By continuing to browse this site you agree to us using cookies as described in About Cookies
Notice: Wiley Online Library will be unavailable on Saturday 7th Oct from 03.00 EDT / 08:00 BST / 12:30 IST / 15.00 SGT to 08.00 EDT / 13.00 BST / 17:30 IST / 20.00 SGT and Sunday 8th Oct from 03.00 EDT / 08:00 BST / 12:30 IST / 15.00 SGT to 06.00 EDT / 11.00 BST / 15:30 IST / 18.00 SGT for essential maintenance. Apologies for the inconvenience.
Martin Todd, Department of Geography, University College London (UCL), 26 Bedford Way, London WC1H 0AP, UK (fax + 44 20 76794293; e-mail email@example.com).
1The brown locust Locustana pardalina is a major agricultural pest in southern Africa, with populations periodically reaching plague proportions. Management and control would benefit from a predictive capacity at seasonal time scales, as yet unavailable.
2The results of a study into the dynamics and potential predictability of locust populations in southern Africa are presented here. The number of districts reporting locust control measures was used as a proxy for swarming brown locust populations.
3Spectral analysis of the annual number of brown locust infestations over southern Africa revealed dominant periodicity at 17·3 years. The data were low-pass filtered and the low-frequency and high-frequency components were retained. The low-frequency component led the observed 18-year cycle in southern African rainfall by about 3 years, and was therefore likely to reflect endogenous controls on populations.
4Variability in the interannual high-frequency component of brown locust infestations was strongly related to rainfall over the Karoo and Eastern Cape regions of South Africa. The highest correlations were with rainfall over the 12 months prior to the locust season (r = 0·64) and in particular with rainfall during December (r = 0·55).
5Evidence is presented that the high-frequency component is related to the Pacific El Niño/Southern Oscillation (ENSO) and that high-frequency locust activity is abnormally high (low) during La Niña (El Niño) events.
6The high-frequency component of locust activity correlates positively and negatively, respectively, with sea-surface temperatures over the tropical western and eastern Pacific Ocean many months in advance of the locust season. Activity also correlates positively (negatively) with sea-surface temperatures over the south-west Indian Ocean and the Southern Ocean (west and north-west Indian Ocean). These relationships occur later than those in the Pacific, developing in the austral winter and peaking in early summer. This pattern of correlations and the associated atmospheric circulation anomalies is consistent with ENSO-related and non-ENSO related patterns of climate variability.
7The results suggest that there may be considerable scope for future development of models for the seasonal prediction of brown locust activity in which high-frequency variability is related to climatic indices.
The brown locust Locustana pardalina (Walker 1870) is a major agricultural pest in much of southern Africa (SA), south of about 20°S. It has been described as the most important agricultural pest in South Africa (Lea 1953), where its main outbreak areas are in the semi-arid Karoo region (Botha 1969; Centre for Overseas Pest Research 1982). Population fluctuations can be dramatic, with plagues spreading from the source region into neighbouring Namibia, Botswana and Zimbabwe. Despite a long history of research into the dynamics of population variability, there remain no definite predictions sufficiently far in advance to plan anti-locust campaigns. This has resulted in the application of insecticides over large areas during outbreaks of the swarming phase (Nailand & Hanrahan 1993). In this context there are potentially significant benefits to understanding the controls on population dynamics with a view to developing a predictive capacity.
