will transform the future of hu- man communication and perception, and how we interact with information and data. It will revolutionize the future of how we work, learn, play and entertain, receive medical care, and socialize. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | To maintain the same resolution, the number of FLSs required by a lumination increases as a function of its size. Assuming today’s 0.264 millimeter pixels, Table 1 shows the number of FLSs required to display a sphere as a starting point for a 3D Earth. As size is increased from a ping-pong to a basketball, the number of FLSs increases to hundreds of thousands. With a room sized (10 feet by 10 feet) lumination, the number of FLSs is hundreds of millions. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | not the assistants, even though they must collaborate. There have been studies that quantify the benefits of providing assistants with their own 3D video monitor using color depth [103]. A large FLS display will enable the surgeon and assistants to view the same 3D lumination to collaborate effectively. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | Entertainment and gaming: FLS displays will transform enter- tainment and gaming applications. An example is multi-player fight games such as Mortal Kombat, Street Fighter, Fortnite and others. FLS displays enable players to view the fight scene from different angles. These luminations might be on a game board that sits on a table. It may incorporate today’s flat screens as its bottom. For example, in a scene that shows multiple characters standing on a wooden flooring, a flat screen will render the wooden flooring while FLSs delineate the 3D fighting characters. Players may par- ticipate either remotely or in-person by either sitting or standing around the game board. Players may use either a traditional game controller or wear body sensors that translate their movement into luminations. With the latter, when the players are in the same room, they may see the posture of their opponent(s) and anticipate their next move, introducing new gaming experiences. Players may also get a good cardiovascular exercise from playing such games. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | Design and manufacturing: Display of a 3D design using flat 2D displays is not natural and may result in human errors. 3D printing empowers a designer to evaluate a final design prior to manufac- turing it. An FLS display will fill the gap between design and 3D print by enabling a designer to see their 3D designs and examine them from different perspectives in real-time. With this application, an FLS display maybe the size of a game board that renders small designs (knots and bolts of an airplane seat) or the size of a room for larger designs (the airplane). Both will enable a designer to zoom in and out to scrutinize components. This may be done using a traditional keyboard and mouse, a gesture based interface, or more novel interfaces (see Section 8). FLS displays may be an integral part of design. For example, they may render luminations of a sim- ulation that stresses a design to analyze its failure thresholds. This may include lumination renderings from real environments with manufactured components, e.g., flight recorder data from a crashed airplane. These concepts apply to manufacturing where an FLS display may render the machine tools on a factory floor and how they manufacture a product, e.g., an automobile. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | To the best of our knowledge, the first two ideas are novel and not described elsewhere. See [1, 21, 84] for a survey of applications using UAVs and swarms. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | The rest of this paper is organized as follows. Section 2 describes related work. Sections 3-9 detail challenges of FLS displays. Sec- tion 10 concludes this paper. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | With outdoor shows, a drone’s GPS enables it to be at the right position and follow the correct path in the night sky to illuminate its light for the desired effect. Millimeter drones of an FLS display must be positioned with micrometer precision. Even if GPS was available, it would be too coarse [58]. The best some GPS devices claim is 3 meter accuracy 95% of the time. The United States govern- ment currently claims 7.8 meter with 95% confidence for horizontal accuracy for civilian (SPS) GPS. Real-Time Kinematic (RTK) GPS provides centimeter-level accuracy using additional base stations. Vertical accuracy is worse with GPS solutions. Thus, an FLS display requires novel positioning systems as described in Section 3.1. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | There are other significant differences between show drones and FLS displays. First, the number of drones in FLS displays will be significantly larger. According to the Guinness World Record, the most unmanned aerial vehicles in simultaneous flight for a light show is 3,281 for tens of minutes spread over a kilometre in Shanghai, China on March 29, 2021 [114]. The number of drones in an FLS display will be several orders of magnitude larger. Second, the area used by an FLS display may be significantly smaller. An FLS display may be as small as a standard game board. This is orders of magnitude smaller than the night sky of outdoor drone shows and concert halls [83] of indoor shows. Third, each FLS is significantly smaller than show drones. This enables FLSs to operate in tightly constrained environments in tight formations with higher accelerations, providing higher stability and more rapid response to changes in luminations [53]. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | Systems such as Northrop Grumman’s Virtual Immersive Portable Environment (VIPE) Holodeck surrounds a user with 2D moni- tors [65]. VIPE uses Kinect to detect a user action such as crawling and jumping to display the appropriate training content on the 2D monitors. Our proposed immersive 3D FLS display may replace VIPE’s 2D monitors, manipulating both the trainee’s depth percep- tion and spatial awareness to provide a more realistic experience. Virtual Reality (VR) [34, 52, 91] is a computer generated sim- ulation of a 3D image or environment. A user may interact with this environment using equipment such as a head-mounted dis- play with a screen inside or gloves fitted with sensors. Augmented Reality (AR) [29, 32, 91] uses technology such as a mobile phone, tablet, or AR glasses to present the user with an enhanced version of the real physical world. The enhancements may include digital visual elements, sound, or other sensory stimuli. An immersive FLS display is different than both by providing the human sense of touch with no gloves and true depth with no mobile phone or head mounts. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | parts) around the display to show motion graphics. This idea is inspired by the T-1000 robot from the Terminator 2 movie. To realize such a display, 3D printing must become orders of magnitude faster and polymer chemistry has been suggested as a solution [26]. An FLS display employs small drones with one or more light/reflective sources to generate color and texture. At the same time, an FLS display may incorporate fast 3D printing to represent objects in a lumination that are stationary enough to match the speed of 3D printing. This hybrid display requires lighting of the 3D printed objects. Alternative solutions may include backlighting or FLSs that project light onto these printed objects. The rest of this paper focuses on challenges of FLS displays and defers hybrids using 3D printing to future work. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | Today’s depthcams can be classified into those that use either Time-of-Flight (ToF), Triangulation, or both. For example, Microsoft’s Kinect One uses a flash of light to illuminate a scene and thousands of sensors measure the return time from different parts of the scene. This pixelated ToF device measures depth of a million points per second. A similar technology is employed by iPhone X and its dot projector [43] to create a 3D map of the user’s facial struc- ture [62, 86]. The LiDAR Intel RealSense depthcam uses a variant of the same technique by employing Lidar used in some automo- biles. This depthcam uses a proprietary MEMS micro scanning technology [112] to acquire depth data for indoor applications. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | The D4xx series of Intel RealSense [112] depthcams use triangu- lation technique via stereoscopic imaging. It uses two cameras to record a scene and compare image patches from one camera with image patches in the other camera to collect depth data. The basic idea is that those objects that are closer to the camera are displaced more in the horizontal axis. A limitation of this technique is with low texture scenes that cause the image patches from the different cameras to become ambiguous, e.g., a white wall with no texture. In this scenario, a ToF technique using infrared (IR) light may be used to acquire depth data. This technique is used by Intel RealSense D43x and D45x depthcams. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | A depthcam may include Inertial Measurement Unit (IMU) to capture 6 degrees of freedom data2, e.g., Intel RealSense D435i. The structure from motion of the depthcam may be used to capture depth data [24]. Moreover, this data may be fused with data from other sensors to refine depth data and make it more accurate. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | Today’s depthcam technologies are used in games, to authenti- cate users, and to navigate a robot in an obstacle course. We foresee their evolution as recording devices for FLS displays. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | must be at the right (x,y,z) coordinate of the display as a function of time and illuminate either (a) light at the pre-specified color, brightness, and texture or (b) provide a reflective surface for an external light source to generate color and texture. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | Figure 1 shows the block diagram for continuous display adapted from the autonomous vehicles studies [14, 73]. Perception will trans- form streams of sensory data and the meta-data describing a lumi- nation to a contextual understanding of the environment around each FLS and its goal. Planning refers to the decisions that must be made by each FLS to execute its goal from a given start position. It will generate streams of decisions that dictate the velocity, flight path, color/texture and brightness of light generated by each FLS. These streams may also control external and back lighting of a scene. A continuous display of a lumination is the ability of FLSs to execute the collision-free flight decisions in a timely manner. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | To illustrate, consider a lumination of a running person. If the pipeline of Figure 1 is processed too slow then the display may show walking instead of running. At its worst, the rendering may result in a distorted collection of lights in a 3D space. A continuous display avoids these undesirable luminations by requiring each FLS or grouping of FLSs to be at the right coordinate at the right time to render the right color. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | Continuous display is challenging for several reasons. First, flight of FLSs must avoid collisions. This challenge is exacerbated by com- plex scenes that may require a large number of FLSs that must render their color in close proximity to one another. This is because no positioning system guarantees 100% accuracy and reported po- sition of FLSs may be erroneous. Second, the scene to be displayed may not be known in advance. An example is a real-time stream generated by a RealSense depthcam. Third, failure of FLSs is the norm rather than the exception, see Section 6. A continuous display technique must render a lumination even though its FLSs are failing frequently. Fourth, battery lifetime of FLSs is finite and power is scarce. Fifth, the network bandwidth is finite and there is a latency for one FLS to communicate with another. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | These challenges motivate algorithms and frameworks in sup- port of a positioning system, a planner to compute flight path and lighting of each FLS as a function of time and space, a communica- tion network for FLSs, and a system to charge FLS batteries. Below, we describe these components in turn. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | due to lack of line of sight with orbiting satellites3. A 3D FLS dis- play requires a fine-grained positioning system to enable FLSs to maintain and control their position with hundreds of micrometer accuracy and 95% confidence. Indoor Location Systems (ILS) have been proposed based on Magnetic signals, Vision-based, Audible sound, Infrared (IR), Ultra-sound (UPS), Radio Frequency (RF) in- cluding RFID and WLAN, Bluetooth, Ultra-Wideband (UWB) and Ultra High Frequency (UHF), or a hybrid of these technologies. See [38, 44, 60, 63, 113] for a survey of these technologies including their strengths and limitations. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | Several studies report optical motion capture systems enable millimeter-level position tracking using multiple infrared cam- eras [53, 75, 80]. This solution requires a unique marker arrange- ment for each drone. The small size of an FLS may challenge the applicability of this solution. For example, [75] reports that it was impossible to form unique markers for 49 drones with each mea- suring 92 millimeters diagonally. They develop a method initialized with known positions that is subsequently updated using frame- by-frame tracking. While such techniques may be a good starting point, we believe decentralized positioning systems that include active participation of FLSs will be developed. In its simplest, FLSs may detect and either repel one another [55, 90] or maintain a fixed geometry. This will minimize the likelihood of FLS crashes when a positioning system provides inaccurate positions to FLSs. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | Several studiesdetail path planning algorithms for UAVs, e.g., outdoor light shows [89], search and rescue [42], services in an urban area [105], and environmental sensing [85]. A challenge is to move from a given initial positions to a set of predefined targets while avoiding collisions with obstacles as well as other UAVs [4, 5, 7, 13, 15, 19, 20, 30, 33, 39, 53, 68–70, 74, 75, 87, 92, 97– 100, 108]. Some studies consider downwash, a region of instability caused by the flight of one UAV that adversely impacts other UAVs entering this region [4, 7, 33, 75, 100, 108], e.g., loss of control or unpredictable behavior. This effect is considered in collision avoidance by modelling each UAV as axis-aligned ellipsoids [4, 7, 75] or cylinders [33, 108] that results in a larger separation along the Z-axis. The latter is appropriate for quadrotors. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | Today’s collision avoidance techniques may be categorized into centralized [5, 15, 19, 39, 53, 70, 70, 74, 75, 89, 92] and decentral- ized [4, 7, 13, 20, 33, 68, 69, 87, 92, 97–100, 108]. A centralized algo- rithm executes on a CPU with abundant amount of memory and storage. A decentralized algorithm may use the limited process- ing and storage capability of each FLS, requiring multiple of them to collaborate by communicating with one another. Centralized algorithms may further be divided into online [15, 19, 75] or of- fline [5, 39, 53, 70, 70, 74, 89, 92]. Offline techniques assume a static environment with known obstacles, a known scene, and reliable UAVs. Online techniques minimize the number of such assumptions. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | To illustrate, [89] presents a centralized, offline algorithm to compute a flight/lighting plan for outdoor light show performances. This algorithm requires drones to be placed in a field and in a spe- cific arrangement. Each drone is provided with a pre-computed flight path and timed light display. All drones take off in a specific order, fly to the performance area, spread out and turn into the first performance pattern, change the formation into the next pat- tern, and repeat until all pre-specified performance patterns are complete. Finally, the drones return into a formation and fly back to the takeoff position. This algorithm is centralized because it as- sumes complete information of the field, intended display patterns and their duration, obstacles, and participating drones. It is offline because it executes once, producing a plan for each drone and com- municating it to that drone. If during the light show, a new obstacle is introduced (e.g., a bird), an online algorithm is required to com- pute a new plan for the impacted drone(s). This online algorithm is more complex than the offline algorithm because it must monitor the environment, detect new obstacles and the impacted drones, compute a new plan with a new flight path for each drone, identify the drones whose flight paths have changed, and communicate the new flight path to these drones. An offline algorithm does not have these steps. It may not even require drones to communicate because the drones are provided with a flight path once at the beginning. It is important to note that centralized offline algorithms are of value because they may guarantee optimality in terms of metrics such as minimum flight time, time to realize a lumination, or amount of power consumed. They can be used as yardsticks to evaluate online algorithms (both centralized and decentralized) that are typically heuristic based. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | Collision avoidance techniques may further be categorized into those that evaluate their technique using a simulation [4, 20, 39, 68, 70, 75, 87, 98], an implementation using either outdoor UAVs, robots, or indoor quadrotors [5, 19, 92], or both [7, 13, 15, 33, 53, 69, 74, 89, 97, 99, 100, 108]. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | Existing techniques and algorithms are a great starting point to develop an online planning component in support of continuous display. An ideal algorithm or framework should provide a scalable execution time as a function of computing resources. This means a centralized algorithm should execute faster as a function of the number of CPU/GPU cores used to execute the algorithm. Similarly, a decentralized algorithm should execute faster as a function of the number of participating FLSs. Both concepts are challenging to realize in practice. |
Processed_Holodeck:_Immersive_3D_Displays_Using_Swarms_of_Fl.txt | The concept of scalability extends beyond algorithms to include the display itself. For example, a scalable positioning system will enable players of a game to stretch their 13 inch game board to 40 inches. To maintain the resolution of luminations, users may add new FLSs. It is interesting to note that with a decentralized algo- rithm for planning, the additional FLSs will increase the processing capability of the system to compute spatio-temporal plans automat- ically. A centralized algorithm may require increased processing capability (e.g., additional cores) to schedule the new FLSs. This may require a shut-down and restart of a game board. An elastic decentralized algorithm will incorporate the new FLSs and continue the game without starving for processing capability. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.