Observing and forecasting the retreat of northern Australia’s rainy season

2.1. Observations and hindcasts

Daily gridded rainfall data are from The Bureau’s Australian Gridded Climate Data (AGCD) (ver. 1; Bureau of Meteorology 2019), provided on a 0.05° (5-km) grid. The rainfall data cover 1 January 1950 through to 31 December 2019, providing an adequate sample size to compare wet season retreats for different ENSO phases. By the late 1950s, there was an increase in the number of operational rain gauges (Jaffrés et al. 2018), although gaps remained in Arnhem Land and the Gulf Country. The spatial analysis is restricted to regions north of 30°S; however, as shown later, regions that are climatologically dry or do not have a well-defined annual cycle of rainfall can be difficult to interpret and often do not feature a retreat date (or an onset date; Smith et al. 2008). A data quality mask is applied to regions, predominantly in central WA, that have a poor daily rainfall coverage (Griffiths et al. 2023).

Hindcast and forecast data are utilised from ACCESS-S2, a model based on the UK Met Office’s GloSea5-GC2 seasonal prediction system (Wedd et al. 2022). The horizontal resolution of the atmospheric component of ACCESS-S2 is 60 km in the midlatitudes and across Australia. ACCESS-S2 also uses its own data assimilation to produce its own initial conditions instead of relying on the Met Office assimilation scheme, and has dynamic soil moisture initial conditions as opposed to climatology. Here, the 5 km calibrated hindcasts from ACCESS-S2 are used, which removes the model biases, e.g. using a quantile–quantile matching approach (Griffiths et al. 2023). This is important because ACCESS-S2 tends to have a wet climatological bias over northern Australia tropical regions in the late summer and early autumn seasons (Li Shi, pers. comm.). The hindcast period is 1981–2018, and for seasonal-scale predictions, there are three 279-day model ensemble members that have been generated on the 1st of every month and on each of the previous 8 days. This produces a 27-member time-lagged ensemble (3 members × 9 start days) with which to evaluate ACCESS-S2’s skill in predicting the NRR over the hindcast period. For the last year of the hindcast record, 2018, we only use eight start dates as there is no 1 January 2019 simulation.

2.2. NRR definitions

Based on a review of recent literature, three different ways to define the NRR are selected:

The above definitions require a fixed ‘cut-off’ date for the wet season, to prevent isolated rain events from being inadvertently considered in the post-wet season. In this study, the selected date is 30 April, in line with The Bureau’s formal definition of the end of northern Australia’s wet season and tropical cyclone season. Later in the study, we test how shifting the cut-off date to 31 May affects the NRR date. We also begin the analysis of each NRR period from 1 January. Below, the three different definitions for determining the NRR date are described in more detail, with reference to Fig. 1.

Fig. 1.

The three different methods for defining the northern rainfall retreat date, based on daily rainfall (blue bars) for Darwin during January–April 1981. The three methods are (1) maximum in the cumulative daily rainfall anomaly before 30 April (black curve and dot), (2) backwards accumulation of daily rainfall from 30 April until 50 mm is reached (red curve and dot) and (3) last ‘25 mm in 3-day’ burst before 30 April (green thick lines and dot). The left y-axis is for methods (2) and (3), and the right y-axis is for method (1). Rainfall is rounded to the nearest 0.2 mm.

2.2.1. Definition #1: cumulative anomaly

The cumulative anomaly definition stems from research into the duration of the northern Australian rainy season by Uehling et al. (2021) and Uehling and Misra (2020). This definition has been used to diagnose the onset and retreat of the Indian summer monsoon (Noska and Misra 2016) and South American rainy season (Liebmann et al. 2007). To determine the NRR using this definition, the cumulative daily anomaly of rainfall C′k for each day i of each year k for each grid point is calculated as follows:

(1)C′k(i)=∑n=1i[[rk(n)−C ¯¯]

where the annual mean climatology of rainfall C¯¯ over N days and K years is as follows:

(2)C¯¯=1NK∑k=1K∑n=1Nr(n,k)

where r(n,k) is the daily rainfall over the multi-year period.