The brown locust life cycle is well understood (Price 1988; Nailand & Hanrahan 1993). Egg hatching is a complex process involving quiescence and diapause but is stimulated by rainfall (Matthée 1951). When hatching is successful and widespread, with the resultant nymphs at high enough population densities, the insects change phase from the solitary to the gregarious condition and occur in swarms. Swarming adults congregate at oviposition sites covering up to 100 ha in the outbreak areas, where egg pods are laid in loose, dry soil often shaded by small Karoo bushes. The eggs occur in two forms: (i) those that hatch within 10–20 days given adequate moisture, e.g. after 15–25 mm of rain (Smit 1939) or, if conditions are unsuitable, become quiescent and hatch after some months; and (ii) those that enter diapause for 1–3 or more months. Both kinds of egg may be present in the same pod and are drought resistant. The ability of brown locust eggs to become quiescent or to enter diapause contrasts with the lack of such adaptations in the eggs of the desert locust Schistocerca gregaria (Forsk), but is similar to the condition found in the Senegalese grasshopper Oedaleus senegalensis (Krauss), a major pest in the Sahel region of West Africa (Fishpool & Cheke 1983; Cheke 1990). The brown locust has five nymphal instars or occasionally four in males. The hopper period of solitary locusts lasts 21–38 days, and at least 42 days for the gregarious phase. In the latter phase hoppers tend to be larger than in the former (Centre for Overseas Pest Research 1982). Years of high rainfall can produce three generations in one season (September–April) and four in a year. Under drought conditions the eggs can remain dormant for up to 15 months (Matthée 1951). As such the brown locust is extremely well adapted to the highly variable climate of the Karoo region. The relationship between locust populations and climate has long been noted. Early work of Du Plessis (1938), Smit (1941) and Lea (1958, 1968) noted a correlation with rainfall, and comprehensive recent studies by Steedman (1990) and Nailand & Hanrahan (1993) noted a positive (negative) correlation between brown locust swarming during high summer and the early summer (previous winter) rainfall.
The region of SA of interest to this study exhibits a pronounced zonal climate gradient, with arid conditions in the west and humid conditions in the east. Over much of north-eastern SA there are distinct wet (summer) and dry (winter) seasons associated with the annual cycle in the meridianal position of the Intertropical Convergence Zone (ITCZ). In the Karoo and Eastern Cape regions rainfall displays a bimodal distribution, with peaks in the transition seasons of spring and autumn (Tyson 1986). The western parts of the Northern and Western Cape provinces experience a winter rainfall maximum. The climate of SA is known to show pronounced variability at a range of time scales from intraseasonal (Todd & Washington 1999), through interannual (Jury 1997; Nicholson & Kim 1997; Rocha & Simmonds 1997) to decadal and multidecadal (Tyson 1986; Folland et al. 1999). Interannual variability is particularly high in the drier region, including the Karoo where the coefficient of variation exceeds 40% (Tyson 1986). For a review of rainfall variability in SA see Mason & Jury (1997).
That the climate system behaves as a coupled ocean/atmosphere system, through the exchange of energy, mass and momentum, is a dominant paradigm in contemporary climatology. The behaviour of the atmosphere is dependent on the ocean, and vice versa. El Niño/Southern Oscillation (ENSO), the dominant mode of global interannual climate variability, exerts considerable influence over SA rainfall during the austral summer and is responsible for modulating the extreme dry and wet years. A number of studies have documented the development of positive (negative) sea-surface temperature (SST) anomalies in the equatorial and northern (southern) Indian Ocean during ENSO warm/El Niño events. During ENSO cold/La Niña events a reversal of this pattern is observed. It has been hypothesized that these SST anomalies modulate the large-scale structure of the atmosphere. This occurs through adjustments to the zonal structure of regions of atmospheric convergence and divergence in the tropics (the Walker circulation), and thus the large-scale convergence of low-level moisture (necessary for rainfall) in the SA region (Goddard & Graham 1999). During ENSO warm events (El Niño) the rising limb of the African Walker cell and associated rainfall is displaced eastward into the Indian Ocean, resulting in anomalously dry conditions over SA (Tyson 1986; Mason & Jury 1997; Goddard & Graham 1999; Reason et al. 2000). In addition, Rocha & Simmonds (1997) and Preston, Washington & Todd (2000) demonstrate the importance to SA rainfall of ocean/atmosphere variability in the Indian Ocean region that is independent of ENSO.
