Multi-scales and multi-sites analyses of the role of rainfall in cotton yields in West Africa

Transcription

1 INTERNATIONAL JOURNAL OF CLIMATOLOGY Int. J. Climatol. 30: (2010) Published online 5 March 2009 in Wiley InterScience ( Multi-scales and multi-sites analyses of the role of rainfall in cotton yields in West Africa Benjamin Sultan, a * Marthe Bella-Medjo, a Alexis Berg, a Philippe Quirion b and Serge Janicot a a IRD, LOCEAN/IPSL, Université Pierre et Marie Curie, Case 100, 4 Place Jussieu, Paris cedex 05, France b CNRS, CIRED, 45 bis avenue de la Belle Gabrielle, Nogent sur Marne, France ABSTRACT: Cotton is the main tradable crop of West and Central African countries, representing for some countries the main exported agricultural product. Cotton is then of major concern since it represents an important source of income, accounting for more than a tenth of total exports. Moreover, the subsector as a whole is essential for rural poverty reduction. Since cotton is a rainfed crop in these countries, its yield is closely related to climate, in particular to rainfall variability. The objective of this study is to point out the role of rainfall variability in cotton yields. Our approach consists in taking two completely different sites in the analysis of the climate-yields relationships, i.e. an experimental plot in Mali with a long-term historical yield-survey and farmers yields in 28 administrative units in Cameroon. We found that the same rainfall parameters (rainy season onset and length) are major drivers for the year-to-year and the spatial distribution of cotton productivity, even if the role of rainfall variability is strongly reduced in farmers exploitations where other non-climatic factors such as human management, biotic stresses, pests, etc., impact crop productivity. The link between rainfall and cotton yields seems to depend on the mean climate since the driest cotton areas in Cameroon are the most sensitive to climate variability. The coherence of the results from the two very different situations gives us some confidence in the generalization of our findings to the whole West and Central Africa. Our study shows also that the aggregation of yield data from the local scale to the national scale tends to smooth the non-climatic variability and highlight the role of climate in the year-to-year variability of cotton yields. These results are important in working towards the predictability of crop yields using rainfall information, in particular to highlight the most salient rainfall parameters that are needed in a forecast system. Copyright 2009 Royal Meteorological Society KEY WORDS impacts; cotton; West Africa; Climate Received 22 January 2008; Revised 18 November 2008; Accepted 12 January Introduction West Africa is the world s third largest exporter behind the USA and Central Asia (SWAC/OECD, 2005). Cotton production thus represents an important source of income for West African countries. It also benefits to food and livestock production as African farmers often associate cotton production with the cultivation of grain crops, such as millet, sorghum and maize and this second crop benefits from the fertilizers, materials and animals financed by cotton incomes (SWAC/OECD, 2005). However, if the West African cotton production has considerably increased since the 1960s, the cotton yield experienced stagnation since the 1980s with a decrease in some countries such as Cameroon [The Food and Agriculture Organization of the United Nations (FAO) statistics: The cotton production increase, observed since the 1980s, is thus only due to the increase of the cultivated areas. On the opposite, at the world level, cotton yield has risen continuously since the 1960s. An improvement of the cotton yield would be * Correspondence to: Benjamin Sultan, LOCEAN/IPSL, Université Pierre et Marie Curie, Case 100, 4 Place Jussieu, Paris cedex 05, France. benjamin.sultan@locean-ipsl.upmc.fr a more sustainable solution and a better understanding of the determinants of the cotton productivity is first needed to achieve this improvement. The objective of the present study is to contribute to this understanding by focusing on the effects of climate and mainly on rainfall which is one of the determinants of cotton productivity. The reason for this focus is twofold. First, since cotton is mainly produced in rainfed conditions with relatively few resources and family labour, climate should play an important role in the cotton yields variability. Indeed, agriculture is considered as the most weather dependent of all human activities (Oram, 1989; Hansen, 2002) and the impact of climate is amplified where means to control the crop environment are largely unavailable to farmers: irrigation is rarely an option and use of mechanization, fertilizers and other offfarm inputs are low (Ingram et al., 2002). Second, substantial progress has been made in understanding and modelling the global climate (Palmer et al., 2004) to provide a degree of predictability of climate fluctuations at a seasonal lead time in many parts of the world (Hansen, 2002). There are also increasingly more proofs of climate change based on observed increase in global surface temperatures during the 20th century Copyright 2009 Royal Meteorological Society

2 MULTI-SCALES AND MULTI-SITES ANALYSES OF THE ROLE OF RAINFALL IN COTTON YIELDS 59 and significant inter-annual climate variability observed in many regions of the globe (Salinger, 1994; Salinger et al., 1997), especially in tropical latitudes, caused by events such as the 1982/1983 and 1997/1998 El Niño and the 1991 Mt. Pinatubo volcanic eruption (WMO, 1995, 1998). Considering the potential benefits of climate predictions to agriculture (Sivakumar et al., 2000; Hansen, 2002) and the impacts of anthropogenic climate change on the Sahel, in particular changing intensity, frequency and duration of rainy events (Trenberth et al., 2003), it appears crucial to orient the research efforts to the linkage between the two fields of research: climatology and agriculture. The objective of this study is to document the links between climate and cotton productivity to point out the main rainfall parameters that might impact cotton yields and are thus needed to be predicted at a seasonal lead time and be looked at in climate change scenarios. The documentation of this link in West Africa has already been done by Albergel et al. (1985) and more recently by Kangah (2004) and Blanc et al. (2008). But most of these studies are site-specific with a heterogeneous nature of the types of farms and the results are difficult to generalise. Our approach consists in taking two completely