All posts by Libby White

WRSI OUTLOOK FOR THE 2017-18 SOUTHERN AFRICA MAIZE GROWING SEASON

Average to mildly above average crop success anticipated for much of the major maize producing areas of Southern Africa, with pockets of moderately reduced crop success in subsistence farming and livestock grazing regions.

Will Turner and Greg Husak

The Water Requirement Satisfaction Index (WRSI) is a water balance model that incorporates crop parameters such as the length of growing season and phenological information, along with soil characteristics. This model is commonly used to monitor growing seasons throughout Africa and Central America by the Famine Early Warning Systems Network (FEWSNET). Specifically, the model uses gridded precipitation and reference evapotranspiration (RefET) inputs to calculate the ratio of plant-available water to crop-specific water demand at each stage of crop development. This post describes a real-time monitoring application to examine the progress of the Southern Africa 2017-2018 maize growing season, at approximately the halfway point of the season.

The analysis combines September, October, and November final CHIRPS rainfall estimate (reference link, data link), and December CHIRPS Preliminary data (data link) to identify the to-date 2017-18 Start of Season (Figure 1). The dekad identified as the Start of Season (SOS) for a location (pixel) is defined as the first dekad with at least 25 mm of rainfall followed by a total of at least 20 mm in the next two consecutive dekads. For the example shown here, which only has data through the end of December, dekads 2 and 3 of December only use the 25 mm threshold to identify a potential start. Whether or not the second threshold was met cannot be determined until January CHIRPS Prelim is released. All areas that have not met these requirements by the third dekad of December are labeled as ‘Yet to Start’ (see the areas of Figure 1 in light gray color).

Historical CHIRPS (1981-2016) was used to identify the SOS for each of the last 36 years, and calculate the median historical SOS (Figure 2).

Figure 1
Figure 1. 2017-18 start of season as of December 31, 2017. Areas colored gray have yet to receive sufficient rainfall for a season start.
Figure 2
Figure 2. Median start of season based on CHIRPS from 1981-2016. Areas colored pink consistently do not receive sufficient rainfall for a season start.

To examine the potential outcomes of the current season, 36 WRSI scenarios were run using the result from Figure 1 as the SOS. To get credible simulations of the season, hybrid composites of both precipitation and RefET¹ were created using season-to-date information and previous years to estimate one potential remainder to the season. The precipitation input of each scenario used 2017 CHIRPS final through November and 2017 CHIRPS Prelim for December. The RefET input used 2017 data through September. Because RefET is not available for October-December of 2017 due to the latency of the dataset, these 9 dekads were filled with average RefET. The 36 season scenarios were then run to completion using corresponding year-specific CHIRPS and RefET (i.e. Scenario #1 used 2017 data for Sep-Dec, and 1981 data for Jan-May; Scenario #2 used 2017 data for Sep-Dec, and 1982 data for Jan-May; etc.).

To build a library of historical WRSI output, the WRSI calculation was run using the 36 years of historical SOS from Part 1 of the examination and historical CHIRPS and RefET.

Investigating the difference of the median historical SOS (Figure 2) from this season’s SOS (Figure 1) reveals the timeliness of the current season (Figure 5). The median of the 36 2017-18 WRSI scenarios (Figure 3) was then compared to the WRSI median of the historical runs (Figure 4) to find the simulated WRSI anomaly (Figure 6).

WRSI Colormap
WRSI Colormap for Figures 3 & 4.
Figure 3
Figure 3. Median WRSI for the 2017-18 maize growing season, based on 36 scenarios. Areas colored cyan have not yet received sufficient precipitation for a season start.
Figure 4
Figure 4. Historical median WRSI for the Southern Africa summer maize growing season, based on 1981-2016 WRSI. Areas colored pink consistently did not receive sufficient rainfall for a season start.
Figure 5
Figure 5. 2017-18 start of season anomaly with respect to the 1981-2016 median SOS. Areas colored in shades of blue indicate regions where the current season started earlier than the historical median, while areas colored in shades of red started late. Areas colored orange indicate regions that are past the median SOS for that location and still have yet to receive sufficient rainfall for a season start. Areas colored beige have not yet started, but are not beyond the median start date.
Figure 6
Figure 6. 2017-18 simulated end of season WRSI anomaly with respect to the 1981-2016 median WRSI. Areas colored beige have yet to receive sufficient rainfall for a season start in 2017-18.