The coupling of the ocean and atmosphere provides the physical basis for seasonal forecasting of climate anomalies up to several months in advance (Murphy et al. 2001). Seasonal forecasting is based on notions that (i) the lower boundary forcing of the atmosphere, most notably the state of the ocean (commonly represented by SST), evolves relatively slowly and as such is predictable (because SST have a significant degree of persistence from one month to the next), and (ii) the atmosphere responds in a predictable manner to this component of forcing. Therefore, SST in one season can have an impact on the atmosphere in subsequent seasons, thereby providing the basis for probabilistic seasonal predictions of the behaviour of the atmosphere (often rainfall and/or temperature). Remote SST anomalies, for example in the eastern tropical Pacific associated with ENSO, may also take several months to affect the atmosphere over SA, providing the lead times needed for predictability. A number of studies have recently demonstrated the potential for seasonal forecasting of climate or climate-related variables over SA, based at least in part on SST in the remote Indian and Pacific Oceans (Thiaw, Barnston & Kumar 1999; Washington & Downing 1999; Martin, Washington & Downing 2000).
In this context, the primary aim of this study was to further our understanding of the dynamics and predictability of brown locust populations in SA, through analysis of a long-term data set of reported locust outbreaks. Specifically, the following questions were addressed. (i) What is the dominant periodicity of locust outbreaks? (ii) To what extent is variability in locust outbreaks related to exogenous climatic factors? (iii) To what extent is the variability in locust outbreak intensity over SA predictable from climatic indicators?
Materials and methods
One of the major problems facing research into the behaviour of the brown locust is a lack of direct population estimates over the extended areas and time periods necessary to develop generalized population models. Accordingly, we were limited to indirect estimates or indices, for which we used the number of magisterial districts (D) in South Africa, Botswana and Namibia in which brown locust control activity took place in a given year, over the period 1947–98 (Price & Brown 2000; M. E. Kieser, personal communication; see also Milton, Davies & Kerley 1999 for patterns since 1797). Whilst this is a coarse index of actual locust numbers it is a useful indicator of the year-to-year variability, despite the potential effects of control measures on population numbers. Similar data sets have been used for the brown locust (Nailand & Hanrahan 1993) and the desert locust (Cheke & Holt 1993). Although the data do not describe the precise timing of brown locust infestations, observations suggest that swarms tend to occur in the mid- to late wet season (January–March) after populations have grown sufficiently (Steedman 1990; Kellner & Booysen 1999). Hereafter, we use the term ‘locust season’ to refer to this early part of the calendar year.
Periodicity of brown locust outbreaks was analysed by singular spectral analysis (SSA) of the locust data. On the basis of the SSA results, the low-frequency (LF) component of brown locust variability (with periodicities > 11 years; DLF) was separated from the high-frequency (HF) component (with periodicities < 11 years; DHF), by means of an integrated random walk Kalman filter. The Kalman filter (described fully in Young et al. 1991; Chatfield 1992) fits a smooth line through a time series and is known to be less vulnerable to large swings resulting from outliers in the observations than many simpler methods. The filtered data represent the LF component of variability in D and can be considered to represent variability at decadal time scales. The LF component was then subtracted from the raw time series to leave the HF component (DHF), indicative of variability at interannual time scales.
Information on monthly rainfall (R) and near surface air temperature over the same period was obtained from Hulme (1992) and Jones et al. (1999), respectively, which provided observations over global land areas on a grid at 2·5° latitude by 3·75° longitude and 5 × 5° resolution, respectively. Monthly anomalies of SST in the Niño-3·4 region of the central equatorial Pacific (5°N–5°S, 170–120°W) were used as an index of the state of the ENSO system. Gridded global fields of key atmospheric and surface climatic variables (SST, low-level winds and sea-level pressure), indicative of ocean boundary forcing and the atmospheric circulation, were obtained from the National Center for Environmental Prediction (NCEP) Reanalysis data set (Kalnay et al. 1996). This provides monthly data on a 2·5° global grid for the period 1948–present.