different sites in the analysis of the climate-yields relationships, e.g. an experimental plot in Mali with a long-term historical yield-survey and farmers yields in 28 administrative units in Cameroon (Figure 1). If, using very simple rainfall parameters, we find that cotton yield is sensitive to the same rainfall parameters over these two different situations, we can thus be confident in the role of climate in cotton productivity and in its extrapolation to the whole West Africa. This approach is a good way to go beyond data heterogeneities and idiosyncratic details in order that only the most general forcing remains. According to the different nature of the data, e.g. a long-term time series in a single plot in Mali and a short-term in multiple administrative units in Cameroon, our study will explore the role of rainfall variability at different space and time scales by looking at the climate footprint in the inter-annual variability of cotton yields, at several embedded aggregation levels and in the spatial distribution of the mean cotton yields. 2. Materials and methods 2.1. Cotton and meteorological data Mali data This study is based on a long-term trial conducted by Section Recherche Cotonnière et Fibre Jutière de l Institut d Économie Rurale (IER/SRCFJ) from 1965 to 1990 on a cultivated plot in the region of Koutiala in Mali (Crétenet et al., 1994; Jardot, 1988; Figure 1). Three kinds of fertilizations were used: organic manure, mineral manure and a combination of both. We will only use the yields obtained under the latter condition since Crétenet et al. (1994) has shown that this combination gives the best productivity for the crop and that the obtained yields can thus be considered as the closest to maximum attainable yield. Under such ideal conditions, one can expect an amplification of the signature of climate since the effect of many other factors in human management is suppressed. Both agronomical (yield and sowing dates) and climate data (a synoptic station is near the plot at N Tarla) over the period (1968, 1972 and 1975 are missing). Figure 2 shows the time series of yield data. Notice that N Tarla is a typical Soudano-sahelian station with a mean annual rainfall of around 770 mm Cameroon data We will also base this study on climate and yield data from Cameroon. In this country, the cotton production is Figure 1. Map of West Africa. The cross depicts the Koutiala region where is located the cotton experimental plot Mali. The shaded area depicts the cotton zone in Cameroon. This figure is available in colour online at

3 60 B. SULTAN ET AL Yield (kg/ha) Time (years) Figure 2. Time series of the yield data at the experimental plot in Mali. The cotton yields (kg/ha) is measured by IER/SRCFJ from 1965 to 1990 on a cultivated plot in the region of Koutiala in Mali (1968, 1972 and 1975 are missing). Figure 3. Map of the cotton zone in Cameroon. Both agronomical (yields and sowing dates) and rainfall data are available for 28 of the 38 administrative units over the period (2002 is missing for every administrative units). Notice that for half of the administrative units only 3 years of rainfall ( ) are available. Two synoptic stations are available at Maoua and Garoua. This figure is available in colour online at supervised by the SODECOTON agency that is involved in from technical support to farmers (providing seeds, fertilizers and advices on the cultural calendars) to the international cotton trade. The crop production area is located in North Cameroon (Figure 1) in a Soudanosahelian area, typically between the 600 mm and the 1200 mm isohyets which corresponds to the limits of the land suitability for rainfed cotton defined by Organisation for Economic Co-operation and Development (OECD) (SWAC/OECD, 2005). Both agronomical (yields and sowing dates) and rainfall data are available for 28 of the 38 administrative units (see map on Figure 3) over the period (2002 is missing for every administrative units). Notice that for half of the administrative

4 MULTI-SCALES AND MULTI-SITES ANALYSES OF THE ROLE OF RAINFALL IN COTTON YIELDS 61 units (Figure 3) only 3 years of rainfall ( ) are available. Two synoptic stations are available in North Cameroun at Maoua and Garoua. The 28 administrative units are part of seven cotton production regions defined by SODECOTON. Notice that for most of the southern administrative units only 3 years of data are available (Figure 3). As each administrative unit includes several plots, there is not a unique sowing date per year and per administrative unit, but we obtained from SODECO- TON the percentage of sowing surface per dekad from mid-may to the end of July Climate indexes Since most of the studies on the impact of climate on cotton yields have shown that the total rainfall amount does not explain a large part of the cotton yields variability (Kangah, 2004; Blanc et al., 2008), one needs to define more accurate rainfall parameters that describe the seasonal and intra-seasonal variability of the monsoon. In addition, the use of process models is also a useful tool to assess yield response to various crop management sequences and environmental conditions Rainfall and other meteorological indexes We first define the onset and the end of the rainy season. Several methods exist for calculating the date of onset of the rains that may be taken as the start of the growing season (Ati et al., 2002). We use the agroclimatic approach proposed by Sivakumar (1988) which reduces false starts due to the occurrence of damaging dry spells. Sivakumar (1988) described the onset date in the southern Sahelian and Sudanian climate zones of West Africa as the date after 1st of May when the rainfall accumulated over three consecutive days was at least 20 mm and when no dry spell within the next 30 days exceeded 7 days. This method is based on the fact that a dry spell of >7 days shortly after planting would be disastrous for the crop. We will retain only onset dates between the last dekad of May and the last dekad of July in order to match the cotton sowing date interval in Africa. In the same way, to detect the end date of the rainy season we applied the criterion defined by Stern et al. (1981) and Sivakumar (1988): the day from the 1st of September when consecutive 20 days are dry (recording <1 mm). We then compute the length of the rainy season, the rainfall amount and the number of dry spells within the rainy season. According to the definition by Sivakumar (1988), a dry spell event is defined as more than seven consecutive days with <1 mm/day of rainfall occurring between the onset and the end of the rainy season. As crop yield is not only