Comparing the graphics shown in Figure 5 and 6 we identify some interesting patterns, and can summarize how changes in SOS may impact this season’s crop production in Southern Africa.

  • Major production zones in the Maize Triangle of South Africa, Northern Zimbabwe, Central and Northern Mozambique and Malawi experienced slightly early to on-time SOS (Figure 5). As a result of this, these areas are all showing average to above-average WRSI (Figure 6) using this analysis.
  • More marginal production zones in Southern Zimbabwe and Southern Mozambique, where subsistence farmers are dependent on local production for their food security, experienced more extreme SOS anomalies (early in Zimbabwe, late in Mozambique). These anomalies are leading to below-average WRSI anomalies.
  • Throughout much of Botswana and West-central South Africa, there has been a lack of seasonal onsets. These delays are anticipated to have negative impacts on the crop growing season, even if rain begins in early January.

This analysis is currently experimental, but we think it should be insightful in identifying areas which are likely to experience crop production anomalies early in the season. We will continue to monitor throughout the growing season.

The next logical step would be to start including CHIRPS-GEFS forecast to get projected rainfall information, identify the chances of SOS in the upcoming dekad, and then use that to forecast rest-of-the-season progress.

¹Dataset presented at the 2017 FEWS NET Science Update Meeting. Hobbins MT (2017), A global reference evapotranspiration service for the FEWS NET Science Community: derivation, FEWS applications, next steps. FEWS NET Science Update Meeting: Agro-Climatology for Food Security Assessment, Washington, D.C., 6-8 September. (Oral)

Reformulated NMME simulations indicate probable dryness for the 2017 and 2018 short and long rains in eastern East Africa

Chris Funk, Andy Hoell, Wassila Thiaw, Diriba Korecha, Gideon  Galu, Miliaritiana Robjhon, Endalkachew Bekele, Shraddhanand Shukla, Pete Peterson, and James Rowland

Introduction

This analysis has two objectives:  to i) provide a timely multi-agency assessment of the forthcoming short and long rainy seasons for eastern East Africa, while also ii) advancing towards a multi-agency paper describing a ‘FEWS NET Integrated Forecast System’ (FIFS). The main ideas behind FIFS are to improve the accuracy of seasonal precipitation forecasts by using statistical ‘Model Output Statistic’ (MOS) reformulations based on output from coupled ocean-atmosphere models, such as the North American Multi-Model Ensemble (NMME).  Note that while the work presented here is ‘experimental’, it is backed by years of FEWS NET research and successful forecasts based on similar approaches. The basic relationships used to drive these forecasts involve La Niña-like modifications of the Walker Circulation, which FEWS NET research, motivated by the 2011 famine in Somalia, has shown to be a key driver of droughts during both the October-November-December ‘short’ rains and the March-April-May ‘long’ rains.  Please see the list of papers at the end of this report.

The method explored here is called ‘adaptive regression’ (AdReg), which is an outgrowth of the modified constructed analog approach explored by Shukla and Funk.  AdReg is also similar to the process that an expert statistical climate forecaster might carry out, i.e. similar to the approaches used to derive outlooks for climate outlook fora.  AdReg has two steps: selection and estimation. AdReg employs three selection criteria: correlation, activity, and dynamic significance. In this application, ensemble average NMME precipitation fields will be used as predictors. We will use October-November-December (OND) NMME forecasts to predict OND eastern East African (EEA) rainfall and OND NMME forecasts to predict MAM EEA rainfall.