The relationship between DLF and low-frequency climate variability was assessed by comparison of the time series of DLF and low-frequency rainfall variability (RLF) over the SA region. We used rainfall time series (RLF) for early (OND) and late (JFM) summer. These were the eigenvector time coefficients of the leading empirical orthogonal functions (EOF) of seasonal rainfall. The EOF were selected on the basis that they have loadings over the central interior of SA (Washington 1998). In addition, we used the OND RLF over the Karoo (grid cell centred on 32·5°S, 26·25°E).
To assess the influence of climate on the high-frequency component of brown locust variability, the relationship of DHF and climate variables (R, temperature, SST, sea-level pressure and low-level winds) was analysed by means of correlation and composite analysis. In the former case the time series DHF was correlated with the time series of climate variables at each grid cell in the global field. To assess predictability of DHF from climate, the DHF time series was lagged by a number of months. The aim of composite analysis was to identify the characteristic structure of the ocean and atmosphere associated with the major years of high and low brown locust activity. The DHF data were ranked and the five most extreme years of high and low locust activity were identified. Mean anomalies of NCEP climate data were then calculated for these samples at each grid cell and statistical significance was tested using a t-test.
To evaluate the predictive strength of the relationship between brown locust variability and rainfall, estimates of DHF were derived from a linear regression with preceding annual rainfall at certain grid cells over SA. The skill of these ‘hindcasts’ was tested through a ‘jack-knife’ procedure, involving 52 regression analyses. In each case a single year’s data were omitted and its DHF value predicted from the regression of DHF and annual rainfall derived from the remaining data. Given the absence of serial autocorrelation in the data (see the Results) this ensured that hindcasts were made using independent data. The accuracy of the predicted DHF relative to the observed DHF was compared using the correlation coefficient, mean bias and root mean squared error (RMSE). The estimates were also compared in terms of broad categories using the Heidke skill score (Wilks 1995) and linear error in probability space (LEPS) score (Potts et al. 1996). Three categories were selected (above normal, normal and below normal), defined by the appropriate tercile values of the DHF distribution. The Heidke and LEPS scores defined the percentage improvement in the accuracy of estimate classification into these three categories over a reference strategy with little ‘skill’, such as random guessing or climatological persistence. The LEPS included a weighting to account for the magnitude of errors between class boundaries.
periodicity of brown locust outbreaks
Singular spectral analysis of the raw number of districts reporting brown locust control (D) revealed a dominant peak at 17·3 years, with lesser peaks at 3·7, 2·9, 10·4 and 7·4 years in decreasing order of importance (Figs 1 and 2). The peaks at 17·3, 3·7 and 2·9 years in the power spectrum were statistically significant at the 0·05 level or higher, based on the Bartlett–Kolmogorov–Smirnov test where the null hypothesis maintains that the time series was white noise (Fuller 1976). Kalman filtering of the data indicated that the variance of the high-frequency component of D was approximately twice that of the low-frequency component. The new derived time series of DLF and DHF are shown in Fig. 1. Analysis of the autocorrelation function of DHF at various lags from 1 to 20 years revealed no statistically significant correlation (data not shown).
Kalman filtering of the SA rainfall data [OND and JFM EOF1 from Washington (1998) and OND rainfall at grid cell centred on 32·5°S, 26·25°E] showed multidecadal variability. A pronounced 18-year cycle occurred in JFM RLF and, although the periodicity of the two OND RLF time series was less clear, there was some evidence of periodicity near 18 years (Fig. 3). However, the DLF cycle led both the JFM and OND RLF cycle by about 3–7 years (Fig. 3). Our confidence in the phase of this low-frequency variability was not large owing to the short data series.