dependent on rainfall and soil water storage, we use also the meteorological parameters available in synoptic stations such as global radiation and temperatures (maximum, minimum and mean) as the sum from 1st May to the end of October Cotton water requirement satisfaction index The use of process models is a good way to assess yield response to various crop management sequences and environmental conditions. However, most of the cotton crop models used in literature (McKinion et al., 1989 for the GOSSYM/COMAX model; Jallas et al., 1999 for the COTONS model) require a large number of parameters to describe the crop environment (soil hydrologic characteristics, soil carbon and nitrogen contents and daily climatic variables) and management (cultivar, dates and characteristics of cultural practices). For instance, the variety file of the GOSSYM/COMAX model contains 50 parameters that modify the growth and development of cotton cultivars (Boone et al., 1993). Owing to this large number of parameters, such models are difficult to calibrate and their results difficult to extrapolate from one site to another. We will thus orient this study to the use of much more simple water use efficiency models that have already shown their ability to provide a reliable indicator of annual crop performance (FAO, 1998; Senay and Verdin, 2003; Verdin and Klaver, 2002; Leplaideur, et al., 2002). They are usually based on the computation of an explicit WRSI based on the availability of water to the crop during a growing season and related to crop production using a linear yieldreduction function specific to a crop (FAO, 1977; FAO, 1979; FAO, 1986). Such indexes have been widely documented by the FAO and are up to now widely used. For instance, they have been recently adapted to be used in an operational mode with the Famine Early Warning Systems Network (FEWS NET; program which provides early warning and food security information to prevent famine and ameliorate food insecurity. The advantage, but also the limit of such model, is the very low requirement of input parameters to describe the crop and the site. It then can be quickly adapted from a region to another and from a crop to another. However, due to this low input parameters requirement, the model does not allow studying some specificities of a variety or a region. In our study, we define a specific WRSI to reproduce the year to year variability of the cotton crop, latter called cotton WRSI (CWRSI ). This index is based on the ratio between the water supply and demand during the cotton growing season. It is computed as: CWRSI = ( d=15 d=1 ) ET / a ET c ET c is the crop evapotranspiration under standard conditions referring to the evaporating demand from crops that are grown in large fields under optimum soil water, excellent management and environmental conditions and achieve full production under the given climatic conditions. It is calculated from the Penman Monteith reference crop evapotranspiration (ET o ) using the cotton crop coefficients (K c ) to adjust for the growth stage of the cotton crop: ET c = ET o K c

5 62 B. SULTAN ET AL. The calculation of ET o by means of the Penman Monteith equation requires radiation, air temperature, air humidity and wind speed data (FAO, 1998). The cotton crop coefficients K c are set according to FAO guidelines (FAO, 1977; 1979). The actual evapotranspiration (ET a ) is based on the Eagleman (1971) polynomial function for which a recent description is given by Affholder (1997): ET a = ET c + (4.97ET c 0.661ETc 2 ) MR (8.57ET c 1.56ETc 2 ) MR2 + (4.35ET c 0.88ET 2 c ) MR3 Soil water availability is calculated as a moisture ratio (MR), ranging from 0 to 1 as: MR = R/AWC The maximum available water capacity (AWC )ofthe soil and crop root depth is the range of available water that can be stored in soil and be available for growing crops (Richards and Wadleigh, 1952). R is the rainfall of a given time step. When R exceeds the AWC, themr value is changed to 1 to simulate water losses by deep drainage. The model time step is the dekad. All the climate inputs of the model (rainfall and ET o ) are summed by dekad. We consider 15 dekads as growing season length. This choice is determined by the length of the rainy season (between 14 and 15 dekads in Mali). The ratio ET a /ET c is computed for each dekad from the sowing dekad (d = 1) to the last dekad (d = 15) and the CWRSI is defined as the average of this ratio over the 15 dekads of computation. The CWRSI varies from 0 (the water requirement of the crop has never been satisfied) to 100% (the water requirement of the crop has always been satisfied). Notice that using a daily time step to compute the CWRSI gives very similar results. 3. Results 3.1. In experimental station Results from the climate indexes Table I reports the correlations between the cotton yield in Mali and the meteorological indexes defined in section and computed using the synoptic data observed at the N Tarla station. The correlation coefficients in bold are significant at the 1% level. The most salient parameters explaining the cotton yield are the length of the rainy season, the onset date and the rainfall amount during the rainy season. Their correlations with the cotton yield are, respectively, 0.68, 0.67 and Notice that the onset and the length of the rainy season are strongly linked, with a correlation coefficient between each other reaching The cessation dates of the rainy season show a significant relationship with the cotton yield at the 5% level, not at the 1% level. Table I. Correlation coefficients between cotton yields in Mali and climate indexes such as the length of the rainy season, the onset date, the annual rainfall amount, the global radiation (Grad), the end of the rainy season, the maximum (T max ), mean (T moy and minimum (T min ) temperatures and the number of dry spells (Dry). Yield Length 0.68 Onset 0.67 Rainfall 0.56 Grad 0.51 End 0.46 T max 0.41 T moy 0.41 T min 0.32 Dry 0.12 The most significant values (at the 1% level) are shown in bold. The correlation coefficients are computed over the period (1968, 1972 and 1975 are missing). Figure 4 shows the dispersion of the yield according to the onset dates (Figure 4(a)), the length of the rainy season (Figure 4(b)) and the rainfall amount within the rainy season (Figure 4(c)). In Figure 4(a), the dashed lines depict the mean yield (horizontal line) and onset date (vertical line). One can notice that 11 of 13 early onsets (before the day 159, i.e. 8th June) are characterized by a yield greater than the average while six of eight late onsets shows yields lower than the average. It is also interesting to notice that