Background

This work has been motivated by the severe food crisis in EEA. As announced in a recent FEWS NET alert, the prospect of a fourth, and perhaps even a fifth, consecutive drought looms over East Africa due to the emerging La Niña-like sea surface temperature conditions. At present, the world faces an extremely high level of food insecurity, with an estimated 81 million people requiring urgent assistance. In East Africa, consecutive droughts have pushed millions to edge of starvation. In Somalia, food insecurity remains very high, with close to a million people (more than 800,000) facing near-famine (Emergency) conditions, another 2.3 million people in a severe food crisis, and another 3.1 million facing substantial food stress; an estimated 388,000 children under the age of five are acutely malnourished. In Kenya, persistent high food prices continue to drive high levels of food insecurity. In the Somali region of Eastern Ethiopia food security conditions are approaching famine conditions, and over all there are ~5-10 million Ethiopians in facing acute food insecurity.

Figure 1 shows the early October Integrated Phase Classification (IPC) food insecurity outlook map from www.fews.net. Orange and red areas in this map indicate areas experiencing a severe food crisis (orange) or emergency (red). These shades indicate severe food stress. For this study, we have selected the area outlined in black. This region was selected because of its high level of food insecurity, regional homogeneity, and fairly good data support.

Figure 1. FEWS NET Food Security Outlook from Early October
Figure 1. FEWS NET Food Security Outlook from Early October

Over the past thirty years, rainfall in this region during the OND season has remained fairly steady, probably due to the beneficial influence of warming western Indian Ocean SSTs. Rainfall during the MAM long rains, however, has declined substantially. Figure 2 shows a 1900-2017 March-April-May Standardized Precipitation Index (SPI) for this area, based the Centennial Trends (1900-1980) and CHIRPS (1981-2017) data sets, centered on a 1981-2010 baseline. The Centennial Trends and CHIRPS datasets are highly correlated (0.93) over the 1981-2014 time period.  What is striking about Figure 2 is large frequency of dry seasons since 1999. Only 2010 and 2013 experienced good rainfall performance. The onset of La Niña-like climate conditions combined with warm SSTs in the western Pacific could set the stage for another sequence of back-to-back droughts during the OND 2017 and MAM 2018 long rains.

Figure 2. Eastern East Africa MAM Standardized Precipitation Index time series.
Figure 2. Eastern East Africa MAM Standardized Precipitation Index time series.
Data

Here, we make AdReg forecasts based on the ensemble average of the latest NMME model simulations. Bias corrected precipitation simulations have been provided by Brent Roberts and the NASA SERVIR program. These forecasts have been bias corrected using Global Precipitation Climatology Center precipitation fields. In general, the NMME models predict moderate La Niña conditions (Figure 3) during the fall and winter of 2017/18, followed (in most models and simulations). In Figure 4 we see the NMME OND precipitation forecast. Consistent with a moderate La Niña forecast and moderately warm West Pacific SSTs, we see strong subsidence at the dateline near the equator. Such subsidence is typically a robust indicator of an enhanced La Niña-like amplification of the Walker Circulation. Over East Africa, we see a general tendency to slightly above normal rainfall.  An examination of historical correlation between the NMME and OND rainfall in our target region, will question and revise this assessment (below).

Figure 3. NMME NINO3.4 forecasts, from http://www.cpc.ncep.noaa.gov/products/NMME/current/images/nino34.rescaling.ENSMEAN.png
Figure 3. NMME NINO3.4 forecasts, from http://www.cpc.ncep.noaa.gov/products/NMME/current/images/nino34.rescaling.ENSMEAN.png
Figure 4.Standardized OND NMME ensemble forecasts of precipitation.
Figure 4.Standardized OND NMME ensemble forecasts of precipitation.
Methods

Here we explore the application of ‘adaptive regression’, a new type of statistical reformulation being developed by FEWS NET. AdReg filters potential global climate predictors by three criteria: correlation, activity, and dynamic significance. Correlation is based on the historical correlation between our target and each locations historical SST or precipitation forecast time series. Our main target is 1997-2016 OND rainfall in our Eastern East Africa study site (Figure 1). We begin by using a threshold of |0.5| as our criteria for correlation.

Figure 5 shows the correlation between OND EEA precipitation and the NMME forecast SST and Precipitation. Correlation values less than 0.5 and greater than -0.5 have been set to 0. For OND we see a positive-negative-positive relationship to SST in the Western Indian, Western Pacific and Eastern Pacific Ocean. The precipitation correlation pattern looks similar to that shown in diagnostic analyses of the short rains. Note, however, the details of the correlation structure over Eastern Africa and the Indian Ocean. During dry EEA OND years, the model produces strong subsidence and drying to the east of East Africa, primarily over the Indian Ocean. In the actual NMME forecast (Figure 5), conditions are fairly neutral in this area, and fairly active over the ENSO-related regions of the western and eastern Pacific.

To generate AdReg forecasts, we follow two steps: selection and estimation. AdReg employs three selection criteria: correlation, activity, and dynamic significance. Here we use correlation threshold of 0.5, a standardized anomaly threshold of 0.5Z, and a mean precipitation threshold of 30 mm for these criteria. NMME locations a historical correlation of less than -0.5 or greater than +0.5 are retained.

We next screen for activity. Areas that have a standardized anomaly, in this season, of less than -0.5Z or greater than +0.5Z are retained. We then screen for ‘dynamic significance’; areas with long term average NMME precipitation of less than 30 mm are not used. Figure 6 shows an example of masking, based on the 2017 OND forecast process.

Figure 5. Correlation between observed 1982-2016 EEA OND precipitation and 1982-2016 NMME Precipitation forecast made in early October.
Figure 5. Correlation between observed 1982-2016 EEA OND precipitation and 1982-2016 NMME Precipitation forecast made in early October.
Figure 6. Mask used to predict the OND 2017 rains. Orange areas are positively correlated with observed EEA OND rainfall. Blue areas are negatively correlated.
Figure 6. Mask used to predict the OND 2017 rains. Orange areas are positively correlated with observed EEA OND rainfall. Blue areas are negatively correlated.

The next step in the AdReg process calculates the first and second principal components (PCs) of the NMME forecasts. As an example, Figure 7 shows the correlation between the first AdReg principal component and the NMME precipitation predictions. This is an ENSO signal. The final step uses regression to translate the PCs into forecasts of the predicted target.

Take-one-away cross-validation is used to assess the out-of-sample skill of the AdReg procedure. For each year, that year’s data is withheld, the AdReg process calculated and then forecast accuracy is evaluated.

It should be noted that selection processes can create artificial skill, even when using cross-validation. The AdReg process attempts to minimize this effect by working with the low order principal components of the NMME forecasts. In the NMME models, this often corresponds with ENSO-like patterns of variability. The results presented here tend to focus on ENSO-like patterns similar to those studied in FEWS NET diagnostic analyses (see papers below).

Figure 7. Correlation between AdReg PC1 and the OND NMME precipitation forecasts.
Figure 7. Correlation between AdReg PC1 and the OND NMME precipitation forecasts.

 

Results 1. The OND Outlook

We begin by presenting (Figure 8) results for OND 2017 based on a model with masking based on a historical correlation of |0.5|, as standard anomaly of at least ±0.5Z and a mean precipitation rate of 30 mm. This model performs fairly well (cross-validated correlation of 0.68 with standard error of 0.8Z). While some droughts are missed, most are captured fairly well, including droughts in 2010 and 2016. A brief sensitivity analysis (Table 1) indicates fairly robust performance across our model criteria, except for when we set the correlation value to a high value (0.7). These strict criteria likely induces over-selection which performs poorly under cross-validation. Also shown in table 1 are cross-validated correlation and standard error values based on the raw NMME OND forecasts over the study area. While the models have some skill (r=0.45), AdReg skills are higher. A correlation of 0.45 implies explaining ~20% of the variance, AdReg correlations (~0.68%) imply explaining 46% of the variance. Raw NMME forecasts, regressed against the target, are shown in Figure 8 (triangles). These results are based on the bias corrected NMME forecasts, more experimentation with the underlying NMME precipitation fields will be examined in later analyses.

The outlook for the OND EEA rains based on AdReg is -0.8Z ± 1Z, based on 80% confidence intervals. Below normal appears to be the most likely outcome.

Figure 8. AdReg forecasts for EEA OND rains. Cross-validated results for 1982-2016, and the 2017 forecast. Bars denote 80% confidence intervals. Also shown are estimated based on regression with the raw EEA OND NMME (triangles).
Figure 8. AdReg forecasts for EEA OND rains. Cross-validated results for 1982-2016, and the 2017 forecast. Bars denote 80% confidence intervals. Also shown are estimated based on regression with the raw EEA OND NMME (triangles).
Correlation Activity Mean Rainfall CV R CV Standard Err
|0.5| >±0.5Z > 30 mm 0.68 0.8Z
|0.7| >±0.5Z > 30 mm -0.4 1.4Z
|0.3| >±0.5Z > 30 mm 0.63 0.8Z
|0.5| >±0.5Z > 0 mm 0.70 0.7Z
|0.5| >±0.5Z > 100 mm 0.70 0.7Z
|0.5| >±0.0Z > 30 mm 0.69 0.7Z
|0.5| >±0.8Z > 30 mm 0.67 0.8Z
Raw EEA OND NMME Precipitation 0.45 0.9Z

Table 1. Sensitivity Analysis for the OND Rains.

Results 2. The MAM Outlook

We next examine the relationship between the NMME OND precipitation forecasts and historical MAM EEA precipitation. We use the past 20 years (1998-2017), because FEWS NET research has shown a much stronger teleconnection with La Niña following the 1997/98 El Niño event, when the western Pacific SSTs increased substantially. Interestingly, we see quite a strong correlation between the OND NMME precipitation forecasts and the observed MAM EEA rainfall in the following spring (Figure 9). This ENSO-like pattern likely indicates the likelihood of a persistent overturning circulation that continues to disrupt rainfall across the tropics. The magnitude of these correlations is quite high (~±0.7). The cross-validated AdReg correlation is high (+0.7), with a fairly low standard error (0.69Z). The associated 2018 forecast would be -0.6Z ± 0.9Z; a below normal outlook with considerable uncertainty. Figure 10 shows the scatterplot of cross-validated forecast results, along with the 2017 and 2018 MAM forecasts. Figure 10 shows the corresponding NMME correlation pattern for the MAM AdReg forecast; again a strong ENSO-like pattern (the PC1 time series is dominated by strong ENSO-like inter-annual variations).

Interpretation

These results suggest, in line with numerous studies, that La Niña-like conditions are associated with dry conditions over the easternmost section of equatorial East Africa. The NMME can successfully anticipate such conditions, and provides a very useful guide to anticipating East African hydrological extremes. By focusing on the models’ large scale representation of ENSO, AdReg can be used to produce skillful automated forecasts, in early October, for both the short and long rains. For the short rains, our level of confidence might be quite high, since we have already begun the season, rainfall to date is slow to start, we can see evidence of a La Niña-like transition in the observed circulation, and near-term weather predictions tend to be pessimistic, at least over large parts of East Africa. If October is dry, the overall season will likely be poor (Figure 11). Our level of confidence in the MAM 2018 outlook is lower, given the large span of time, and the good possibility that La Niña-like conditions may fade. On the other hand, our forecast of -0.6Z is also quite similar to the ‘New Normal’ implied in the long term decline of the EEA long rains (Figure 2). Given the combination of current La Niña-like conditions, the pessimistic but skillful AdReg results shown here, and general tendency for EEA MAM rains to be below normal in the absence of strong forcing from El Niño or Indian Ocean Dipole-like conditions, assumptions of below normal rains for both the short and long rains may be warranted. The MAM outlook value would hover at the edge of the normal and below normal tercile boundary.

Figure 9. Correlation between observed EEA MAM rains and NMME OND precipitation from the preceding October.
Figure 9. Correlation between observed EEA MAM rains and NMME OND precipitation from the preceding October.
Figure 10. AdReg forecasts for EEA MAM rains. Cross-validated results for 1982-2016, and the 2017 forecast. Bars denote 80% confidence intervals.
Figure 10. AdReg forecasts for EEA MAM rains. Cross-validated results for 1982-2016, and the 2017 forecast. Bars denote 80% confidence intervals.
Figure 11. Correlation between NMME OND Precipitation and PC1 for MAM EEA forecast.
Figure 11. Correlation between NMME OND Precipitation and PC1 for MAM EEA forecast.
Figure 12. EEA October and OND rainfall, 1981-2016.
Figure 12. EEA October and OND rainfall, 1981-2016.
Prior research

While not detailed here, the reader should be aware of numerous studies supported by USAID, the USGS and NASA, to inform decision making in context like this. This work seeks to better understand and anticipate EEA droughts.*

Williams P. and Funk C. (2010) A Westward Extension of the Tropical Pacific Warm Pool Leads to March through June Drying in Kenya and Ethiopia, USGS Openfile Report 1199, http://pubs.usgs.gov/of/2010/1199/pdf/ofr2010-1199.pdf

Williams P. and C. Funk (2011) A westward extension of the warm pool leads to a westward extension of the Walker circulation, drying eastern Africa, Climate Dynamics, V37.11-12, p. 2417-2435.  http://www.springerlink.com/content/u0352236x6n868n2/fulltext.pdf

Liebmann, B., Bladé I., Kiladis G. N. , Carvalho L. M. V., Senay, G., Allured D., Leroux S., Funk C (2011) African Precipitation Seasonality Based on Daily Satellite Data from 1996-2009, J. of Climate, Journal of Climate 25, no. 12 (2012): 4304-4322.

Hoell, A. and C. Funk (2013): The ENSO-related West Pacific Sea Surface Temperature Gradient, Journal of Climate. J. Climate, 26, 9545-9562. doi: http://dx.doi.org/10.1175/JCLI-D-12-00344.1

Hoell A. and C. Funk (2013) Indo-Pacific Sea Surface Temperature Influences on Failed Consecutive Rainy Seasons over Eastern Africa, Climate Dynamics, DOI: 10.1007/s00382-013-1991-6. p1-16.

Hoell, A., Funk, C., & Barlow, M. (2014). La Niña diversity and Northwest Indian Ocean Rim teleconnections. Climate Dynamics, 1-18.

Liebmann, B., Hoerling M., Funk C., Dole, R. M., Allured A., Pegion, P., Blade I., Eischeid, J.K, (2014) Understanding Eastern Africa Rainfall Variability and Change, Journal of Climate.

Funk, C., Hoell, A., Shukla, S., Bladé, I., Liebmann, B., Roberts, J. B., and Husak, G. (2014). Predicting East African spring droughts using Pacific and Indian Ocean sea surface temperature indices. Hydrology and Earth System Sciences Discussions, 11(3), 3111-3136.

Shukla, S., McNally, A., Husak, G., & Funk, C. (2014). A seasonal agricultural drought forecast system for food-insecure regions of East Africa. Hydrology and Earth System Sciences Discussions, 11(3), 3049-3081.

Shukla S., Funk C. and Hoell A. (2014) Using constructed analogs to improve the skill of March-April-May precipitation forecasts in equatorial East Africa , Env. Res. Letters, Environmental Research Letters 9.9 (2014): 094009.

Funk, C. and Hoell A. (2015) The leading mode of observed and CMIP5 ENSO-residual sea surface temperatures and associated changes in Indo-Pacific climate, J. Climate, 28, 4309-4329, http://journals.ametsoc.org/doi/abs/10.1175/JCLI-D-14-00334.1

Shukla S., Roberts, J., Hoell A., Funk C., Robertson F. and Kirtmann, B. (2016) Assessing North American Multimodel Ensemble (NMME) Seasonal Forecasts, Climate Dynamics. 1-17.

Funk, C and Hoell A. (2017) Recent Climate Extremes Associated with the West Pacific Warming Mode, AGU Monograph-Climate Extremes: Patterns and Mechanisms. Ed. by Simon Wang,  Jin-Ho Yoon, Chris Funk and Robert Gillies, published by Wiley Press 226 (2017): 165.

Liebmann B., Bladé I., Funk C., Allured D., Hoerling, M., Hoell, A., Peterson, P., Thiaw, M.T. (2017) Climatology and Interannual Variability of Boreal Spring Wet Season  Precipitation in the Eastern Horn of Africa and Implications for its Recent Decline,J. Climate, Journal of Climate 30.10 (2017): 3867-3886.

Hoell A, Hoerling M, Eischeid J, Quan W-X, Liebmann B (2017) Reconciling theories for human and natural attribution of recent East Africa drying. J Clim 30:1939-1957