climate variability and high-frequency variability in brown locust infestations
The highest correlations (up to 0·55) between DHF and surface rainfall preceding the brown locust swarming season occurred over a restricted area of the Karoo and Eastern Cape region of South Africa (notably two grid cells centred on 32·5°S, 22·5°E and 32·5°S, 26·25°E) in December (Fig. 4). There were significant correlations over a broader region of SA in October. Correlations of DHF and austral winter rainfall were weak, although locally significant (at the 0·05% level) positive correlations occurred in July over the Western Cape province and in August over the Eastern Cape province. Significant positive correlations also occurred with late summer (January and February) rainfall some 10–12 months in advance of the locust season. As a result, the highest statistical relationship (r = 0·64) was observed between DHF and annual (January–December) rainfall over the Eastern Cape region (32·5°S, 26·25°E). That annual rainfall in the year leading up to the locust season explained a substantial proportion (42%) of variability in DHF provides potential for predictability using linear regression (Table 1). To test the validity of the posterior selection of a target cell at 32·5°S, 26·25°E, the same procedure was conducted using rainfall data at all grid cells surrounding it. For brevity only the ‘best’ and ‘worst’ results from surrounding cells are shown (Table 1).
Table 1. Accuracy assessment of hindcasts (n = 52) of the high-frequency component of brown locust populations estimated from linear regression with preceding annual (January–December) rainfall at individual grid cells over southern Africa, climatological persistence and random guessing
Hindcast estimation method
Linear regression with annual rainfall at cell 32·5°S, 26·25°E
Linear regression with annual rainfall at cell 32·5°S, 22·5°E
Linear regression with annual rainfall at cell 30°S, 30°E
Heidke skill score
The relationship of D and DHF with rainfall at the cell centred on 32·5°S, 26·25°E, indicating that only the HF component of locust populations had a strong relation to rainfall (Fig. 5). Correlation analysis with surface temperature fields over the SA region revealed no statistically significant correlations during any month or season within one year preceding the wet season (data not shown).
seasonal predictability of brown locust outbreaks
The extreme years of HF brown locust activity were 1985–86, 1950–51, 1970–71, 1963–64, 1971–72, of which four corresponded to ENSO ‘cold’ events (La Niña) in the Pacific (on the basis of January SST anomalies in the Niño-3·4 index). Strong SST anomalies occurred throughout the previous year in these cases. The extreme years of low brown locust activity were 1972–73, 1992–93, 1990–91, 1949–50, 1987–88, of which two (1972–73 and 1987–88) corresponded to major ENSO ‘warm’ conditions (El Niño). The events in the 1990s coincided with the prolonged occurrence of moderate El Niño conditions throughout the early 1990s.
Correlations between DHF and the Niño-3·4 index of Pacific SST at various lags (Fig. 6) were statistically significant (at the 0·05% level) for up to 12 months prior to the brown locust season (assumed to occur in mid/late summer). Highest correlations were observed with the Niño-3·4 index during February–May (austral summer/autumn, peaking at –0·43 in February, significant at the 0·01% level) prior to the brown locust season (Fig. 6). The sign of the correlations indicated that ENSO warm (cold) events generally preceded years of below (above) average HF locust irruptions.
Correlation analysis of DHF and gridded SST at various lags (Fig. 7a–i) indicated statistically significant positive (negative) correlations (up to 0·5) in seasons prior to the brown locust season over extensive regions of the tropical western (eastern) Pacific. Negative correlations (up to 0·5) also occurred over the extensive regions of the western Indian Ocean from austral winter onwards (feature A in Fig. 7d–i), associated with positive correlations (up to 0·5) over the south-west Indian Ocean (feature B in Fig. 7d–i). Broadly, there was a north/south dipole in the correlation sign over the north-west/south-west Indian Ocean. This correlation structure in the Pacific and Indian Ocean basins was consistent with a persistent ENSO signal, represented by correlations of the opposite sign in Fig. 7j.