for onset dates varying from 140 (20 May) to 150 (30 May), the yield remains constant (see the left upper square of Figure 3(a)) may be due to sowing dates that are always later than 30 May. Figure 4(b) shows the relationship between the length of the rainy season and the cotton yield. It points out a very clear linear increase of the yield with the length of the rainy season varying from 90 to 140 days. The correlation between the yield and the length of the rainy season reaches 0.89 if we consider only lengths lesser than 140 days. Above 140 days, the length of the rainy season does not explain the yield variability anymore as the average duration of the cotton growth cycle does not exceed 150 days in Africa. Notice that the year 1980 appears as an outlier with a rainy season duration of around 140 days and a very low yield <1500 kg/ha. Figure 4(c) gives elements to explain this singularity as there is positive relationship between the annual rainfall amount and the yield and as the year 1980 reveals a rainfall deficit with only 615 mm observed during the rainy season. The link between the yield and the rainfall amount is very clear for the years with <900 mm occurring during the rainy season. The correlation coefficient 0.56 for the whole time series (Table II) and reaches 0.75 if we consider only the rainfall amounts less than 900 mm. For the three wetter years exceeding

6 MULTI-SCALES AND MULTI-SITES ANALYSES OF THE ROLE OF RAINFALL IN COTTON YIELDS 63 (a) (b) (c) (d) Figure 4. Relationships between (a) cotton yields and onset dates, (b) cotton yields and the length of the rainy season, (c) cotton yields and the annual rainfall, (d) the annual rainfall amount and the length of the rainy season. The cotton yields (kg/ha) is measured by IER/SRCFJ from 1965 to 1990 on a cultivated plot in the region of Koutiala in Mali (1968, 1972 and 1975 are missing). The rainfall (mm) is measured at the N Tarla station. The full line in (b), (c) and (d) represents the regression line. The dashed lines in (a) depict the mean yield (horizontal line) and onset date (vertical line). The dashed line in (b) depicts the length 140. The dashed lines in (c) and (d) depict the 900 mm value. Table II. Correlation coefficients between CWRSI, cotton yields in Mali and climate indexes such as the length of the rainy season, the onset date, the annual rainfall amount. Sim1 Sim2 Yield Length Onset Rainfall End Two CRWSI s are considered: Sim1 is the CWRSI computed using the observed sowing date and Sim2 the CWRSI using the onset date. The most significant values (at the 1% level) are shown in bold. The correlation coefficients are computed over the period (1968, 1972 and 1975 are missing). 900 mm 1988, 1965 and 1967 the yield appears to be much lower than expected by the linear relationship between yield and rainfall. Moreover, Figure 4(d) shows that although there is a positive correlation between the length of the rainy season and the rainfall amount, these three wet years are not characterized by an abnormally long rainy season. It means that for these 3 years the distribution of rainfall is non-homogeneous with periods of heavy falls and the occurrence of dry spells. A closer look at the daily rainfall time series for these years confirms the occurrence of rainy days with >90 mm per day and prolonged dry spells occurring in the second part of the cotton growth cycle. For instance, the rainfall time series of the year 1967 with the highest rainfall amount (1087 mm) reveals one day with 92 mm and several days with >60 mm. It reveals also a late sowing date (1st July) around 1 month after the onset of the rainy season detected by the Sivakumar method. This late sowing might explain the relatively low yield observed this year as the crop growth cycle has missed 1 month of rainfall in June and has experienced a prolonged dry spell in October and November. Our results presented in Table I also show that the yield does not seem to be affected by the occurrence of dry spells within the rainy season. The correlation coefficient remains insignificant while varying the length of the detected dry spells from 6 to 10 days and while taking into account only dry spells occurring from the sowing date to the end of the cotton growth cycle 150 days later. However, as the sensitivity to water stress varies along the growth cycle, the detection of dry spells over the whole rainy season can lead to miss the potential effect of dry spells occurring on critical stages for the crop. For instance, Sultan et al. (2005) have shown a differential impact of dry spells on millet yield depending on the phenological stages of the crop affected by the drought. One thus needs to point out, if any, the critical stages for the cotton crop in which a dry spell would have a negative impact on the crop yield. Figure 5 shows the correlation between the cotton yield and the number of dry spells detected within a moving window of 30 days from the first day to the last day of the year. For instance, the correlation of the day 150 represents the correlation between the yield and the number of dry

7 64 B. SULTAN ET AL. Figure 5. Correlations between the cotton yield in Mali and the occurrence of dry spells detected within a moving window of 30 days from the first day to the end of the year. Several variations in the criterion of the detection of dry spells have been made in order to have an ensemble of correlations for each day. These variations concern the minimum length of the dry spell varying from 6 to 8 days and the minimum rainfall threshold for a dry day that can be either 0 mm/day or 1 mm/day. The correlations are computed over the period (1968, 1972 and 1975 are missing). Horizontal lines depict the negative and positive correlation values significant at the 1% level of confidence. spells between the day 150 and the day 179. Several variations of the definition of Sivakumar have been used to detect dry spells by varying the minimum length of the dry spell (from 6 to 8 days) and the minimum rainfall threshold characterising a dry day (1 or 0 mm/day). The results show that there are two critical periods for cotton in which a dry spell has a negative impact on yield in early and in late season. Significant negative correlations between 0.50 and 0.69 are observed with dry spells occurring between the last dekad of May and the first dekad of June (lasting until the last dekad of July), i.e. during the sowing period. Significant negative correlations are also observed for dry spells occurring from the second part of August to mid-september (lasting until mid-october). This critical period occurs generally between 60 days and 