Hence, the cumulative anomaly definition describes the removal of a multi-year daily mean amount from each day’s rainfall total and then this anomaly is accumulated from a given start date (i.e. 1 January). To illustrate this, we use daily rainfall from Darwin in 1981 from 1 January to 30 April (blue bars, Fig. 1). As the wet season progresses in this year, the cumulative daily anomaly C′k increases following each rainfall event and slowly declines during break (in rain) periods due to the negative daily anomalies (e.g. 19 February to 3 March; see black ‘anomaly’ curve in Fig. 1). The point at which C′k reaches its maximum before 30 April is designated as the NRR (black dot, Fig. 1). In the Darwin 1981 example, the NRR occurs on 25 March. Uehling and Misra (2020) begin the summation of day n from 1 September, as they also focus on the onset; however, as our interest is solely on the retreat, we begin our anomaly summation from 1 January. If there is no discernible wet season in any given year and peak C′k is reached at day 1, then this location is represented by a special value. This is more likely to happen in climatologically dry regions (e.g. parts of WA) or regions with highly variable rainfall, e.g. southern Queensland (Qld). This is also why the definition is better suited to the far northern regions of Australia (region highlighted in Fig. 2a) that experience monsoon rainfall, as suggested in Uehling and Misra (2020).

Fig. 2.

Median retreat day over 1950–2019 using the three definitions. Shown are (a) the peak daily cumulative anomaly, (b) the backward accumulation of daily rainfall from 30 April and (c) the last day of the last burst event (25 mm in 3 days) before 30 April. The black contour indicates 1 March (day 60). Retreat dates for nine locations are shown across northern Australia (based on a non-leap year). Red, grey and blue labels indicate the retreat definition that produces the earliest, middle and latest median retreats respectively. Any year for which the retreat day is undetermined is referred to as a non-retreat year. Regions with sparse rainfall coverage or with 30% non-retreat years across the 70-year period are masked out in grey. These mainly represent climatologically dry regions. The area encompassed by the blue dashed line represents the Australia’s monsoonal north region.

2.3. Defining ENSO and the MJO

To separate the NRR dates based on El Niño, La Niña and neutral years, we use the Oceanic NINO index (ONI) for 1950–2019. The ONI is a 3-month running mean of sea surface temperature (SST) anomalies in the NINO3.4 region (5°S–5°N, 170–120°W), calculated from the Extended Reconstructed Sea Surface Temperature dataset (ver. 5, ERSST.v5; Huang et al. 2017). Here El Niño and La Niña events are defined as when the ONI threshold is at least ±0.5°C for four consecutive 3-month seasons from September–November (year (−1)) through to December–February (year (0)). This allows us to associate the NRR dates with ENSO events that peak in austral summer, not in the development stage. This results in 25 El Niño events, of which 21 events also exceed +0.5°C in the January–March season (year (0)). The number of La Niña events selected is 22, of which 2 years (2005–06, 2008–09) do not meet the −0.5°C threshold in September–November, but well exceed the −0.5°C threshold in the January–March season and are included because their timing aligns with the timing of the NRR. As the ±0.5°C ENSO threshold is weaker than The Bureau’s operational threshold of ±0.8°C in the NINO3.4 region, our sample groups of El Niño and La Niña events contain both strong and weaker events, as well as central and eastern Pacific flavours.

The MJO is measured by the Wheeler and Hendon (2004) Real-time Multivariate MJO (RMM) index; an index calculated using daily equatorially averaged (15°S–15°N) Outgoing Longwave Radiation (OLR) (Liebmann and Smith 1996) and NCEP–NCAR zonal winds at 850 and 200 hPa (Kalnay et al. 1996). We use the RMM index from January 1975 through to April 2019 (inclusive), noting that there are missing dates from 17 March to 31 December 1978. The objective of this part of the study is to investigate if NRR dates tend to cluster within the same MJO phase (e.g. when the MJO moves into the western Pacific causing suppressed convection over northern Australia; see Evans et al. 2014). This follows the composite approach of Uehling et al. (2021), who isolated the intraseasonal oscillation using empirical orthogonal functions on the 30–60-day bandwidth and looked at both the onset and retreat of the northern rainy season in each MJO phase. We extend the composite approach for all three NRR definitions.

3.1. Median retreat date

We first compare the 1950–2019 median NRR dates for the three definitions (Fig. 2). Additionally, the dates from nine locations across northern Australia are labelled in each map. As expected, the more northward tropical locations and those dotted along the east coast (i.e. east of the Great Dividing Range) experience a later retreat well into April, whereas inland arid regions have retreats in late January. The masked (grey) areas of northern South Australia, central WA and south-west Qld are regions that do not receive substantial wet season rainfall, with the retreat unable to be determined in at least 30% of years. In general, the cumulative anomaly definition produces earlier retreat dates for the Top End and much of Qld, and the burst definition displays later retreats. For example, Darwin shows a range of 10 days in median retreat dates, with the cumulative anomaly, backward accumulation and burst definitions showing 30 March, 5 April and 9 April respectively (Fig. 2). Further south, in the town of Katherine, the retreat range is 8 days, with the latest date of 24 March (burst definition). Across south-east Qld (e.g. Rockhampton and Roma in Fig. 2), the three retreat definitions diverge – this suggests that, like the NRO, defining a retreat date in coastal subtropical latitudes is difficult as rainfall is not overly seasonal (Drosdowsky and Wheeler 2014).

Interestingly, the pattern of the median NRR dates is not the direct opposite of the median NRO (Cowan et al. 2020) or onset based on the Uehling et al. (2021) definition. Although Cape York and far eastern Arnhem Land have quite late retreats (aside from eastern coastal regions that commonly experience dry season rain), the earliest onsets occur over Darwin and the western Top End (Drosdowsky and Wheeler 2014; Cowan et al. 2020; Uehling et al. 2021). The contrasting spatial retreat pattern could be indicative of the active phase of the MJO passing into the western Pacific – this is explored later in the study. The far west of WA (e.g. Pilbara) have the latest onsets and earliest retreats, as found using other monsoon definitions (Zhang 2010). Further north-east into the Kimberley region, late onsets c. late December–early January are typically followed by late retreats in mid-March, likely a result of a mid-autumn tropical cyclone or depression (e.g. TC Fay in March 2004).³

The strongest agreement between the three definitions occurs north of 20°S; these regions, including the Kimberley, northern NT and northern Qld, all exhibit a strong seasonality in their rainfall and reliable wet seasons (Sharmila and Hendon 2020). Regions south of 20°S either exhibit a higher coefficient of variation (west and central Australia) or do not have a well-defined wet season (e.g. south-east Qld).

For the northern red meat producers, the Green Date marks the date by which a producer would normally have received 50 mm over a 3-day period for their location in 70% of all years. Taking this approach to the NRR, we calculate the long-term 70th and 30th percentiles (for 1950–2019) for each definition, which characterises the latest (70th percentile) and earliest (30th percentile) NRR dates (Fig. 3).

Fig. 3.

Observed (left) 70th and (right) 30th percentile retreat day over 1950–2019 using three definitions. Shown are (a, b) the peak daily cumulative rainfall anomaly, (c, d) the backward daily rainfall accumulation from 30 April and (e, f) the last day of the last burst event (25 mm in 3 days) before 30 April. The black contour in (a, c, e) represents 1 April (day 91), and in (b, d, f) 1 March (day 60). Regions with sparse rainfall coverage or with 30% non-retreat years across the 70-year period are masked out in grey. An expansive list of 70th percentile values at various red meat and sugar locations across northern WA, NT and Qld are shown in Supplementary Tables S1–S3.

An expansive list of individual 70th percentile dates, based on the three definitions for locations across northern Australia that produce red meat or sugar, are shown in Supplementary Tables S1–S3 (for NT, WA and Qld respectively). These locations are taken from over 550 test locations introduced from the Forewarned is Forearmed project (Hudson et al. 2022). The purpose of these dates is to provide a producer with an approximate ‘decision date,’ similar to the Green Date, deriving the timing of the last meaningful rain event in the majority of years, from which dry season planning can begin (i.e. forage budgeting for livestock producers). For a location like Darwin Airport, considering the range of possible retreat dates based on the three NRR definitions, the retreat typically occurs by 9–14 April (70th percentile; Fig. 3a, c, e) but sometimes can occur as early as 24–30 March (30th percentile, Fig. 3b, d, f). These Darwin dates are much later than average estimates of early-mid March, derived from wind (Drosdowsky 1996) and seasonal rainfall total (Smith et al. 2008), but closer to other objective retreat definitions that give early April retreats (Janowiak and Xie 2003; Zeng and Lu 2004). At more southern latitudes in Qld’s Outback town of Cloncurry, the typical retreat happens by 13–19 March, but in 30% of years, can occur in late February. As with the medians, the cumulative anomaly definition generally captures the earliest retreats in the NT and Qld, irrespective of the percentile. Over much of inland eastern Qld, the burst definition gives April retreats in the 70th percentile, in contrast to the early-mid March retreats given by the other two definitions. This suggests that April rainfall events are not of sufficient magnitude to exceed the 50-mm accumulation or overcome the steady decline in the cumulative anomaly, dictated by dry spells between wet events (Fig. 1).

3.2. ENSO and retreat dates

It is well known that ENSO has a strong control on the onset of the wet season rainfall (Smith et al. 2008; Kajikawa et al. 2010; Evans et al. 2014) and the NRO (Lo et al. 2007; Drosdowsky and Wheeler 2014), and this feeds into a stronger or weaker magnitude start to the wet season (Cowan et al. 2020). These studies all utilised the Southern Oscillation Index (SOI) to describe ENSO. Using the NINO3.4 index to describe ENSO, Lisonbee and Ribbe (2021) showed that, in fact, strong La Niña events (NINO3.4 SST anomalies <−0.99°C) are associated with earlier onsets; however, little separated average onsets from weak La Niña events, neutral or El Niño events. For retreat dates based on the cumulative anomaly definition, Uehling et al. (2021) concluded that there was a patchy negative relationship to ENSO (defined as Sep–Nov NINO3.4), with peaks over the Cape York and southern Top End.

Fig. 4 displays the difference between the 1950 and 2019 median NRR and the median NRR in the El Niño and La Niña subsets. Brown shading indicates where the NRR median in ENSO years is earlier than the all-year median. For the three NRR definitions, it appears that there is very little deviation in retreats from the long-term median (±5 days) for El Niño years (Fig. 4a, c, e). Using Darwin as an example, the retreat occurrences for El Niño range between 2 days earlier to 1 day later. Further south in Katherine (south-east of Darwin), the retreat shift is 0–2 days earlier. The largest influence of El Niño is observed across southern NT and southern Qld, with the retreat dates shifting 5–10 days earlier. The El Niño-induced earlier shifts in the retreat are most striking in the cumulative anomaly definition, with the northern Qld regional centre of Charters Towers (Fig. 2) showing a retreat shift to 8 days earlier (Fig. 4a). These shifts represent ~25% of the interannual variability, as defined by the interquartile range. The typical retreat variability is ~20–25 days over northern Australia (Supplementary Fig. S1b, d, f), so the shifts due to El Niño are not overly remarkable. The earlier shifts in the southern regions in the cumulative anomaly definition stem from the fact that this definition is capturing earlier NRR dates, during which ENSO is likely to be stronger (i.e. closer to its peak).

Fig. 4.

Observed difference between the 1950 and 2019 median retreat day and the median for (a, c, e) El Niño years and (b, d, f) La Niña years for each northern rainfall retreat definition. Shown are (a, b) the peak daily cumulative rainfall anomaly, (c, d) the backward daily rainfall accumulation from 30 April and (e, f) the last day of the last burst event (25 mm in 3 days) before 30 April. El Niño and La Niña events cover the January–March period (i.e. demise of event), with events from https://origin.cpc.ncep.noaa.gov/products/analysis_monitoring/ensostuff/ONI_v5.php. The El Niño years include 1952, 1954, 1958–1959, 1964, 1966, 1969–1970, 1973, 1977–1978, 1980, 1983, 1987–1988, 1992, 1995, 1998, 2003, 2005, 2007, 2010, 2015–2016 and 2019. The La Niña years include 1950, 1955–1956, 1965, 1971–1972, 1974–1976, 1984–1985, 1989, 1996, 1999, 2000–2001, 2006, 2008–2009, 2011–2012 and 2018. Regions with sparse rainfall coverage or with 30% non-retreat years across the 70-year period are masked out in grey.

Unsurprisingly, for the La Niña subset, we see later retreats across much of northern Australia (Fig. 4b, d, e). The magnitude of the later retreats is far stronger than the El Niño-related earlier retreats across the northern tropics, suggesting an asymmetric relationship with ENSO (e.g. Chung and Power 2017; Heidemann et al. 2023). Although Uehling et al. (2021) showed that retreat dates (using the cumulative anomaly definition) across northern Australia are negatively correlated with ENSO, we suspect that this relationship is dominated by the strong association with La Niña, like the wet season onset (Lisonbee and Ribbe 2021). Of the three NRR definitions, the backward accumulation definition shows the weakest shifts over northern Australia (north of 20°S) with respect to La Niña, with an average of 7 days later, compared to 8 and 10 days for the burst and cumulative anomaly definitions respectively.

For south-east Qld, La Niña has a weaker control on the retreat dates, whereas in the north-west (e.g. Dampier Downs in the Kimberley region, and Karratha in the Pilbara) retreat dates are pushed back by 8–12 days (~30–50% of the interquartile range). For many regions, the El Niño influence (earlier retreats) is much weaker than that for La Niña (later retreats), meaning there must be a compensation during the neutral ENSO years. Indeed, we see over large swaths of northern Australia, particularly southern Top End and Cape York, that neutral years push the retreat towards earlier dates (Supplementary Fig. S2). A good example of this is in Katherine, which typically experiences a retreat date up to 12 days earlier (burst definition) in neutral ENSO years, compared to 2 days earlier for El Niño and 3–7 days later for La Niña. It is possible that the earlier retreats co-occurring with neutral ENSO conditions are related to intra-seasonal climate variations in the western Pacific in the lead-up to a developing ENSO event. Of the 23 neutral years, nine developed into El Niño events by the austral winter and five events had already surpassed the +0.5°C ONI threshold by the April–June season. Only one neutral ENSO year developed into a La Niña (2017 into 2017–18); however, this event did not surpass the −0.5°C ONI threshold until the austral spring. A lag correlation analysis between the seasonally averaged and linearly detrended ONI and NRR dates, averaged over northern Australia (north of 20°S for all Australian longitudes) suggests that there is a significant negative relationship between the NRR date and the following ONI from austral winter following the retreat (JJA₊₁) through to the next autumn (MAM₊₁) (Supplementary Table S4). A developing ENSO event (from JJA₀) is also correlated with NRR, although the strength of this relationship is weaker in the burst and backward accumulation definitions. Although exploration of this relationship is beyond the scope of this paper, Evans et al. (2014) suggest that the mechanism is related to deep convection over the Maritime Continent and variations in the Walker Circulation. In summary, following the peak of La Niña events, NRR tends to occur later, whereas for El Niño events, NRR typically does not change much from normal in the far north, but occurs earlier in the subtropical east. Neutral ENSO years are associated with earlier NRR dates over the far northern tropics, although the mechanism for this is unclear and requires further exploration.