The dipole structure of negative (positive) correlations over the north-west and central southern (south-west) Indian Ocean (features A and B in Fig. 7d–i) evolved from the austral winter season and peaked in strength during early summer (December; Fig. 7i). An arc of negative correlations extended from the north-west Indian Ocean to the subtropical southern Indian Ocean from winter onwards (feature A) and the highest correlations moved southwards to lie at 30°S, 55°E in December. Positive correlations (feature B) propagated westward from the south-west Indian Ocean, with the highest correlations in this region (up to 0·5) located immediately south of SA in the Southern Ocean (at 40°S) during December. An index of SST over this region (37·5–42·5°S, 11·5–22·5°E) for the OND season had a correlation of 0·5 with DHF. Throughout the austral early summer period there were negative correlations between DHF and SST over the subtropical south Atlantic, centred on 20°W, 35°S (feature C in Fig. 7g–i). The correlations with SST in the Southern Ocean (south of SA), the south-west Indian Ocean and south-east Atlantic were higher for locust activity than for the Niño-3·4 index (Fig. 7j), suggesting that in these regions the observed SST structure related to DHF may not be entirely ENSO related.
In December prior to high DHF events, associated with La Niña conditions, an anomalous continental low was located over SA (Fig. 8a). An anomalous SLP high was centred over the south-west Indian Ocean near 50°S, 50°E (Fig. 8a). These features led to anomalous low-level easterlies from the subtropical south-west Indian Ocean peaking at 40°S (Fig. 8b). During low DHF events, associated with El Niño conditions, these anomalies were reversed and the moist easterlies were weakened over SA. The inference that such mechanisms are directly related to rainfall is supported by the close similarity with the correlation structure between these fields and December rainfall over the Karoo and Eastern Cape region (31–33°S, 22–26°E) (data not shown).
Understanding the nature of brown locust populations over SA so that effective control measures can be implemented, could potentially result in considerable benefits to agriculture but has remained a problem. First, it is likely that the insect’s capacity for rapid population growth represents the interaction of both endogenous and exogenous factors. Secondly, the development of mathematical population models is difficult because of the locust’s phase change, which can result in locusts being barely noticeable as solitary populations in one generation and then gregarious swarms in the next. Finally, there is a lack of quantitative field data of actual population numbers. In comparison to the desert locust, the population dynamics of the brown locusts have received relatively little attention despite the potential benefits.
This study has focused on analysing the nature of brown locust populations and the possible exogenous control exerted by climate. From this we were able to assess the potential predictability of populations of brown locust, on the basis of the evolution of the climate system. The data set used was a proxy index of brown locust populations (D) in which the precise nature of the relationship to actual locust numbers cannot be specified, although we assume that the data were indicative of late austral summer swarming populations.
The dominant 17·3-year cycle (Fig. 2) in brown locust populations is substantially longer than those identified previously (Lounsbury 1915; Lea 1968, 1972) but close to the 16-year cycle identified by Cheke & Holt (1993) for the desert locust in West Africa. A key question is what drives this LF periodicity. There is evidence that the climate of SA experiences decadal variability dominated by an 18-year periodicity (Mason & Jury 1997), possibly related to global low-frequency SST anomalies (Washington 1998; Folland et al. 1999). However, as the LF component of the D time series leads that of SA rainfall by about 3–7 years (Fig. 3), it is unlikely that LF variability in brown locust populations results from decadal variability in rainfall. Although it is possible that other climate variables and/or interaction with pathogens may be involved, it is likely that the observed LF variability in brown locust populations may be an expression of endogenous controls. In any case, the strong LF cyclicity in brown locust outbreaks suggests that about one-third of the total variance may be predicted on the basis of a 17·3-year oscillation.
Working with logistic equations governing population growth rates, May (1974, 1976) suggested that for populations with particular intrinsic rates of generational population increase (r, where 2·685 < r < 2·692) population numbers can exhibit stable cyclic behaviour with ‘period doubling’. The spectral peaks of D (Fig. 2) show little evidence of this, perhaps indicating that the brown locust has chaotic ‘boom and bust’ population dynamics, characteristic of higher growth rates and determined by endogenous factors (May 1974, 1976). Further research is therefore required into the precise cause of the low-frequency cyclicity in brown locust outbreaks.
The dominant proportion of total variance of D is contained in the HF component and is of primary interest in terms of interannual variability and predictability of locust populations. There is little temporal autocorrelation in the HF component (data not shown). This is suggestive of ‘boom and bust’ dynamics, although our subsequent analysis suggests that there is substantial exogenous control of brown locust population numbers. The absence of serial autocorrelation is in contrast to the desert locust over West Africa, where positive autocorrelation at 1 year is significant (Cheke & Holt 1993). It is also possible that populations are dependent on some other unidentified precedent population characteristic. Price (1988) suggests that swarms arise after the build-up of the solitary phase in the previous year. Unfortunately, our data are best seen as an index of the swarming populations and thus do not support investigation of this hypothesis.
The raw brown locust data exhibit only a weak relationship with rainfall, characterized by heteroscedasity (Fig. 5a). Cheke & Holt (1993) observed a similarly heteroscedastic relationship between rainfall and desert locust populations in West Africa, and found that simulations of populations using a logistic model with high growth rates (characteristic of chaotic dynamics related to endogenous factors) revealed similar patterns. For brown locusts, our results indicate that much of the scatter in the raw data/rainfall relationship can be removed by separating the variability at low-frequency (decadal) time scales from the high-frequency (or interannual) component, and treating the latter separately (Fig. 5b).
High-frequency brown locust variability is most strongly associated with December rainfall over the Karoo region and to a lesser extent the Eastern Cape region (Fig. 4). This confirms that brown locust outbreaks in the wider SA region originate from a relatively restricted source region where locusts are known to breed, and that this process is most sensitive to rainfall in the early summer period, particularly December rainfall. In addition, we observe significant correlations of DHF with rainfall over the same region during the previous late wet season. As such, a substantial proportion (49%) of DHF variance can be explained from annual rainfall prior to the locust season.
We find no evidence of a connection between brown locust irruptions and previous austral winter rainfall (Nailand & Hanrahan 1993; Kellner & Booysen 1999) nor temperatures (Kellner & Booysen 1999) over SA. Much of the interannual variability unrelated to rainfall may therefore be endogenous. It is also important to note that there is evidence that locust population breeding regions can change over time, possibly as a result of changes in local vegetation (L.J. Rosenberg, personal communication). As our study is based on a long-term data set, the results may reflect historical conditions rather than those in the present day, at least in regions where ecological changes have been pronounced.
That the high-frequency component of the number of districts reporting brown locust control (DHF) exhibits a strong correlation with both annual and, in particular, preceding December rainfall over a relatively small region may indicate that there is scope for developing a predictive capacity at interannual time scales. First, monitoring of rainfall in real time may facilitate short lead-time predictability of likely irruption rates in the remainder of the wet season following December, using a simple linear regression. The results (Table 1) show that hindcasts based on regression of DHF and rainfall over the Karoo (32·5°S, 26·25°E) are accurate relative to (i) climatological persistence or random guessing and (ii) hindcasts based on rainfall in neighbouring grid cells (highlighting the importance of this region of the Karoo). In practice, such predictions may facilitate more efficient planning, preparation and resource allocation for subsequent locust control.
In addition, forecasts with longer seasonal lead times may be possible. There is substantial evidence that the high-frequency component of brown locust populations is abnormally high (low) during La Niña (El Niño) phases of the Pacific ENSO system. This is consistent with the documented relationship between ENSO and SA rainfall. In addition, the spectral peaks at 3·7 and 2·9 years identified from spectral analysis of the raw locust data are within the interannual component of the ENSO signal (Allan 2000). High-frequency locust variability shows significant associations with SST over extensive regions of the Pacific and Indian Oceans in the seasons prior to the locust plague season (Fig. 7a–i). In the tropical Pacific there is a clear east/west dipole of negative/positive DHF/SST correlations representing the major centres of action of ENSO. In addition, SST anomalies develop in the Indian and southern Atlantic Oceans some months later than those in the Pacific. This space/time structure is highly characteristic of ENSO-related variability in the major ocean basins (Fig. 7j). A north/south dipole in correlations over the western Indian Ocean similar to that observed here, during the JAS and OND season (lagging the peak SST anomalies in the Pacific), has been noted in composites of major ENSO events (Nicholson & Kim 1997; Reason et al. 2000). Given that the SST structure over much of the Indian Ocean lags that in the Pacific, there is scope to develop a statistically based prediction of the former on the basis of canonical correlations (Goddard & Graham 1999).