90 days after sowing, i.e. during the blooming phase. The same results are obtained using the 30-days moving rainfall average (not shown) instead of the number of dry spells to build Figure 5, pointing out the importance of the months in margin of the rainy season. It is coherent with the high correlation observed in Table I between the length of the rainy season and the cotton yield. It is also coherent with the conclusions of Parry (1982) showing that the cotton plant is especially vulnerable to water stress at two stages of its growth, right after it has been sown and during the blooming phase. If we consider only the effects of climate variables and not those of on-farm parameters (e.g. mineral inputs and crop protection), attainable yield is mainly limited by water, energy or both. However, Table I does not show any significant correlations at the 1% level between cotton yield and the other climatic parameters such as solar radiation and temperatures. The correlation with solar radiation is the highest and significant at the 5% level. The maximum and minimum temperature sums show a weak negative correlation barely significant at the 5% level. However, it does not imply a causal link between temperature and yield since this correlation is associated with rainfall amount variations. Indeed, these two temperatures indexes are negatively correlated with the rainfall amount occurring during the rainy season. Moreover the computation of the partial correlation between temperature and yield, removing the rainfall effect, gives a positive but non significant correlation value around Results from the CWRSI We then compute the CWRSI using the synoptic data recorded at the N Tarla station and the cotton sowing dates for the years (with three missing years: 1968, 1972 and 1975). The first step is to choose the value of AWC which is also an input of the model (see Materials and methods section). Affholder (1997) deduced the AWC value from the soil texture by measuring the soil clay and fine silt content (%) in the top 1-m soil layer according to the work of Hamon (1980) and Imbernon (1981). As such data are not available in N Tarla, we use the observed cotton yield to calibrate the model. We calculate the evolution of the correlation coefficient between cotton yield and CWRSI, computing with an AWC value varying from 1 to 150 mm. The correlations are significant at the 1% level with AWC values ranging from 4 to 70 mm with the highest correlations exceeding 0.70 obtained with AWC values ranging from 35 to 60 mm. For the highest AWC values (>70 mm), the rainfall amount per dekad recorded at the N Tarla station is most of the time (in particular for the early and late dekads of the cotton growing season) weaker than AWC leading to a weak soil moisture availability and an unrealistic water stress for the crop. The AWC value can thus be chosen among the values maximising the correlation between the observed

8 MULTI-SCALES AND MULTI-SITES ANALYSES OF THE ROLE OF RAINFALL IN COTTON YIELDS 65 yield and the CWRSI. The robustness of the model can be assessed using the simple leave-one-out cross-validation procedure describe below: - We choose the AWC value from a portion of the data (called training period) comprising all years minus one and we use this value to compute the CWRSI for the remaining year. - We repeat this step for each year to compute a crossvalidated correlation between the observed yield and the CWRSI, which can be considered as a realistic representation of the skill of the model applied to unseen years. The cross-validated correlation is very high, up to 0.78 (Table II). The stability of the model is attested by very low variations of the AWC values over the different training periods (around 41 mm). These values of AWC correspond to a rooting depth of about 35 cm in such typical alfisol with very weak soil clay content (Wybrecht et al., 2002). Notice that the use of the CWRSI improves clearly the link between climate and cotton as the obtained correlation coefficient is greater the ones observed in Table I. Figure 6 shows in the left axis the CWRSI for the 23 years of simulation and in the horizontal axis the observed yield (a), the onset (b), the length (c) of the rainy season and the rainfall amount within the wet season (d). The CWRSI reproduces well the cotton yield variability without any bias (Figure 6(a)). However, it underestimates the yield variance: the variation coefficient is near 15% for the CWRSI while it reaches 25% for the observed yield. The relationships between the CWRSI and the onset dates as well as the length of the rainy season are strongest and more linear than the that shown in Figure 4 using the observed yield. The observed asymptotic relationship between rainfall amount and yield (Figure 4(c)), as would be expected from crop physiology, is revealed by the CWRSI (Figure 6(c)) with the same threshold of 900 mm across that with which the dependency with rainfall amount seems to disappear. It is particularly true for the year 1967 showing a relatively low yield, although the seasonal rainfall amount is very high in coherence with observations. One of the advantages of the computation of the CWRSI is that it allows documenting the sensitivity of the simulated yield to the sowing date as this date is an input of the model. As the model captures well the observed variability of cotton yield and its relationships with the characteristics of the rainy season (onset, length, etc.), it gives greater confidence to the model s ability to capture the response of the cotton yield to a change of the sowing dates. For instance, we can compute a new CWRSI, but using the onset date instead of the observed sowing date and examine the differences between the two simulated yields. Figure 7(a) shows the time series of the two CWRSI. For most of the years the CWRSI is greater using the onset date as the sowing date. However, several years such as 1973, 1981 and 1986 show a significant decrease of the simulated yield. It is interesting to notice that these 3 years are characterised by a late onset, respectively, 13 July, 30 June and 12 June. More generally, the difference between the two CWRSI shows a linear relationship with the difference between the sowing and the onset (a) (b) (c) (d) Figure 6. Relationships between (a) cotton yields and CWRSI, (b) CWRSI and the onset dates, (c) CWRSI and the length of the rainy season, (d) CWRSI and the annual rainfall amount. The CWRSI is expressed in percentage. The rainfall (mm) is measured at the N Tarla station from 1965 to 1990 (1968, 1972 and 1975 are missing). The full line in (a) represents the regression line.