3.3. MJO and retreat dates

As shown by Uehling et al. (2021), the relationship between retreat date and the MJO phase is not homogeneous across tropical northern Australia, with phases 7 and 8 more frequently associated with retreats for the Gulf region, whereas phase 6 is linked to retreats across the Top End and along Qld’s north-east coast. In general, rainfall bursts across northern Australia that feature a non-southerly vorticity flux prior to the rain event (i.e. tropical origins) are more related to the MJO when in its active phases of 3–6 (moving from the Maritime Continent to northern Australia) (Narsey et al. 2017). To assess the state of the MJO during NRR dates, we first focus on retreats for Darwin, plotting the amplitude and phase of the MJO on the day of the NRR, for all years of 1975–2019 (Fig. 5). For the three NRR definitions when the MJO is strong (amplitude > 1), the MJO is more likely to be in phase 6 when the retreat date occurs. For the cumulative anomaly definition, the retreat dates fall in MJO phase 6 in 20% of years, whereas for the burst definition, it occurs in phase 6 in 24% of years. The cumulative anomaly and burst defined retreats occur in 20 and 26% of years respectively, when combining MJO phases 2 and 3. These phases are when convection is strong in the central Indian Ocean and suppressed conditions exist over the north-north-east of Australia in austral autumn (Cowan et al. 2023). This drops to 13% for the backward accumulation definition. One reason for this disparity might be because the backward accumulation definition does not necessarily require discrete rainfall events to surpass the 50-mm threshold, as opposed to the cumulative anomaly and burst definitions. Any rain event leading to the 50-mm accumulation could be unrelated to the MJO. Interestingly, the retreat dates based on the cumulative anomaly and burst definitions almost never or very rarely fall in MJO phases 8, 1 and 4. During autumn, phases 8 and 1 are associated with suppressed convection (i.e. positive OLR anomalies) over northern Australia (Cowan et al. 2023), so perhaps this is unsurprising. Phase 4, however, is related to strong convection stretching from the Sumatra–Java region, east across the Maritime Continent and over the northern reaches of Australia’s Top End and Cape York. Given the MJO’s strong influence on rainfall over northern Australia in phases 5–7, this infers that the last meaningful rain events of the wet season are likely to occur after phase 4. The fact that most retreats (33–40%) occur when the MJO is weak (amplitude < 1) means that the MJO may not be the dominant forcing, at least in Darwin.

Fig. 5.

Phase of the MJO on the date of the NRR for each year of 1975–2019 for (a) the peak daily cumulative anomaly definition, (b) the backwards accumulation of daily rainfall from 30 April and (c) the last day of the last burst event (25 mm in 3 days) before 30 April. Blue and red coloured dots represent March and April dates respectively.

The same analysis is applied to each of the nine northern Australian locations that feature in Fig. 2, with the percentage of retreat dates (for each NRR definition) that fall into each MJO phase shown in Fig. 6. Like Darwin, the locations of Katherine, Normanton, Cloncurry and Charters Towers all show MJO phase 6 as the dominant phase, although the percentages vary between the definition dates. The percentages associated with the dominant phase are also generally weaker at those locations compared to Darwin, and there is a high degree of spread across all phases for Cloncurry and Charters Towers. In the Gulf town of Normanton, MJO phase 8 is associated with a retreat in 10% of years, approximately half of what Uehling et al. (2021) found. For Dampier Downs in WA’s Kimberley region, as well as Roma in Qld’s south-east, the retreat date is more frequently linked with MJO phase 7, occurring 15–20% of the time, although phases 6 and 7 feature in 13% of years for Dampier Downs. Further south-west, at Karratha, it is clear that MJO phase 6 is when the retreats are more likely; however, the backward accumulation definition displays more frequent dates in phase 5. This is likely because this definition captures more retreats in February (for Karratha) than for the other two definitions. When in MJO phase 5, during January–March, convection is the strongest over Karratha,⁴ so it is likely the backward accumulation definition is capturing the last substantial rainfall event of the wet season. By contrast, the cumulative anomaly and burst definitions show retreats in March and April for Karratha. Some locations, like Rockhampton, on Qld’s far east coast, do not show any strong association with the MJO. Indeed, what is not shown in Fig. 6 is that the majority of retreats are most associated with weak MJO activity (amplitude < 1), at close to 40% for Darwin, Charters Towers and Rockhampton. Most other locations show retreats occurring ~30% of the time when the MJO is weak, similar to other studies (Evans et al. 2014). From these results, we conclude that MJO phases 6 and 7 are associated with retreats during February–April; however, retreats can happen in nearly all phases, and the fact that these occur most often when the MJO is weak infers that other stochastic processes are more often responsible for the retreat of the rainy season.

Fig. 6.

Frequency of northern rainfall retreat that falls into each phase of the MJO, excluding the weak phase (when MJO < 1) for the nine northern Australian locations for 1975–2019. The percentages are calculated by the number of years the retreat occurs in a particular phase, divided by the total number of years.

3.4. Northern rainfall retreat in ACCESS-S2

To answer the question of whether it is possible to forecast the NRR, we first need to evaluate the model biases and skill in the ACCESS-S2 hindcasts. In general, there have not been any great improvements in forecasting late summer to early autumn rainfall for northern Australia in ACCESS-S2 over its predecessor ACCESS-S1 (Wedd et al. 2022). In terms of short multi-day rainfall bursts, ACCESS-S2 shows a degeneration in prediction skill over most of the tropical north beyond a lead time of 3 weeks (Cowan et al. 2022a, 2022b). Despite this, we utilise 27 hindcasts from ACCESS-S2 for each year from 1981–82 to 2018–19 (noting that an ensemble of only 24 hindcasts is used for 2018, as explained in Section 2). The hindcast start dates cover each day from 24 December to 1 January (year + 1) for three different ensemble members. For each ensemble member and start date, we determine the NRR using each definition and then calculate the long-term model median, which we compare to the observed NRR median for 1982–2019. Fig. 7 displays the biases (i.e. difference between the model and observations) and Brier skill scores in ACCESS-S2, the latter smoothed 10× with a Gaussian (moving average low-pass) filter to remove noise. Starting with the biases, we see that ACCESS-S2 performs well at capturing the retreat dates over the far northern reaches, within ±6 days. The biases are greatest for the cumulative anomaly definition and of higher magnitude in southern Qld and the NT.

Fig. 7.

ACCESS-S2 biases and skill in median retreat date over 1982–2019 using the three northern rainfall retreat definitions. Shown are (left) the biases and (right) Brier skill scores across 38 years and 27 ensemble members for (a, b) the cumulative anomaly definition, (c, d) backward cumulative definition and (e, f) burst definition. The ACCESS-S2 biases are calculated as the difference between ACCESS-S2 and AGCD. The 27 members consist of three ensemble members that include 1 January (except 2019) and the eight prior days of the previous month to 24 December. The Brier skill score shows the percentage improvement over the climatological forecast, using the cross-validation method with the patterns smoothed 10× using a Gaussian smoothing function. Stippling indicates where the Brier skill score is above +10%, usually considered significant. Regions with sparse rainfall coverage or with 30% non-retreat years are masked out in grey.

Taking a broad view, ACCESS-S2 simulates positive (i.e. later retreat) biases in both the backward accumulation and burst definitions, and given the similarities in their bias patterns, this suggests that the two definitions capture the same retreat date in both the hindcasts and observations. As the model medians are based on almost 1024 simulations the biases are small in comparison to the observed deviations seen during ENSO events. The skill of ACCESS-S2 to predict the retreat dates is good across the Kimberley, Cape York and patches of the Top End for the cumulative anomaly and backwards accumulation definitions (Fig. 7b, d), with all these regions showing a large areal extent above the 10% improvement over a climatological forecast (i.e. 50% chance of either a later or earlier retreat). Given the retreats for these regions occur between early and late March, this infers that ACCESS-S2 shows reasonable skill out to ~2.5 months from a late December forecast. The best skill is seen for the cumulative anomaly definition, whereas the burst definition has the poorest skill. This possibly reflects the fact the cumulative anomaly retreats are earlier than the burst retreats and are less susceptible to timing and magnitude errors, whereas the burst definition needs to reach at least 25 mm and it must be over consecutive days, making it more difficult to forecast. These results suggest that there could be some utility in a forecast of the NRR across the far northern tropics; however, further south, given ACCESS-S2 skill is poor (or patchy at best) and biases are relatively large, forecasting the retreat is more difficult.