An important question is whether the evolving SST structure in the south-west Indian Ocean (Fig. 7a–i) is typical of that associated with ENSO. That correlations with SST in the south-west Indian Ocean, the Southern Ocean (south of SA) and south-east Atlantic (Features A, B and C, respectively, in Fig. 7i) are notably higher for locust activity than for the Niño-3·4 index (Fig. 7j) suggests that locust activity may be related to ENSO but that the specific structure of SST in the oceans immediately surrounding SA may also be crucial in determining the climate and response of brown locusts. There is growing evidence of patterns of Indian Ocean SST that are independent of ENSO (Rocha & Simmonds 1997; Preston, Washington & Todd 2000).
Although it is beyond the scope of this paper to establish the physical mechanisms by which the evolution of the ocean thermal structure influences SA climate, it is notable that early summer sea-level pressure anomalies associated with DHF extremes (Fig. 8a) resemble characteristics of ENSO-related modulation of the atmospheric Walker circulation noted by Reason et al. (2000). The composite mean low level (850 hPa) wind anomaly field associated with extreme DHF years (Fig. 8b) is broadly consistent with both observed surface wind anomalies associated with La Niña events (Reason et al. 2000) and with general circulation model simulations of the effect of ENSO-related Indian Ocean SST anomalies (Goddard & Graham 1999). Thus, during La Niña events and periods of high DHF activity, anomalous easterlies flow into SA from the south-west Indian Ocean (the dominant moisture source for SA) advecting large quantities of moisture over SA, facilitating the development of convective rainfall systems (Tyson 1986). Thus, in accordance with previous work, our results indicate that it is the combination of SST and atmospheric circulation anomalies that dictates the nature of climate anomalies in the SA region, to which there appears to be a consistent response in the HF component of locust infestations.
In this study we have identified that an index of annual brown locust infestations over SA consists of a low-frequency component, possibly controlled by endogenous factors, and a high-frequency component, strongly related to rainfall in the Karoo (and Eastern Cape) regions. About one-third of the total variance can be represented by the 17-year cycle, while much of the remaining high-frequency variability can be related to indices of the evolution of the large-scale climate system. As such, there appears to be considerable scope for developing statistical models for seasonal prediction of brown locust activity many months in advance. Such forecasts may be useful to optimize resource allocation and preparation for locust control activities. The key predictor indices are likely to be SST in the tropical Pacific and western Indian Oceans, the south-west Indian Ocean and the Southern Ocean immediately south of SA, indicative of both ENSO and non-ENSO modes of variability. December rainfall over the Karoo region is an important control on locust populations and thus may provide a valuable late ‘check’ on the likely accuracy of any seasonal forecasts. Seasonal forecasting of the high-frequency component of brown locust infestations (rather than climate variables as in previous work) would certainly represent a novel development in this field.
The authors are grateful to the UCL and the University of Oxford for support. NCEP reanalysis data were obtained from the National Centre for Atmospheric Research. The Niño-3·4 time series of Pacific Ocean SST anomalies was obtained from the NOAA Climate Prediction Center http://www.cpc.ncep.noaa.gov/data/indices. R.A. Cheke is also grateful for support from programme development funds (NRI project ZA0394) of the Crop Protection Programme of the UK Department for International Development (DFID) for the benefit of developing countries. The views expressed are not necessarily those of DFID. Thanks also to Dr Jane Rosenberg of the NRI for helpful comments.