9 66 B. SULTAN ET AL. CWRSI CWRSI difference Time (years) Sowing date minus Onset date Figure 7. (a) Time series of CWRSI (%) in Mali using the observed sowing date (black line) and the onset of the rainy season (grey line) as the sowing date for the CWRSI computation over the period. (b) Differences between the CWRSI computed using the observed sowing date and the CWRSI computed using the onset date. This difference is expressed in percentage as the ratio to the CWRSI computed using the observed sowing date. Positive (negative) values mean that the CWRSI is greater (weaker) while using the observed sowing date instead of the onset date. These differences are related to the difference between the observed sowing date and the onset date. Positive (negative) values in the horizontal axis mean that the observed sowing date follows (precedes) the onset date. The full line in (b) represents the regression line. This figure is available in colour online at date (Figure 7(b)): the yield tends to increase (decrease) while sowing before (after) the onset date. Table II shows that the correlation between the new CWRSIs using the onset date and the observed yield is weakened, but the relationship with the seasonal rainfall amount increases with a coefficient correlation of Taking into account the onset of the rainy season, thus seems to improve the relationship between water available and water used by the plant and thus seems to potentially increase crop water use On-farm situation A similar analysis of the link between climate and cotton yield is applied using farmers yields in North Cameroun. This application will allow us (1) to assess the role of climate in an area where many factors such as human management, biotic stresses and/or pests impact crop productivity (mean and variability) and (2) analyse the role of climate from a spatial to an inter-annual point of view as both climate and agronomical data are available over 10 years and 28 administrative units. Notice that only rainfall will be considered for climate as it is the most salient parameter in the experimental station (section 3.1) and only two synoptic stations are available in North Cameroun A spatial point of view The role of rainfall in the spatial distribution of the mean cotton yield in North Cameroun is investigated by averaging yields and rainfall indexes over the period for the 28 administrative units (Figure 8). Notice that for half of the administrative units (white points in Figure 8) only 3 years of rainfall are available. Figure 8 shows a clear linear relationship between the spatial distribution of the mean cotton yield and the mean rainfall indexes. As for the Malian experimental plot, the most salient parameters are the onset date (R = 0.53), the length of the rainy season (R = 0.55) and the rainfall amount in June (R = 0.63). As the end of the rainy season is not correlated with the spatial distribution of the mean yield, we can argue that these three rainfall parameters (onset date, length of the rainy season and rainfall amount in June) point out the same relationship: the administrative units with a late (early) onset have low (high) rainfall in June, a shorter (greater) rainy season duration and tend to have lower (higher) yields. Figure 9 shows the spatial distribution of both June rainfall and cotton yield. There are several similarities in the two maps. The administrative units in the centre of Cameroon (Hama, Pitoa, Padem, Bibem) are the most productive area and also those with the wettest conditions in June. We can also depict a similar latitudinal gradient with a decrease of cotton yields and June rainfall. Similar to the results from the experimental station but for a spatial point of view, the relationship between the total rainfall amount and the yield is very low with a correlation coefficient around An inter-annual point of view The role of rainfall on the year-to-year variability of farm cotton yields is investigated in this section. This investigation is associated with a great limitation which is the less number of years (10). When computing a correlation with only 10 points, the result is highly instable (the correlation coefficient can be completely different by removing or changing only a single value) and the link needs to be very strong to pass a significance test (R = 0.75 at the 1% level). To go through this limitation, we combine the spatial and temporal variability in yield and rainfall in a single analysis. To keep only the inter-annual variability, we first remove the spatial mean and variance by removing to each administrative unit its mean and dividing the values by the standard deviation computed over the 10 years. Notice that only the 14 complete administrative units have been considered. We then build a single vector from the 10 years and the 14 complete administrative units (140 points). Table III shows the correlations between the rainfall parameters and the cotton yields. All the parameters except the seasonal rainfall amount are significantly correlated with the cotton yields at the 1% level. As for the Malian experimental plot,

10 MULTI-SCALES AND MULTI-SITES ANALYSES OF THE ROLE OF RAINFALL IN COTTON YIELDS 67 (a) (b) (c) (d) (e) Figure 8. Spatial relationships between cotton yields in Cameroon and (a) the onset dates, (b) the length of the rainy season, (c) the end of the rainy season, (d) the June rainfall amount, (e) the annual rainfall amount. The cotton yields (kg/ha) and the climate indexes are averaged over the period (2002 is missing) for the black points and over the period for the white points. Figure 9. Spatial distribution of the cotton yields in kg/ha (left) and the June rainfall amount in mm/month (right). Both yields and rainfall are averaged over the period (2002 is missing) when 10 years of data are available and over the period when 3 years of data are available (see Figure 3 for the map of data availability).

11 68 B. SULTAN ET AL. Table III. Correlations between cotton yields in Cameroon and four rainfall parameters: the onset dates, the length of the rainy season, the June rainfall amount and the annual rainfall amount. Local Regional Cotton area June rainfall Length Onset Annual rainfall The correlations are computed first at the local scale (the administrative unit scale), the regional scale and the whole cotton area (see map of Figure 3). Only the administrative units where >10 years data were available over the period are considered for the computation of the correlation. The most significant values (at the 1% level) are shown in bold. in coherence with the spatial analysis, the most salient parameters are the length of the rainy season (R = 0.37) and the June rainfall (R = 0.36). To describe the range of correlations in each administrative unit, we compute the correlations between the rainfall parameters and the yield for each administrative unit and show on Figure 10 the distribution of the correlation coefficients. Notice that we use only the 14 complete administrative units to compute the correlation. A large number significant correlations (at the 10% level, i.e. R = 0.50 ) assess the robustness of the relationship between the rainfall indexes and the cotton yield at the inter-annual scale. Figure 10 shows that 6 of 14 administrative units point out a significant correlation (at the 10% level) between yield and either onset date or rainy season duration. Five administrative units show a significant correlation (at the 10% level) between yield and June rainfall. The other indexes present <4 of 14 units with significant correlation. These results are coherent with the ones obtained using the experimental plot while the link between climate and yield seems to be reduced as the maximum number of significantly correlated areas is prior than half of the administrative units. We then classified the administrative units according to their correlations between rainfall parameters and yields to try to explain the differences in the sensitivity to climate variability from a unit to another. Figure 11 shows the correlation coefficients for the three salient rainfall parameters (e.g. the onset date, the length of the rainy season and the June rainfall amount) according to the annual rainfall amount averaged over the period. It seems that there is a link between the intensity of the correlation with climate and the average climate conditions over the administrative units: the driest administrative units are the most sensitive to year-to-year variations of climate and this sensitivity decreases with the increase of the mean rainfall (see the linear decrease of the correlation coefficients from 600 to 800 mm/year). This relationship between the average climate conditions and the intensity of the link with climate is the most obvious while using the length of the rainy season (see triangle in Figure 11). It is noisier while using the other climate parameters but this apparent erratic response to (a) (b) (c) (d) (e) Figure 10. Inter-annual relationships between cotton yields in Cameroon and (a) the onset dates, (b) the length of the rainy season, (c) the end of the rainy season, (d) the June rainfall amount and (e) the annual rainfall amount. For each administrative unit with >10 years data over the period (14 of 28), The correlation between climate indexes and cotton yield was computed and the histogram of the correlation coefficients was represented. The left axis represents the number of administrative units and the shaded area the coefficients values between 0.5 and 0.5 that are not significant at the 10% level.

12 MULTI-SCALES AND MULTI-SITES ANALYSES OF THE ROLE OF RAINFALL IN COTTON YIELDS June rainfall Length Onset Annual rainfall local regional cotton area Figure 11. Correlation coefficients between cotton yields and climate in Cameroon for the three salient rainfall indexes according to the annual rainfall amount (mm/year) averaged over the period (2002 is missing). The three rainfall indexes are the onset date (circles), the length of the rainy season (triangles) and the June rainfall amount (crosses). Only the correlations >0.5, significant at the 10% level, are shown. climate might be due the instable character of the correlation coefficient using very few points and/or to some local characteristics, in human management for instance, that can hide or highlight the climate signal. To discriminate between these local properties and potential large-scale forcing of rainfall in cotton yields, we aggregate yield and climate data first at an intermediate scale by averaging data at the regional level (see the administrative map on Figure 3) and second at the larger scale by averaging the data over the 14 complete administrative units. The aggregation of local data as an upscaling technique is a simple way to go beyond data heterogeneities and idiosyncratic details in order that only the important generalities conditioned by the large-scale forcing, e.g. climate variability, remain. Notice that the regional analysis is performed in the same way as that for the administrative unit analysis: we first remove the spatial mean and variance by removing to each region its mean and dividing the values by the standard deviation computed over the 10 years. We then build a single vector from the 10 years and the five complete regions (50 points) and compute the correlations between the rainfall parameters and the cotton yield. Finally, we consider a larger scale analysis where the correlations are computed between the rainfall parameters and the yield averaged over the 14 complete administrative units. Figure 12 and Table III present the correlations for each scale and each rainfall parameter. The aggregation clearly amplifies the climate signal as we observe a clear linear relationship between the cotton yields since the correlations values increase with the scale. Even if the sample is small (only 10 points corresponding to the 10 studied years), we observe a clear linear relationship between the cotton yield and the onset of the rainy season (R = 0.66), the Figure 12. Absolute correlation coefficients between cotton yields in Cameroon and four rainfall parameters: the onset dates, the length of the rainy season, the June rainfall amount and the annual rainfall amount. The correlations are computed first at the local scale (the administrative unit scale), the regional scale and the whole cotton area (see map of Figure 3). Only the administrative units where >10 years data were available over the period are considered for the computation of the correlation. rainy season duration (R = 0.69) and the June rainfall amount (R = 0.83). The total rainfall amount and the end of the rainy season (not shown) remain apparently unlinked with cotton yield. The most salient parameter is the june rainfall amount. It correlates significantly (at the 1% level of confidence) with cotton yield at each scale. 4. Conclusions and perspectives Using both experimental and on-farm data, this study has pointed out an important role of rainfall variability on cotton yields in West Africa at different time and space scales. First, we have shown a strong link between rainfall variability and cotton yield at the inter-annual time scale. This link is particularly obvious using the experimental dataset in Mali where the yield is obtained under optimal resources and thus close to the maximum yield attainable. In such conditions, the computation of a very simple crop water satisfaction index depicts very well the year-to-year variations of cotton yields. As farmers yields are often far from the attainable yields because of many unknown biotic and abiotic factors (Affholder et al., 2003; Baron et al., 2005), the signature of climate is less clear but remains present. We have shown that the intensity of the link between rainfall variability and farmers yields seems to be dependant on the mean climate. Indeed the signature of climate on the year-to-year variability of cotton yields is more obvious in the driest locations than in the wet locations. We also found that the aggregation of local data is a good way to go beyond data heterogeneities and idiosyncratic details in order that only the important productivity generalities conditioned by the large-scale climate forcing remain. The apparent erratic response

13 70 B. SULTAN ET AL. to rainfall variability at the local scale disappears when using this bottom up approach. Second, this study has pointed out the role of rainfall variability in the spatial distribution of the mean cotton yields using agronomical and rainfall data over 28 administrative units in Cameroon. This role is as important as the ones detected at the inter-annual time scale. In a context of a changing climate with the recent drought over the soudano-sahelian Africa (Nicholson, 1986) and the global change due to the anthropogenic greenhouse forcing, one could expect changes in the spatial distribution of the mean cotton productivity with some lands becoming less or more adequate to cotton cultivation, inducing an increase competition for the use of the more productive lands. It is important to retain the close link between the spatial mean state scale and the inter-annual scale from this study. Figure 11 illustrates clearly this link, showing that the driest locations are not only less productive in terms of mean yield but also the most sensitive to climate fluctuations and thus the locations where the risk of crop failure due to climate is the highest. In the northern part of the cotton production belt, a potential decrease of the mean annual rainfall in the context of climate change would decrease the mean productivity and increase the risk due to climate variability. Notice from the shape of the link shown in Figure 11 that this decrease in the mean rainfall would not increase the climate risk linearly but exponentially. Third, we have shown that, either for the experimental plot in Mali or for the on-farm conditions of Cameroon, the most salient climate variables are rainfall characteristics, in particular the length of the rainy season. The cotton productivity seems to be very sensitive to a rainfall deficit in early season (in May June) and in late season (September October) which leads to a shorter rainy season length. The total rainfall amount occurring during the rainy season appears to be of less importance in Mali and totally disconnected with farmers yields in Cameroon. These results are important in working towards the predictability of crop yields using rainfall information, in particular to highlight the most salient rainfall parameters that are needed in a forecast system. For instance, the last results from the climate scientific literature have been translated into operational forecast methods for the summer rainfall. These seasonal forecasts are performed by national meteorological and hydrological services and included as a part of the West Africa Climate Outlook Forum (PRESAO; see Hamatan et al., 2004; Ward et al. 2004) which has produced seasonal rainfall outlooks for the region for the summer season each year since Since this exercise is focused on the total rainfall amount over the whole rainy season and does not give any predictions of the onset or the length of the rainy season, it is thus not useful to anticipate the year-to-year fluctuations of cotton yields. Such study can thus be used to highlight the critical variables that are needed for this specific application. This study can be used as a basis to examine the climatic scenarios from the Intergovernmental Panel on Climate Change (IPCC) and estimate the potential impact of climate change on cotton productivity in West Africa. The examination of the scenarios of the changes in the spatial distribution of early season rainfall for instance could give an index of the changes in the productivity map of cotton. Moreover, one can look at the changes in the interannual early season rainfall variability to give scenarios on potential changes in the year-to-year variability of cotton yields. This study can also be used to give some recommendations to climate community on the most salient climate parameters that should be included in a weather forecast system to be useful to anticipate the year-to-year variability of cotton yield variability. Another perspective is to use the climate indexes depicted in this study as important determinants for cotton productivity in a weather index-based insurance for the cotton cultivation. Indeed, traditional agricultural insurance products, based on the findings of a bad harvest, are seldom developed because of the difficulty of assessing damages and of the information asymmetry between the farmer and the insurer. For a few years, insurance products based on weather indexes, such as the ones described in this paper, have been launched in few developing countries (Malawi, India, Ethiopia) but not yet in West Africa (Hess and Syroka, 2005; Luo et al., 1994). Yet they could mitigate the potentially tragic impact of a bad rainy season for the cotton farmers since the link between climate indexes and yield is strong enough. Acknowledgements The authors are thankful to M. Crétenet, R. Morel, M. Angokai and P. Asfom for providing data and for useful discussions. The authors thank the two anonymous reviewers who helped to clarify and improve this paper. References Affholder F Empirically modelling the interaction between intensification and climatic risk in semi-arid regions. Field Crops Research 52: Affholder F, Scopel E, Madeira Netto J, Capillon A Diagnosis of the productivity gap using a crop model. Methodology and case study of small-scale maize production in central Brazil. Agronomie 23: Albergel J, Carbonnel JP, Vaugelade J Aléas climatiques et production agricole: le cotton au Burkina. Acta Oecologica- Oecologia Applicata 6: Ati OF, Stigter CJ, Oladipo EO A comparison of methods to determine the onset of the growing season in northern Nigeria. International Journal of Climatology 22: Baron C, Sultan B, Balme M, Sarr B, Lebel T, Janicot S, Dingkuhn M From GCM grid cell to agricultural plot: scale issues affecting modelling of climate impact. Philosophical Transactions of the Royal Society of London. Series B, Biological sciences 360(1463): Blanc E, Quirion P, Strobl E The climatic determinants of cotton yields: evidence from a plot in West Africa. Agricultural and Forest Meteorology 148: Boone MYL, Porter DO, McKinion JM Calibration of GOSSYM: Theory and practice. Computers and Electronics in Agriculture 9: