+ All Categories
Home > Documents > Predictability of seasonal precipitation across major crop ... · Increasing availability in...

Predictability of seasonal precipitation across major crop ... · Increasing availability in...

Date post: 22-May-2020
Category:
Upload: others
View: 1 times
Download: 0 times
Share this document with a friend
12
Contents lists available at ScienceDirect Climate Services journal homepage: www.elsevier.com/locate/cliser Original research article Predictability of seasonal precipitation across major crop growing areas in Colombia Alejandra Esquivel a , Lizeth Llanos-Herrera a , Diego Agudelo a , Steven D. Prager a , Katia Fernandes a,b , Alexander Rojas c , Jhon Jairo Valencia d,e , Julian Ramirez-Villegas a,f, a International Center for Tropical Agriculture (CIAT), km 17 recta Cali-Palmira, 763537 Cali, Colombia b International Research Institute for Climate and Society, Columbia University, Palisades, NY, USA c Federación Nacional de Arroceros, Fedearroz, Carrera 100 No. 25H-55, Bogotá, Colombia d Federación Nacional de Cereales y Leguminosas, Cota, Cundinamarca, Colombia e Food and Agriculture Organization of the United Nations (FAO), Bogotá, Colombia f CGIAR Research Program on Climate Change, Agriculture and Food Security (CCAFS), c/o CIAT, km 17 recta Cali-Palmira, 763537 Cali, Colombia ARTICLE INFO Keywords: Seasonal forecast Predictability Sea surface temperatures Climate services Colombian agriculture ABSTRACT Agriculture is one of the sectors that has greatly benetted from the establishment of climate services. In Colombia, interannual climate variability can disrupt agricultural production, lower farmers' incomes and in- crease market prices. Increasing demand thus exists for agro-climatic services in the country. Fullling such demand requires robust and consistent approaches for seasonal climate forecasting. Here, we assess seasonal precipitation predictability and forecast skill at agriculturally-relevant timescales for ve departments that re- present key growing areas of major staple crops (rice, maize, and beans). Analyses use Canonical Correlation Analysis, with both observed SSTs and modeled (NCEP-CFSv2) SSTs, as well as with CFSv2 predicted pre- cipitation elds (through a Model-Output-Statistics analysis). Some 74.4% of the forecast situations analyzed (5 departments 4 seasons 3 predictors 3 lead times) showed correlation-based goodness index (Kendalls tau, τ ) values above 0.1, 38.8% above 0.2, and 18.8% above 0.3. Predictability was limited towards eastern Colombia, and during wet periods of the year in the Inter-Andean Valleys. Importantly, results were consistent between ERSST and CFSv2-driven forecasts, implying that both can oer valuable outlooks for Colombia. While our study is a rst important step toward the establishment of a sustainable and successful climate service for agriculture in Colombia, further work is required to (1) improve seasonal forecast skill; (2) link seasonal fore- casts to agricultural modelling applications; (3) design appropriate delivery means; and (4) establish stake- holder-driven processes that allow two-way communication between forecast issuing institutions (e.g. IDEAMColombian Meteorological Service) and famersorganizations and farming communities. Practical implications The present study assesses seasonal precipitation predict- ability and forecast skill at agriculturally relevant timescales in Colombia. Here, based on feedback from the Meteorological Service (IDEAM), we empirically dene suitable skill as that in which the correlation-based goodness index (Kendalls tau, τ ) is above 0.1, but also analyze situations with τ above 0.2 and 0.3. There are a number of practical implications that stem from this work. A rst major implication of our work is related to the nding that ca. 75% of the forecast situations analyzed have suitable forecast skill. Moreover, some 38.8% of these situations have τ values above 0.2, and 18.8% above 0.3. This nding, together with existing anecdotal evidence on seasonal forecast impact in Colombia (see https://ccafs.cgiar.org/ research/results/cracking-patterns-big-data-saves-colombian- rice-farmers-huge-losses#.WZNXlK2ZP4c) suggests that sea- sonal forecasts may be suitable for use by farming organiza- tions and farming communities for their decision-making. Both ex-ante and ex-post assessments of forecast use in agri- culture may help in building evidence on the impact and benets of seasonal forecast use for Colombian agriculture. This will ultimately help in establishing climate services for the agricultural sector. https://doi.org/10.1016/j.cliser.2018.09.001 Received 2 March 2018; Received in revised form 3 August 2018; Accepted 12 September 2018 Corresponding author at: International Center for Tropical Agriculture (CIAT), km 17 recta Cali-Palmira, 763537 Cali, Colombia. E-mail address: [email protected] (J. Ramirez-Villegas). Climate Services xxx (xxxx) xxx–xxx 2405-8807/ © 2018 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/BY/4.0/). Please cite this article as: Esquivel, A., Climate Services, https://doi.org/10.1016/j.cliser.2018.09.001
Transcript
Page 1: Predictability of seasonal precipitation across major crop ... · Increasing availability in wide-area climate datasets (Funk et al., 2015; Muñoz et al., 2010), open data policies

Contents lists available at ScienceDirect

Climate Services

journal homepage: www.elsevier.com/locate/cliser

Original research article

Predictability of seasonal precipitation across major crop growing areas inColombia

Alejandra Esquivela, Lizeth Llanos-Herreraa, Diego Agudeloa, Steven D. Pragera,Katia Fernandesa,b, Alexander Rojasc, Jhon Jairo Valenciad,e, Julian Ramirez-Villegasa,f,⁎

a International Center for Tropical Agriculture (CIAT), km 17 recta Cali-Palmira, 763537 Cali, Colombiab International Research Institute for Climate and Society, Columbia University, Palisades, NY, USAc Federación Nacional de Arroceros, Fedearroz, Carrera 100 No. 25H-55, Bogotá, Colombiad Federación Nacional de Cereales y Leguminosas, Cota, Cundinamarca, Colombiae Food and Agriculture Organization of the United Nations (FAO), Bogotá, Colombiaf CGIAR Research Program on Climate Change, Agriculture and Food Security (CCAFS), c/o CIAT, km 17 recta Cali-Palmira, 763537 Cali, Colombia

A R T I C L E I N F O

Keywords:Seasonal forecastPredictabilitySea surface temperaturesClimate servicesColombian agriculture

A B S T R A C T

Agriculture is one of the sectors that has greatly benefitted from the establishment of climate services. InColombia, interannual climate variability can disrupt agricultural production, lower farmers' incomes and in-crease market prices. Increasing demand thus exists for agro-climatic services in the country. Fulfilling suchdemand requires robust and consistent approaches for seasonal climate forecasting. Here, we assess seasonalprecipitation predictability and forecast skill at agriculturally-relevant timescales for five departments that re-present key growing areas of major staple crops (rice, maize, and beans). Analyses use Canonical CorrelationAnalysis, with both observed SSTs and modeled (NCEP-CFSv2) SSTs, as well as with CFSv2 predicted pre-cipitation fields (through a Model-Output-Statistics analysis). Some 74.4% of the forecast situations analyzed (5departments ∗ 4 seasons ∗ 3 predictors ∗ 3 lead times) showed correlation-based goodness index (Kendall’s tau,−

τ ) values above 0.1, 38.8% above 0.2, and 18.8% above 0.3. Predictability was limited towards easternColombia, and during wet periods of the year in the Inter-Andean Valleys. Importantly, results were consistentbetween ERSST and CFSv2-driven forecasts, implying that both can offer valuable outlooks for Colombia. Whileour study is a first important step toward the establishment of a sustainable and successful climate service foragriculture in Colombia, further work is required to (1) improve seasonal forecast skill; (2) link seasonal fore-casts to agricultural modelling applications; (3) design appropriate delivery means; and (4) establish stake-holder-driven processes that allow two-way communication between forecast issuing institutions (e.g.IDEAM–Colombian Meteorological Service) and famers’ organizations and farming communities.

Practical implications

The present study assesses seasonal precipitation predict-ability and forecast skill at agriculturally relevant timescalesin Colombia. Here, based on feedback from the MeteorologicalService (IDEAM), we empirically define suitable skill as that inwhich the correlation-based goodness index (Kendall’s tau,

τ )is above 0.1, but also analyze situations with

τ above 0.2 and0.3. There are a number of practical implications that stemfrom this work. A first major implication of our work is relatedto the finding that ca. 75% of the forecast situations analyzed

have suitable forecast skill. Moreover, some 38.8% of thesesituations have

τ values above 0.2, and 18.8% above 0.3. Thisfinding, together with existing anecdotal evidence on seasonalforecast impact in Colombia (see https://ccafs.cgiar.org/research/results/cracking-patterns-big-data-saves-colombian-rice-farmers’-huge-losses#.WZNXlK2ZP4c) suggests that sea-sonal forecasts may be suitable for use by farming organiza-tions and farming communities for their decision-making.Both ex-ante and ex-post assessments of forecast use in agri-culture may help in building evidence on the impact andbenefits of seasonal forecast use for Colombian agriculture.This will ultimately help in establishing climate services forthe agricultural sector.

https://doi.org/10.1016/j.cliser.2018.09.001Received 2 March 2018; Received in revised form 3 August 2018; Accepted 12 September 2018

⁎ Corresponding author at: International Center for Tropical Agriculture (CIAT), km 17 recta Cali-Palmira, 763537 Cali, Colombia.E-mail address: [email protected] (J. Ramirez-Villegas).

Climate Services xxx (xxxx) xxx–xxx

2405-8807/ © 2018 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/BY/4.0/).

Please cite this article as: Esquivel, A., Climate Services, https://doi.org/10.1016/j.cliser.2018.09.001

Page 2: Predictability of seasonal precipitation across major crop ... · Increasing availability in wide-area climate datasets (Funk et al., 2015; Muñoz et al., 2010), open data policies

Secondly, Canonical Correlation Analysis (CCA), im-plemented through the Climate Predictability Tool (CPT), iswidely used for forecasting due to its low computing re-quirements and robustness in producing skillful seasonalforecasts. Moreover, CPT is a user friendly and publiclyavailable tool, and hence it has become widely used byNational Meteorological Centers and institutions that rely onclimate forecast for specific sectors, such as agriculture. At theIDEAM (Colombian Meteorological Service), CPT is used aspart of the approach to produce operational seasonal fore-casts. On a monthly basis, IDEAM’s Sub-direction ofMeteorology performs CPT runs for all regions of Colombiausing observed SSTs with a zero-month lead time (e.g.November SSTs are used to forecastDecember–January–February precipitation), which is one ofthe model configurations we assess herein. Information on theskill of seasonal forecasts produced by CPT that is specificenough to be used to inform agriculture is, however, limited.This study presents a thorough evaluation of seasonal pre-cipitation forecast in regions of rice, maize and bean cultiva-tion in Colombia, which provides users of both CPT and sea-sonal forecasts with important insight and guidance onstrengths and limitations of seasonal forecasts in the country,which can serve to further improve forecast models.Moreover, the results can also act as a benchmark againstwhich any such improvements are compared, or with whichother forecast systems can be compared.

A final implication of our work is related to the inputweather data. Firstly, low data quality is found for theCasanare department, in addition to low weather stationdensity. Higher density in the weather station network is de-sirable for a better understanding of regional and local cli-mates (Ramirez-Villegas and Challinor, 2012). Since it may bepossible that seasonal forecast model skill improves with ahigher density network, a direct implication of our studywould be for the country to enhance the weather stationnetwork across the Eastern Plains. Regarding data quality, wehave implemented a novel data gap filling algorithm by bias-correcting the Climate Hazards Infra-Red Precipitation withStations (CHIRPS) dataset (Funk et al., 2015) (see Section2.2). Even though CHIRPS already includes some weatherstations, most of the ones included in our study are not part ofthe dataset. The method implemented here for bias-correctingCHIRPS data worked consistently well across the weatherstations used in this study, which cover a wide range of geo-graphies. For IDEAM or other institutions currently main-taining weather stations, the data gap filling method used herecould be of use.

1. Introduction

Increasing availability in wide-area climate datasets (Funk et al.,2015; Muñoz et al., 2010), open data policies in national meteor-ological agencies, and open source prediction tools (Hansen et al.,2011; Stainforth et al., 2005), have resulted in a proliferation of locallytailored climate services (Fraisse et al., 2006; Georgeson et al., 2017;Yang et al., 2016). Due to heavy reliance on favorable weather andclimate conditions (Ray et al., 2015), agriculture is one of the sectorsthat has greatly benefitted from the establishment of such climate ser-vices, in part through the development of seasonal forecasts that sup-port improved decision-making by agricultural producers (Hansenet al., 2006; Jones et al., 2000; Patt and Gwata, 2002). One way inwhich seasonal forecasts support farmers’ decisions is through im-proved understanding of how expected seasonal conditions may

influence crop growth and development and, by extension, whatplanting dates, varieties, and other management practices will offer thegreatest potential yields (Capa-Morocho et al., 2016; Hammer et al.,1996). The determination of this type of information depends on skillfulseasonal forecasts.

Climate services in agriculture require the establishment of rigorousapproaches to ensure that the quality of the seasonal forecasts beinggenerated is as high as possible. This is especially true where associa-tions between climate prediction variables are often complicated bylocal geographies and high regional complexity (Poveda, 2004). InColombia, interannual climate variations and especially those driven bythe El Nino Southern Oscillation (ENSO) can disrupt production, lowerfarmers' incomes and increase market prices (CIAT-MADR, 2015; Iizumiet al., 2014). For instance, Delerce et al. (2016) report that some30–60% of rice yield variation in Meta and Tolima departments can beexplained by seasonal climate variability, with an important role forprecipitation. Similarly, Cortes Bello et al. (2013), using the AquaCropmodel (Steduto et al., 2009), reported that maize growth and devel-opment in the Cordoba, Meta, Tolima and Valle del Cauca departmentsare affected by the El Nino Southern Oscillation (ENSO) phenomenon aswell as by the Inter-Tropical Convergence Zone (ITCZ) atmosphericsystem. In Colombia, there is increasing demand for agro-climatic ser-vices and, scientifically robust and consistent approaches to im-plementing the underpinning seasonal climate forecasts. Notably,farmers require making decisions on whether to plant, when to plant,and what to plant, which could usefully be informed by 3–6monthoutlooks (Dorward et al., 2015; Tall et al., 2018). Furthermore, whenconnected to crop simulation tools, seasonal forecasts can be used toassess implications for crop productivity, as well as to guide cropmanagement decisions (Capa-Morocho et al., 2016).

In Colombia, as in many other Latin American countries [see e.g.Recalde-Coronel et al. (2014); Montecinos et al. (2000)], seasonalforecasting is typically exercised using a statistical technique calledCanonical Correlation Analysis [CCA; Hotelling (1936); Glahn (1968)],implemented through the Climate Predictability Tool (CPT) softwarepackage (Mason and Tippett, 2017). The development of a robust sea-sonal climate forecast is a multi-part process and is dependent on anumber of key inputs and methodological steps. These include (1) thedefinition of a conceptual model that links the local climate conditions(often precipitation) and the large-scale atmospheric or oceanic forcingpatterns; (2) the formal establishment (i.e. fitting) of such modelthrough CCA; (3) the validation of the model's accuracy; and (4) the re-adjustment of the conceptual model (e.g. adjustment of predictor do-main) based on model performance, expert knowledge and insights onthe ongoing conditions in the area of interest and their relationship tolarge-scale circulation patterns.

Operational national-scale seasonal climate forecasts in Colombiaare generated primarily by the national meteorological service(Institute of Hydrology, Meteorology and Environmental Studies,IDEAM) for large geographic areas known as ‘natural regions’ (i.e.Amazonas, Eastern Plains, Andean, Caribbean and Pacific). At morelocal levels, farmer organizations, namely, the rice growers association(FEDEARROZ) and the maize and legumes growers association(FENALCE), are also producing operational seasonal climate forecasts.The latter forecasts are of particular relevance since they currently feedinto decision making processes at the farm level (CIAT-MADR, 2015).Despite progress at these different institutions, to date, no analysisexists that shows how predictable seasonal climate conditions are, orwhether predictability and seasonal forecast model skill are sufficient tosupport decision making in agriculture.

In this paper, we use CCA to assess local seasonal precipitationpredictability and forecast skill using both observed [ExtendedReconstructed SST version 4, Huang et al. (2015)] and modeled [Na-tional Centers for Environmental Prediction Climate Forecast Systemversion 2, NCEP CFSv2, Saha et al. (2014)] sea surface temperature(SST) data, as well as CFSv2 modeled regional-scale precipitation. We

A. Esquivel et al. Climate Services xxx (xxxx) xxx–xxx

2

Page 3: Predictability of seasonal precipitation across major crop ... · Increasing availability in wide-area climate datasets (Funk et al., 2015; Muñoz et al., 2010), open data policies

focus on local precipitation as it is one of the main drivers of agri-cultural yields in the region and is, likewise, the variable with thegreatest inter-annual variation. Analyses are conducted separately forfive departments (Cordoba, Santander, Valle del Cauca, Tolima, andCasanare) with major growing areas for cereals (rice, maize) and le-gumes (dry beans), and for four seasons that represent wet and dryperiods (DJF: December–January–February, MAM: March–April–May,JJA: June–July–August, SON: September–October–November). ForSSTs, we explore the use of a standard tropical region (30°S–30°N),whereas for modeled NCEP-CFSv2 precipitation we use a region aroundColombia (Recalde-Coronel et al., 2014). The latter choice of area isbecause in using precipitation we are interested in downscaling andbias-correcting the CFSv2 precipitation field through a Model-Outputs-Statistics (MOS) analysis. Using two different training and holdoutstrategies (i.e. cross-validation and retrospective evaluation), we com-pare forecasts generated using the aforementioned observed and si-mulated SST relative to different holdout subsets.

In the following section on materials and methods, we describe thestudy area, input data and models used. In the results section, we firstpresent the results of the predictability analysis using observed SSTs,then compare observed SST results with NCEP-CFSv2 SST and pre-cipitation model results, and finally investigate changing predictiveskill across lead times. The results are finally discussed in light of ex-isting research and future prospects for climate information services inColombia.

2. Materials and methods

Developing a seasonal climate forecast is a multi-step process. Themajor steps include identification of the study areas (Section 2.1), thecollection and cleaning of input data (Sections 2.2, 2.3, 2.4), the es-tablishment of experimental design with respect to the relationship andlags between predictor and response variables including the develop-ment of a strategy for training and validation of the developed model(Sections 2.5 and 2.6).

Fig. 1 shows the detailed workflow for the analysis we conductedhere. The first step of the analysis is to gather data. Input data includesobserved precipitation provided by IDEAM, the Extended Re-constructed SST version 4 (Huang et al., 2015), and the various en-sembles of NCEP CFSv2 (Saha et al., 2014). IDEAM meteorologicalstation data was quality controlled and gap filled (see Section 2.2).Predictor variable data was downloaded for specific lead times andseasons (see Sections 2.3 and 2.4). Next, prediction models are thendeveloped using Canonical Correlation Analysis (CCA) within the Cli-mate Predictability Tool (Mason and Tippett, 2017), and evaluated intwo ways (cross-validation and retrospective validation). These twoprocesses allow the calculation of performance indicators to illustratethe overall predictive skill associated with the modeling approach. Wedeveloped CCA models using observed SSTs (ERSST), NCEP-CFSv2hindcast SSTs and regional precipitation as predictors of local quarterly

precipitation for four different lead times (simultaneous, 1-month, 3-month, and 5-month). Our choice of predictor variables, domains andlead times is based on the following rationale: (1) observed SSTs areused as part of the operational system at the National MeteorologicalService (IDEAM, see Section Practical Implications); (2) in regions withsubstantial ENSO influence such as Colombia, SSTs (both observed andpredicted) often deliver the greatest forecast skill at seasonal timescales(Goddard et al., 2001; Torrence and Webster, 1998; Zebiak et al.,2015); and (3) regional precipitation downscaling through a ModelOutput Statistics (MOS) analysis can improve CCA-based model pre-dictive skill as compared to models that use SSTs (Landman andGoddard, 2002; Recalde-Coronel et al., 2014).

2.1. Study areas

Analyses were conducted in five departments in Colombia, namely,Cordoba, Valle del Cauca, Tolima, Casanare and Santander (Fig. 2). Wefocus on these departments because they are major producers of keystaple crops for Colombia (rice, maize, dry beans). Casanare and Tolimaare the top two departments in harvested area and production of rice,alone accounting for 48.2% and 47.1% of national harvested area andproduction, respectively (Fedearroz, 2017). Cordoba and Valle delCauca are the top producers of maize (DANE, 2014), whereas Santanderis usually within the top-two bean producers (DANE, 2014). Croppingsystems in these regions are highly varied, ranging from large scale,highly technologically developed and irrigated rice systems in some ofthe major valleys, through medium scale producers of maize in bothvalley and moderately hilly areas, to small scale producers of beans,sometimes in very remote, steep hillsides.

The five departments also represent the three major ecoregionswhere cropping is highly prevalent in Colombia: Andean (Valle delCauca, Tolima and Santander), Eastern Plains (Casanare) andCaribbean (Cordoba). The seasonal rainfall patterns also vary across thedifferent regions, resulting in different prediction requirements. Theother two natural regions (Amazonas and Pacific) are mostly forestedareas with very limited farming compared to the Andean, Eastern Plainsand Caribbean regions.

The five departments have very distinct annual precipitation pat-terns and amounts. Precipitation in Casanare and Cordoba is unim-odally distributed, with the rainy period being April–October, andprecipitation amounts of 250–400mmmonth−1 for Casanare and120–200mmmonth−1 for Cordoba. Tolima, Valle del Cauca andSantander, on the other hand, have a bimodal pattern, with rainy sea-sons in March-April-May (MAM) and September-October-November(SON).

2.2. Meteorological data

Observed daily precipitation data for the period 1982–2013 weregathered from the meteorological station network of IDEAM, mostly

Fig. 1. Overview of the process used for understanding seasonal precipitation predictability and forecast skill.

A. Esquivel et al. Climate Services xxx (xxxx) xxx–xxx

3

Page 4: Predictability of seasonal precipitation across major crop ... · Increasing availability in wide-area climate datasets (Funk et al., 2015; Muñoz et al., 2010), open data policies

representative of agricultural and urban landscapes (see Fig. 1).Weather station data, however, often contain missing or wrongly re-ported values (Ramirez-Villegas and Challinor, 2012; Van Wart et al.,2015). Therefore, data quality control and gap-filling algorithms comehand-in-hand with the implementation of all the analyses. We qualitycontrolled the observed meteorological data using the RClimTool soft-ware [Llanos-Herrera (2014); http://www.aclimatecolombia.org/rclimtool-free-application-analyzing-climatic-series/]. The quality con-trol consisted of three filters aimed at flagging and removing wrongly

reported values, as follows:

(1) Range check (filter 1): precipitation below zero or above350mmday−1

(2) Outliers (filter 2): data values greater than the third quartile (Q3)plus five times the inter-quartile range (IQR), or lower than the firstquartile (Q1) minus five times the IQR

(3) Constant values (filter 3): more than 3 non-zero equal consecutivevalues in the whole series

Fig. 2. Target areas for analysis. Point locations in the maps are the weather stations used in the analyses. The barplots show the 1982–2013 climatological monthlymeans as averages of all weather stations within each department.

A. Esquivel et al. Climate Services xxx (xxxx) xxx–xxx

4

Page 5: Predictability of seasonal precipitation across major crop ... · Increasing availability in wide-area climate datasets (Funk et al., 2015; Muñoz et al., 2010), open data policies

Values that meet any of the three conditions are considered un-realistic since they are unlikely to follow or are likely outside theprobability distribution of precipitation. Table 1 shows the number ofstations by department used after filtering, and their quality controlindicators. Once all meteorological stations were quality controlled, weproceeded to filling gaps in the monthly precipitation time series. Onlystations with less than 10% total gaps in the period 1982–2013 wereselected for analysis, except for those stations in Casanare, for which amaximum of 15% gaps was allowed due to overall lower data avail-ability in that department.

Several data gap filling methodologies exist (Eischeid et al., 2000;Harktamp et al., 2000; Tsidu, 2012). The performance of these methodsis limited for tropical and sub-tropical regions in which rainfall ismostly of convective nature (Acock and Pachepsky, 2000; Pickeringet al., 1994). Novel methods that use satellite information offer skillfulalternatives to previously existing methods (Dinku et al., 2014;Huffman et al., 1995). Here, we used a satellite-based precipitation datagap-filling methodology. The methodology combines meteorologicalstation data with the Climate Hazards Infrared Precipitation with Sta-tions [CHIRPS, Funk et al. (2015), freely available at ftp://ftp.chg.ucsb.edu/pub/org/chg/products/CHIRPS-2.0] dataset, to produce gap-filledmeteorological station records at the monthly scale. CHIRPS is a quasi-global precipitation dataset developed using a combination of satelliteand rain gauge data (see Funk et al., 2015 for more details). Wedownloaded CHIRPS data for the period 1982–2013 at 0.05° resolution,and fitted a linear model using the weather station data at the monthlyscale as the dependent variable and the distance-weighted average ofthe closest four CHIRPS pixels to the meteorological station in questionas the independent variable. Two procedures are performed. First, amodel using all data is first constructed per weather station to estimatemissing station data. The R2 coefficient of these models is used as ameasure of performance. Secondly, we perform a leave-out cross-vali-dation for each station to check the robustness of the gap filling models(see Supplementary Table S1).

2.3. Observed sea surface temperature data

Sea Surface Temperature (SST) data were obtained from theNational Oceanic and Atmospheric Administration (NOAA) ExtendedReconstructed Sea Surface Temperature Dataset version-4 (Huang et al.,2015; Liu et al., 2015) (ERSST hereafter), freely available through theInternational Research Institute for Climate and Society (IRI) data li-brary (at http://iridl.ldeo.columbia.edu/SOURCES/.NOAA/.NCDC/.ERSST/.version4/.sst/). The monthly ERSST data are distributed at 2-degree spatial resolution and available from January 1854 to currenttime at monthly timescale. November, February, May and Augustmonthly ERSSTs were used as predictors (see experiment description,Sections 2.5 and 2.6) of local seasonal precipitation. In addition, theERSST averaged over the trimesters DJF, MAM, JJA and SON were alsoused as predictors to precipitation in concurrent seasons.

2.4. NCEP climate forecast system version 2 data

We used predicted SST and regional precipitation fields from the

NCEP CFSv2 (Saha et al., 2014) as predictors to the selected Colombiandepartments seasonal precipitation (see Section 2.6). The NCEP-CFSv2(CFSv2 hereafter) model retrospective forecast is available from Jan-uary 1982 to March 2011, whereas the real-time forecast is availablefrom April 2011 to current. The NCEP-CFSv2 model produces predic-tions with a 9-month lead time, and with four initial conditions(starting at 0000, 0600, 1200, and 1800 h UTC), every fifth day. Thisresults in 24 ensemble members for each month, except for November,for which there are 28 members. For a more complete description of theNCEP-CFSv2 forecast system the reader is referred to Saha et al. (2014).Here, we use the ensemble mean for retrospective and real-time fore-casts from the CFSv2. These are freely available through the IRI DataLibrary at http://iridl.ldeo.columbia.edu/SOURCES/.NOAA/.NCEP/.EMC/.CFSv2/.ENSEMBLE/.

2.5. Assessment of seasonal precipitation predictability limit, and observedSST-based predictive analyses

Canonical Correlation Analysis (CCA) is a multivariate linear sta-tistical technique used to generate seasonal climate predictions, both ina deterministic and probabilistic manner (Glahn, 1968; Goddard et al.,2001; Hotelling, 1936). Probabilistic forecasts are typically expressed interciles (below normal, normal, above normal), whereas deterministicones are expressed as a single value with a confidence interval. CCAaims at finding the best relationships and correlations between two setsof data, with each set of data containing many variables. In seasonalforecast models, it is used to exploit the teleconnection between remoteareas (e.g. in the ocean) and the local climate conditions. CCA modelsfor seasonal forecasting can use either SSTs or any atmospheric variablethat can be related to local precipitation. CCA has been used widely inpredictability studies in Latin America and elsewhere [e.g. Montecinoset al. (2000); Korecha and Barnston (2007); Diaz et al. (1998)].

Here, the development of a CCA-based seasonal forecast modelfollows a series of steps. First, predictors (e.g. SSTs) and predictands(e.g. local precipitation) are separately pre-filtered through a principalcomponent analysis (PCA), also referred to as an Empirical OrthogonalFunction (EOF) analysis. As for most climate patterns only a few EOFscan capture most of the entire variability, only few EOFs are selected forthe predictor and the predictand. These resulting patterns after pre-filtering are used as inputs into a CCA. Once the canonical componentsare determined, a linear regression is performed between the canonicalmodes of the predictor and the predictand, including a cross-validationand a retrospective validation procedure. Here, we implement CCAmodels through the Linux version of the Climate Predictability Tool(CPT, Mason and Tippett, 2017). CPT is widely used by meteorologicalservices globally, as it provides an efficient and user-friendly im-plementation of CCA-based seasonal forecast models (Mason, 2011;Muñoz et al., 2010; Recalde-Coronel et al., 2014). CPT also connectsdirectly with the IRI data library, which facilitates data download,processing and use into the CCA models implemented here.

Using CCA, two types of analyses can be performed. The diagnosticanalysis, in which the predictor and the predictand are from the sameperiod (e.g. January to predict January). The second type is calledpredictive analysis, in which the predictor and the predictand arelagged in time (e.g. January to predict March). The diagnostic casemaximizes the relationship between predictors and predictands, and,when observed SSTs are used it represents the upper limit of predict-ability for the SST as predictor. In the predictive analysis, one relies onthe persistence of the SST in its teleconnection with the local climateconditions, and hence skill is generally lower (Recalde-Coronel et al.,2014). In some cases, due to remote teleconnection, the predictiveanalysis can produce greater forecast skill as compared to the diagnosticdue to lags in the response of the climate system.

ERSST analyses here are conducted for different lead times, chosento represent both diagnostic (predictability limit) and predictive ana-lyses (to assess SST persistence). Lead time refers to the time difference

Table 1Number of stations per department, and total percentage of missing data.

Department Number ofstations

Average initialmissing data(%)

Average dataflagged atfilters (%)

Totalmissing data(%)

Valle del Cauca 27 5.0 0.0 5.0Tolima 34 5.0 1.0 6.0Santander 76 3.0 0.0 3.0Cordoba 34 6.3 0.7 7.0Casanare 9 11.8 0.2 12.0

A. Esquivel et al. Climate Services xxx (xxxx) xxx–xxx

5

Page 6: Predictability of seasonal precipitation across major crop ... · Increasing availability in wide-area climate datasets (Funk et al., 2015; Muñoz et al., 2010), open data policies

(in months) between month of issue of the predictor variable (in thiscase the ERSST) and the month of the observations. Due to the persis-tence of ERSSTs, a maximum lead time of 3months is typically possible.We implement CCA models for three lead times, namely, simultaneous(LT-Sim, predictability limit), 0-month lead (LT-0) and 3-month lead(LT-3) (see Table 2 for model configuration). We implement all analysesseparately for four standard seasons, namely, December–Januar-y–February (DJF), March–April–May (MAM), June–July–August (JJA),September–October–November (SON), that are representative for thedry and rainy periods in all target study areas (see Fig. 1). Table 2shows predictor configuration for the four seasons analyzed.

All models are fitted using the wide global tropics (30°S–30°N) asthe domain for the predictor. For SSTs, this choice is a compromisebetween having a domain that covers all areas in the tropics that havean effect on the precipitation in Colombia, while also excluding theextratropics, which are less relevant for rainfall prediction in tropicalSouth America (Córdoba-Machado et al., 2015; Montecinos et al., 2000;Recalde-Coronel et al., 2014). Since all analyses focus on statisticalmodels, which could potentially suffer from over-fitting, all models areassessed using both cross-validation and retrospective validation (seeSupplementary Text S1 for details). We use the spatially-averagedGoodness Index (Kendall correlation,

τ hereafter) as the main perfor-mance metric. We use

τ as it is a suitable indicator for forecasts that are

of probabilistic nature, and is unlikely to be affected by outlying values.

2.6. Predictive analysis using NCEP CFSv2 fields

We conducted analyses using CFSv2 SSTs forecasts, as well as withregional precipitation predictions. These are used for the diagnosticdesign at lead times commensurate with those needed for agriculturaldecision making (4–6months, typically) (Davey and Brookshaw, 2011;Hansen et al., 2011). Analyses with CFSv2 SSTs were conducted simi-larly to those of ERSST, though with different lead times, chosen toallow for assessing the skill in predictive situations. Specifically, we usea 0-month (LT-0), 3-month (LT-3), and a 5-month (LT-5) lead time. Thismeans that, for instance, for a prediction of MAM, LT-0 corresponds tothe CFSv2 SST prediction of MAM with initial conditions of February,LT-3 with initial conditions of November, and LT-5 to September.Table 2 shows predictor configuration for all four seasons and threelead times analyzed for CFSv2.

As stated above, CFSv2 regional precipitation predictions were alsoused. Following Recalde-Coronel et al. (2014), we conducted a ModelOutput Statistics (MOS) analysis using predicted precipitation from theCFSv2 model. In MOS, instead of generating a statistical model basedon the teleconnection between a largescale predictor and a local pre-dictand, the CCA both statistically corrects the systematic bias of theCFSv2 model and downscales the prediction (Ndiaye et al., 2011,2009). The selected predictor domain is a rectangle between latitudes16°N–16°S and longitudes 90°W–64°E; this includes Colombia but alsoneighboring countries, which may be relevant as sometimes the modelcan misplace rainfall areas of interest. As with ERSST analyses, weconduct both cross-validation and retrospective validation, and use

τ asthe performance metric.

3. Results

3.1. Predictability limit in seasonal precipitation

Fig. 3 presents the results for the predictability limit (that is, leadtime LT-Sim using ERSST), for each department analyzed. Several im-portant findings become apparent. Foremost, despite variation in

τacross departments, results from both validation procedures (cross-va-lidation and retrospective validation) are consistent, suggesting modelsare not overfitted. Secondly, we note that the Casanare departmentshows the lowest

τ values, perhaps due to the low density of the me-teorological stations (Fig. 2), the more limited data quality in this de-partment (Table 1), a lack of physical plausibility in the captured

Table 2Input data used into CCA models per lead time.

Input predictordataset

Predictedseason

Predictor per lead time1

LT-Sim2 LT-0 LT-3 LT-5

ERSST DJF DJF Nov Aug –MAM MAM Feb Nov –JJA JJA May Feb –SON SON Aug May –

CFSv2 (SST,precipitation)

DJF – DJF_(Nov)

DJF_(Aug)

DJF_(Jun)

MAM – MAM_(Feb)

MAM_(Nov)

MAM_(Sep)

JJA – JJA_(May)

JJA_(Feb)

JJA_(Dec)

SON – SON_(Aug)

SON_(May)

SON_(Mar)

1 LT-Sim corresponds to simultaneous, LT-0 to a zero-month lead time, LT-3to a 3-month lead time, and LT-5 to a 5-month lead time. Months in brackets forCFSv2 predictions indicate the month in which the model run is initialized.

2 LT-Sim refers to the diagnostic situation.

0.0

0.1

0.2

0.3

0.4

Casanare Cordoba Santander Tolima Valle del Cauca

Mea

n Ke

ndal

l cor

rela

tion

()

Cross validation

Retrospective validation

Fig. 3. Barplot of the goodness index (Kendall correlation,−

τ ) by department. Only the LT-Sim (ERSST) is shown, as it depicts the predictability limit. Error barscorrespond to the

τ minimum and maximum across seasons.

A. Esquivel et al. Climate Services xxx (xxxx) xxx–xxx

6

Page 7: Predictability of seasonal precipitation across major crop ... · Increasing availability in wide-area climate datasets (Funk et al., 2015; Muñoz et al., 2010), open data policies

association between local precipitation and SSTs, or a combination ofall these. When the CCA model shows a moderate or high goodnessindex, the loads for the canonical modes in the CCA model show what isnormally expected for a tropical region (i.e. tropical Pacific SSTs beingmost important, see Supplementary Figs. S1–S4). Where the converse istrue (i.e. low goodness index), the SST signal is weaker and therefore itis evident that there are other non-SST influences that would explainprecipitation variability better in Casanare.

The rest of the departments show−

τ values generally above 0.1,suggesting that predictability is sufficient for generating seasonalforecasts. Cordoba (Caribbean region), shows lower predictability ascompared to the departments in the Andean region (Tolima, Santanderand Valle del Cauca). Valle del Cauca, on the contrary, shows thegreatest seasonal predictability. This is mostly attributed to the de-partment having a coast in the Pacific Ocean, and thus being moreinfluenced more directly by Pacific SSTs [Córdoba-Machado et al.(2015); also see Supplementary Figs. S5–S8]. Nevertheless, significantprecipitation variability remains unexplained, primarily owing to thecomplex topographic and landscape features of the department (twomountain chains with a semi-arid valley in between, and a tropicalrainforest to the west).

It is noteworthy that Andean departments show seasons with bothhigh and low

τ values. This is due to the bimodal nature of the pre-cipitation in these departments. The climatology of these regions isrepresented by two rainy and two dry periods. In these cases, thegreatest predictability is found in the dry seasons (DJF, JJA), and thelowest predictability occurs in the wet seasons (MAM, SON). Fig. 4 il-lustrates this more clearly, with MAM (first rainy season) and SON(second rainy season) showing the least overall predictability; and DJFand JJA (dry periods) showing the greatest predictability.

In particular DJF shows the greatest−

τ average in the cross-valida-tion ( =

τ 0.325) for Andean region departments. Conversely, theCasanare department shows lower predictability in this season( =

τ 0.03), which is the driest in Casanare. We also note that, forCasanare, the greatest predictability is found in the MAM season, whichcorresponds to the transition between the main dry (DJF) and wet (JJA)seasons.

3.2. Performance of ERSST predictions

As stated above, we developed predictive CCA models using ERSST;that is, models that reflect a ‘real’ forecasting situation in which SSTsare lagged with respect to the local observations and then used toconstruct the CCA model. As expected, the performance of the model,

measured using the mean Kendall correlation (−

τ ), decreases as the leadtime increases, reflecting the expected natural decrease in persistence ofthe SSTs (Torrence and Webster, 1998) (Fig. 5). The departments inwhich model performance decreased most with increased lead timewere Valle del Cauca, Tolima and Santander, particularly for the JJAseason (see Fig. 6).

In addition, there is a tendency for longer lead times to show low−

τvalues. This was, however, not always the case. There are cases inwhich

τ increases with increased lead time. This may be due to a laggedeffect from SSTs on the seasonal precipitation (Córdoba-Machado et al.,2015) (Fig. 6). This is particularly so for the Cordoba and Casanaredepartments for the DJF season. For Cordoba,

τ values were 0.33 forthe simultaneous lead time (LT-Sim) and 0.36 for 0-month lead time(LT-0). In Casanare,

τ values were 0.03 (LT-Sim) and 0.11 (LT-0).

3.3. Performance of seasonal forecasts driven by NCEP-CFSv2

At the seasonal scale, agricultural decision making often requirespredictions with longer than 3-month lead time. This is important incases in which an end-of-season yield outlook is necessary to inform the

0.0

0.1

0.2

0.3

0.4

DJF MAM JJA SON

Mea

n Ke

ndal

l cor

rela

tion

()

Cross validation

Retrospective validation

Fig. 4. Barplot of the goodness index (Kendall correlation,−

τ ) by season. Only the LT-Sim (ERSST) lead time is considered here, as it corresponds to the predictabilitylimit. Error bars correspond to the

τ minimum and maximum across departments.

0.0

0.1

0.2

0.3

0.4

0.5

Cross validation Retrospective validation

Kend

all c

orre

latio

n (

)

Lead TimeLT Sim

LT 0

LT 3

Fig. 5. Boxplot of the goodness index (Kendall correlation,−

τ ) for different leadtimes for CCA models using ERSST, for both cross-validation and retrospectivevalidation procedures. Thick horizontal line indicates the median, boxes extendto the interquartile range, and whiskers extend to 5% and 95% of the dis-tribution. Lead times are specified in Table 2.

A. Esquivel et al. Climate Services xxx (xxxx) xxx–xxx

7

Page 8: Predictability of seasonal precipitation across major crop ... · Increasing availability in wide-area climate datasets (Funk et al., 2015; Muñoz et al., 2010), open data policies

decision, the crop's growing cycle is longer than 3months, or the de-cision needs to be made several months in advance (Capa-Morochoet al., 2016; Fraisse et al., 2006; Hansen and Indeje, 2004). Analysis ofthe CFSv2 model included up to a 5-month lead time, which is com-mensurate with the duration of the growing cycles of rice, maize andbeans (Fig. 7).

In general, we find that the skill of SST-based forecasts is generallysuperior to that of precipitation-based ones (Figs. 6 and 7). The

variation of−

τ across seasons for CFSv2-based forecasts is also lowerthan that of ERSST models (both diagnostic and predictive). Ad-ditionally, the performance of the CFSv2 forecast is consistent acrosslead times. This is particularly evident for the MOS analysis with CFSv2precipitation, for which not only the median

τ across seasons staysrelatively constant or increases, but also the likelihood of very low

τvalues reduces substantially (Fig. 7). The converse occurs for predictedSSTs, for which the longer lead time (5-month) consistently shows

0.03 0.11 0.12

0.33 0.36 0.37

0.27 0.29 0.25

0.44 0.46 0.4

0.26 0.28 0.24

0.25 0.25 0.22

0.1 0.12 0.13

0.12 0.01 0.06

0 0.08 0.03

0.12 0.08 0.06

0.07 0.1 0.02

0.18 0.1 0.11

0.4 0.35 0.31

0.43 0.37 0.31

0.3 0.24 0.19

0.02 0.04 0.12

0.15 0.14 0.14

0.13 0.15 0.15

0.16 0.16 0.07

0.11 0.12 0.1

ERSST

DJF

MA

MJJA

SO

N

LT Sim LT 0 LT 3

Casanare

Cordoba

Santander

Tolima

Valle del Cauca

Casanare

Cordoba

Santander

Tolima

Valle del Cauca

Casanare

Cordoba

Santander

Tolima

Valle del Cauca

Casanare

Cordoba

Santander

Tolima

Valle del Cauca

Lead Time

0 0.01 0.04

0.32 0.31 0.3

0.26 0.22 0.16

0.42 0.39 0.32

0.29 0.26 0.28

0.13 0.31 0.22

0.14 0.11 0.1

0.06 0.13 0.15

0.08 0.12 0.12

0.07 0.04 0

0.02 0.05 0.03

0.13 0.1 0.18

0.36 0.22 0.16

0.35 0.27 0.23

0.32 0.1 0.14

0.09 0.1 0.06

0.1 0.07 0.18

0.18 0.18 0.11

0.16 0.23 0.14

0.16 0.17 0.12

0.22 0.12 0.03

0.41 0.4 0.36

0.31 0.28 0.26

0.46 0.44 0.38

0.3 0.29 0.28

0.27 0.29 0.27

0.1 0.21 0.23

0.07 0.06 0.08

0 0.08 0.07

0.12 0.12 0.12

0.03 0.04 0.08

0.22 0.16 0.16

0.42 0.31 0.2

0.46 0.34 0.16

0.34 0.23 0.13

0.06 0.11 0.09

0.12 0.11 0.07

0.16 0.22 0.14

0.14 0.17 0.15

0.1 0.16 0.09

CFSv2 recipitation CFSv2 SST

DJF

MA

MJJA

SO

N

LT 0 LT 3 LT 5 LT 0 LT 3 LT 5

Casanare

Cordoba

Santander

Tolima

Valle del Cauca

Casanare

Cordoba

Santander

Tolima

Valle del Cauca

Casanare

Cordoba

Santander

Tolima

Valle del Cauca

Casanare

Cordoba

Santander

Tolima

Valle del Cauca

Lead Time

0.0

0.1

0.2

0.3

0.4

Fig. 6. Cross-validation goodness index (Kendall correlation,−

τ ) CCA models disaggregated by season, department and lead time. Shown are−

τ values for observedSSTs (ERSST) and predicted SST and regional precipitation (NCEP-CFSv2). Seasons included are DJF: December–January–February; MAM: March–April–May; JJA:June–July–August; and SON: September–October–November. Lead times are specified in Table 2.

CFSv2 Precipitation CFSv2 SST

Cross validation Retrospective validation Cross validation Retrospective validation

0.0

0.1

0.2

0.3

0.4

0.5

Kend

all c

orre

latio

n (

)

Lead TimeLT 0

LT 3

LT 5

Fig. 7. Boxplot of the goodness index (Kendall correlation,−

τ ) for different lead times for CCA models using CFSv2 predictors (SST and precipitation), for both cross-validation and retrospective validation procedures. Thick horizontal line indicates the median, boxes extend to the interquartile range, and whiskers extend to 5%and 95% of the distribution. Lead times are specified in Table 2.

A. Esquivel et al. Climate Services xxx (xxxx) xxx–xxx

8

Page 9: Predictability of seasonal precipitation across major crop ... · Increasing availability in wide-area climate datasets (Funk et al., 2015; Muñoz et al., 2010), open data policies

lower−

τ values than the shorter (0-month; and 3-month) lead times.There are cases in which, as with ERSST (see Section 3.2),

τ in-creases with increased lead time, though increases are generally small.This is the case for Cordoba and Casanare (Fig. 6). The reasons for thisare difficult to discern, but it may be that the CFSv2 model is per-forming better for those lead times for SSTs, leading to greater

τ valuesin the CCA models. The same is not seen for CFSv2 predicted pre-cipitation forecasts, which would suggest that the greater skill in pre-dicting SSTs does not result in greater predicted precipitation skill.

4. Discussion

4.1. Predictability of local precipitation across Colombia

The relationship between the regional climate and SSTs via atmo-spheric teleconnection is well established in the tropics, especially in-terannual variability of rainfall in response to SST anomalies in thetropical Pacific (Cane et al., 1986; Coelho et al., 2002; Ropelewski andHalpert, 1987), or tropical Atlantic (Fernandes et al., 2011; Ndiayeet al., 2011; Yoon and Zeng, 2010), or Indian Ocean Dipole (Jourdainet al., 2013; Kripalani and Kumar, 2004). At the same time, the re-levance of different SST predictor regions may vary seasonally or basedon other concurrent phenomena (Recalde-Coronel et al., 2014; Uvoet al., 1998). In Colombia, the relevance of SST variability is generallywell understood, but the focus of research to date has largely beencentered around longer-term trends and variation under ENSO regimes(Álvarez-Villa et al., 2011; Poveda et al., 2011; Poveda and Mesa,2000). Rainfall in Colombia exhibits a coherent behavior with respectto ENSO, with generally negative precipitation anomalies during the ElNiño (warm) phases, and positive precipitation anomalies during LaNiña (cold) phases (Poveda et al., 2001). Our analyses illustrate howSST teleconnections can be used as a basis for statistical climate forecastmodels and, specifically, how SST data (observed or simulated) can beused to develop statistical climate forecasts for Colombia.

We find that statistical forecast models generally, though not al-ways, serve to predict seasonal precipitation across the studied de-partments (see Figs. 3 and 4). In general, our results are consistent withresults of previous studies in Latin America (Montecinos et al., 2000;Recalde-Coronel et al., 2014) and elsewhere (Korecha and Barnston,2007; Ndiaye et al., 2011, 2009; Paeth and Hense, 2003) in terms offorecast skill, with values of

τ typically in the range 0.1–0.3, mostforecast situations (ca. 75%) having

τ at or above 0.1, 38.8% at orabove 0.2, and 18.8% at or above 0.3.

In our analysis, the greatest predictability is found during dry sea-sons across inter-Andean valleys, when the spatial loadings of themodes of variability used to construct the CCA models are strongest inthe Pacific basin (see Supplementary Fig. S1–S8). Even during ENSO-neutral years, positive or negative anomalies in the El Niño region canaffect local precipitation. These results are consistent with those ofCórdoba-Machado et al., 2015, who studied the influence of tropicalPacific SSTs [El Niño and El Niño Modoki, Weng et al. (2007)] onseasonal precipitation. They report that precipitation can be predictedusing tropical Pacific SSTs with suitable skill with lead times of up tofour seasons, though with greatest prediction skill during dry periods.The same study indicated that the most important mode of variabilityfor Colombian precipitation is El Niño (vs. El Niño Modoki). Similarfindings are reported by Poveda and Mesa (1997), Poveda et al. (2011),and Tootle et al. (2008) who analyzed and reviewed hydro-climaticvariability in Colombia with respect to ENSO. Their results, similar toours, suggest that, when predictability is moderate, the SST modes as-sociated with El Niño has the greatest influence on seasonal precipita-tion.

Low predictability was found towards the Eastern Plains region(Casanare department). As stated above (Section 3.1) data qualitycould, at least partly, explain our results for Casanare. Previous studieshave also suggested that precipitation in the Eastern Plains (where

Casanare is situated) is influenced by high-level winds (BustamanteLozano et al., 2013; Pacheco and León-Aristizabal, 2001). NeitherAmazon- or wind-related drivers were explicitly considered in thisstudy, but a comparative analysis of their skill compared to SSTs is atopic for future research. Furthermore, all users of climate services inthe region would benefit from improvements in the quality and densityof the weather station network in Casanare.

Predictability was limited during wet periods of the year.Particularly for inter-Andean valleys (i.e. Valle del Cauca and Tolima),we find that predictability is low during MAM and SON (rainy seasons,see Fig. 7). Foremost, we note that the effects of ENSO in the ColombianAndes are phase-locked, with strong (weak) signals during De-cember–February (March–May) (Poveda et al., 2001), which wouldconfirm our results. Additionally, according to Guzman et al. (2014)and Poveda et al. (2001), there are many atmospheric systems thataffect precipitation in Colombia, including variability in the position ofthe Inter-Tropical Convergence Zone (ITCZ), transient high pressuresystems, Caribbean easterly waves, tropical cyclones, the South AtlanticConvergence Zone (SACZ), cold fronts, and low-level jets [e.g. theCHOCO low-level jet, Poveda and Mesa (2000)]. The interactions ofthese large-scale phenomena with meso-scale climate and land surfacedynamics (e.g. coasts in both Atlantic and Pacific oceans, complexAndean orography patterns) can substantially limit seasonal precipita-tion predictability (Poveda, 2004; Poveda et al., 2001).

4.2. Forecast skill across seasons and geographies

As stated above, in a ‘real’ prediction situation the issued forecast islagged with respect to the month of the predictor. Results for the pre-dictive analysis were consistent with respect to those for the predict-ability limit (i.e. at LT-Sim for ERSST), though CCA model skill de-creased with increasing lead time, as it is naturally expected (Torrenceand Webster, 1998). This consistency across lead times and predictors isan encouraging result, and implies that both SST persistence is sub-stantial and that the CFSv2 model offers predictions (both SSTs andprecipitation) that are useful for seasonal forecasting in Colombia.

Of particular relevance here are the forecasts driven by CFSv2, sincethese are the ones that can be issued at the longest lead times (here5months). We find that CFSv2-driven forecast skill is similar for SSTsand regional precipitation, suggesting that both types of predictors areof possible use for Colombia. Importantly, CFSv2 SST-driven forecastsgenerally performed better than precipitation-driven ones, suggestingthat CFSv2 skill is better for SSTs than for precipitation. The reasons forlower skill in simulated CFSv2 precipitation are difficult to discern, butcould be associated to the parameterized physics (deep moist convec-tion, boundary layer processes), and/or to the horizontal resolution ofthe model, which does not fully resolve land surface processes and theeffect of topography (Gallée et al., 2004; Garcia-Carreras et al., 2015).There may, however, be opportunities to improve precipitation-drivenforecast skill if different models are used. For instance, MOS applied toprecipitation as predictor has been reported to perform better than SSTsfor the ECHAM4.5 model in Ecuador (Recalde-Coronel et al., 2014),and the same is true for MOS applied to low-level wind in the Sahel(Ndiaye et al., 2011). By assessing other GCMs using MOS, future stu-dies could assess whether precipitation-driven forecasts using MOSoffer good opportunities for seasonal prediction in Colombia and morebroadly across Latin America.

4.3. Towards a climate service for improved agricultural decision making

We have assessed seasonal forecast skill at lead times commensuratewith seasonal agricultural decision making. These may be useful tomake decisions on whether to plant, what varieties to plant, and whento plant. We find that predictive skill of CCA models both using ob-served and modeled SSTs, as well as using predicted precipitation isgenerally, yet not always, suitable for issuing forecasts. Some 134

A. Esquivel et al. Climate Services xxx (xxxx) xxx–xxx

9

Page 10: Predictability of seasonal precipitation across major crop ... · Increasing availability in wide-area climate datasets (Funk et al., 2015; Muñoz et al., 2010), open data policies

(74.4%) out of the 180 departments (5) * seasons (4) * predictors (3) *lead times (3) analyzed here showed

τ values equal or above 0.1, 38.8%equal or above 0.2, and 18.8% equal or above 0.3. An illustrative ex-ample using one weather station per department for the DJF season alsoshows moderate to high forecast skill (Supplementary Fig. S9).Importantly, results were consistent between ERSST and CFSv2-drivenstatistical forecasts. Similar results have been reported in other tropical,and extra-tropical regions. For instance, Ramírez-Rodrigues et al.(2016) tested an ‘always-correct-season’ forecast for wheat in Mexico.They find that the value of seasonal forecasts in informing the decisionof whether to plant can save farmers up to 123 USD ha−1 season−1 inrainfed systems. Similarly, Roudier et al. (2016) report that seasonalforecasts for millet in Niger can generate income gains of 1.8–13%.Hammer et al. (1996) estimate profit gains of up to 20% and/or riskreduction of up to 35% with forecast-informed tactical adjustments inwheat crop management in Australia. Capa-Morocho et al. (2016) re-ported that forecasts for wheat in Spain have significant potentialespecially if they are issued together with market price information.

According to our results, therefore, agriculturally-relevant seasonalforecasts are feasible in most sites in Colombia, though some sites showlimited forecast skill suggesting that further work is required to im-prove seasonal forecast models. Seasonal forecast models have anumber of uses for on-farm agricultural decision-making. However, thevalue of the forecast may vary per location, farming system, livelihoodstrategy, and lead time (Davey and Brookshaw, 2011; Hammer et al.,1996). For instance, the value of forecasts for farmers in Burkina Faso isgreater for sorghum-millet farmers in semi-arid areas compared to li-vestock keepers in arid lands (Ingram et al., 2002). For these farmers, itis important that forecasts can be used to determine the start and end ofthe rainy period, as well as the occurrence of drought spells during theseason are the most useful (Ingram et al., 2002; Roncoli et al., 2002).Similarly, 3–6month seasonal outlooks (such as the ones assessed here)can help identify situations in which farmers should not plant due toincreasingly risky climatic conditions (Phillips et al., 2002; Ziervogeland Calder, 2003). Other uses of seasonal forecasts include the defini-tion of optimal planting dates, the choice of variety or crop (e.g. short-vs. long-cycle varieties or crops), and decisions related to stock man-agement, pricing and trade (Hammer et al., 1996; Soler et al., 2007).However, in order to realize the full potential of seasonal predictions,further improvement in the skill and usability of forecasts is required.

One clear avenue toward improving skill is to improve predictordomain selection. Although the analyses presented here used a single(the narrow global tropics) predictor domain, previous studies in LatinAmerica have reported sensitivity to domain choice. Recalde-Coronelet al. (2014) report high sensitivity in the skill of seasonal forecasts tothe selection of the SST domain to construct the prediction. Results ofCórdoba-Machado et al. (2015) with respect to the relevance of El Niñovs. El Niño Modoki in Colombia also imply sensitivity in the selection ofpredictor domain for statistical forecast models. Future efforts thatcapitalize on the evidence presented here to maximize predictive skillby better domain selection are warranted. Other potential avenues toimprove forecast skill include the use of other predictors [e.g. geopo-tential height, zonal winds; Córdoba-Machado et al. (2015); Guzmanet al. (2014)], or use of alternative prediction methods such as partialleast squares [e.g. Hernández-Barrera and Rodríguez-Puebla (2017)],regional dynamical models that account for mesoscale processes [e.g.Baigorria et al. (2008)], or the use of weather types (Moron et al., 2015;Muñoz et al., 2016, 2015). Similarly, future studies could also analyzethe predictability of wet- and dry-day frequencies and compare thesewith forecasts of total precipitation.

For a climate service for agriculture to be successful, it is importantthat the seasonal forecasts underpinning the decision-making processare skillful, and communicated in a transparent and appropriatemanner to relevant stakeholder groups. While our study contributesimportant evidence to addressing the question of predictability andprediction skill at agriculturally-relevant timescales, for key

departments of importance for major staple crops (rice, maize, andbeans) for the country, it is only the first step towards the establishmentof a climate service. Future work will be required to link seasonalforecasts to crop modelling applications [e.g. Capa-Morocho et al.(2016); Roudier et al. (2016)], to design appropriate delivery means,and to establish stakeholder-driven processes that allow two-waycommunication between forecast issuing institutions (e.g. IDEAM–Colombian Meteorological Service) and the famers’ organizations andfarming communities that are using such forecasts (Ndiaye et al., 2012;Vaughan et al., 2016).

5. Conclusions

In this study, we have assessed precipitation predictability andforecast skill at agriculturally-relevant timescales for key growing areasof three major staple crops, namely, rice, maize, and beans. The fivedepartments analyzed represent the three major natural regions ofColombia where agriculture is most prevalent: Valle del Cauca (Andeanregion), Santander (Andean), Tolima (Andean), Casanare (EasternPlains), and Cordoba (Caribbean). Analyses concentrate in under-standing potential predictability as well as 3- to 5-month forecast skillusing observed SSTs and modelled (NCEP-CFSv2) SSTs and precipita-tion. Results indicate that forecast skill at the three different lead timesanalyzed is both consistent and suitable for operational forecasting.Specifically, only 25.5% of the forecast situations (5 departments * 4seasons * 3 predictors * 3 lead times) analyzed here showed unsuitable( <

τ 0.1) forecast skill. Consistency between observed and modeledpredictors indicates that there CFSv2 model is suitable for seasonalprediction in Colombia, though there may be opportunities for usingmodels that have been proven suitable in other countries in LatinAmerica (e.g. ECHAM4.5, Recalde-Coronel et al., 2014). Likewise,given the complexity of Colombian precipitation generation, it is im-portant to explore both additional predictors, potential refinements tothe selection of predictor domains, as well as other forecasting methods,in order to improve on the forecast models presented here. Finally,further work is warranted to determine the relevance and usefulness ofseasonal climate forecasts with respect to the establishment of a climateservice for agriculture.

Ultimately, the ability to consistently generate reliable seasonalforecasts is the critical first step in the development of agro-climaticforecasts and corresponding agro-climatic services. Agro-climatic fore-casts serve to contextualize seasonal climate forecasts in relation tospecific crops through the approximation of climate effects on cropgrowth and yields (Ramírez-Rodrigues et al., 2016; Roudier et al.,2016). When implemented in a rigorous and replicable manner, with afocus specifically toward the users of the resulting information, this inturn becomes a climate service supporting an important economicsector (Hewitt et al., 2012; Vaughan et al., 2016; Vaughan and Dessai,2014).

Conflict of interest

Authors declare no conflict of interest.

Acknowledgments

This work was carried out under the Climate Services for ResilientDevelopment (CSRD, http://www.cs4rd.org/) program, and under theCCAFS project Agroclimas (http://bit.ly/2i3V0Nh). We thank theUnited States Agency for International Development (USAID) and theCGIAR Research Program on Climate Change, Agriculture and FoodSecurity (CCAFS) for financial support. CCAFS is carried out withsupport from CGIAR Fund Donors and through bilateral fundingagreements. For details please visit https://ccafs.cgiar.org/donors. Theviews expressed in this paper cannot be taken to reflect the officialopinions of these organizations. We also gratefully acknowledge the

A. Esquivel et al. Climate Services xxx (xxxx) xxx–xxx

10

Page 11: Predictability of seasonal precipitation across major crop ... · Increasing availability in wide-area climate datasets (Funk et al., 2015; Muñoz et al., 2010), open data policies

Insituto de Hidrologia, Meteorologia y Estudios Ambientales (IDEAM)for providing access to their weather station data, and for discussion onthe CCA-based forecast model results presented here. Authors thankSimon J. Mason from the International Research Institute for Climateand Society (IRI) for advice on the implementation and evaluation ofCCA models, Diana Giraldo from CIAT for literature on seasonal fore-casting for agriculture, and Angel G. Muñoz (IRI) and Ousmane Ndiaye(ANACIM, Senegal) for comments and feedback on a previous versionof this manuscript. Authors thank the Ministry of Agriculture and RuralDevelopment (MADR) of Colombia, the Colombian National RiceFederation (FEDEARROZ) and the National Cereals and LegumesFederation (FENALCE), for their financial support, and for their con-tribution with data and insights for this study. Finally, we thank twoanonymous reviewers for their insightful comments.

Appendix A. Supplementary data

Supplementary data to this article can be found online at https://doi.org/10.1016/j.cliser.2018.09.001.

References

Acock, M.C., Pachepsky, Y.A., 2000. Estimating missing weather data for agriculturalsimulations using group method of data handling. J. Appl. Meteorol. 39, 1176–1184.https://doi.org/10.1175/1520-0450(2000) 039<1176:EMWDFA>2.0.CO;2.

Álvarez-Villa, O.D., Vélez, J.I., Poveda, G., 2011. Improved long-term mean annualrainfall fields for Colombia. Int. J. Climatol. 31, 2194–2212. https://doi.org/10.1002/joc.2232.

Baigorria, G.A., Jones, J.W., O’Brien, J.J., 2008. Potential predictability of crop yieldusing an ensemble climate forecast by a regional circulation model. Agric. For.Meteorol. 148, 1353–1361. https://doi.org/10.1016/j.agrformet.2008.04.002.

Bustamante Lozano, Á.M., Páez Martínez, A., Espitia Barrera, J.E., Cárdenas Castro, E.,2013. Análisis de datos meteorológicos para identificar y definir el clima en Yopal.Casanare. Rev. Med. Vet. 85. https://doi.org/10.19052/mv.2301.

Cane, M.A., Zebiak, S.E., Dolan, S.C., 1986. Experimental forecasts of El Nino. Nature321, 827.

Capa-Morocho, M., Ines, A.V.M., Baethgen, W.E., Rodríguez-Fonseca, B., Han, E., Ruiz-Ramos, M., 2016. Crop yield outlooks in the Iberian Peninsula: connecting seasonalclimate forecasts with crop simulation models. Agric. Syst. 149, 75–87. https://doi.org/10.1016/j.agsy.2016.08.008.

CIAT-MADR, 2015. Logros y retos de la agricultura colombiana frente al cambioclimático.

Coelho, C.A.S., Uvo, C.B., Ambrizzi, T., 2002. Exploring the impacts of the tropical PacificSST on the precipitation patterns over South America during ENSO periods. Theor.Appl. Climatol. 71, 185–197. https://doi.org/10.1007/s007040200004.

Córdoba-Machado, S., Palomino-Lemus, R., Gámiz-Fortis, S.R., Castro-Díez, Y., Esteban-Parra, M.J., 2015. Influence of tropical Pacific SST on seasonal precipitation inColombia: prediction using El Niño and El Niño Modoki. Clim. Dyn. 44, 1293–1310.https://doi.org/10.1007/s00382-014-2232-3.

Cortes Bello, C.A., Bernal Patino, J.G., Diaz Almanza, E.D., Mendez Monroy, J.F., 2013.Uso del modelo AquaCrop para estimar rendimientos para el cultivo de maíz en losdepartamentos de Cordoba, Meta. Tolima y Valle del Cauca, Rome, Italy.

DANE, 2014. Censo Nacional Agropecuario 2014. Bogotá D.C, Colombia.Davey, M., Brookshaw, A., 2011. Long-range meteorological forecasting and links to

agricultural applications. Food Policy 36, S88–S93. https://doi.org/10.1016/j.foodpol.2010.10.009.

Delerce, S., Dorado, H., Grillon, A., Rebolledo, M.C., Prager, S.D., Patiño, V.H., GarcésVarón, G., Jiménez, D., 2016. Assessing weather-yield relationships in rice at localscale using data mining approaches. PLoS One 11, e0161620. https://doi.org/10.1371/journal.pone.0161620.

Diaz, A.F., Studzinski, C.D., Mechoso, C.R., 1998. Relationships between precipitationanomalies in Uruguay and Southern Brazil and sea surface temperature in the Pacificand Atlantic Oceans. J. Clim. 11, 251–271. https://doi.org/10.1175/1520-0442(1998) 011<0251:RBPAIU>2.0.CO;2.

Dinku, T., Hailemariam, K., Maidment, R., Tarnavsky, E., Connor, S., 2014. Combined useof satellite estimates and rain gauge observations to generate high-quality historicalrainfall time series over Ethiopia. Int. J. Climatol. 34, 2489–2504. https://doi.org/10.1002/joc.3855.

Dorward, P., Clarkson, G., Stern, R., 2015. Participatory integrated climate services foragriculture (PICSA): Field manual. CCAFS, Reading, UK.

Eischeid, J.K., Pasteris, P.A., Diaz, H.F., Plantico, M.S., Lott, N.J., 2000. Creating a seriallycomplete, national daily time series of temperature and precipitation for the WesternUnited States. J. Appl. Meteorol. 39, 1580–1591. https://doi.org/10.1175/1520-0450(2000) 039<1580:CASCND>2.0.CO;2.

Fedearroz, 2017. IV Censo Nacional Arrocero 2016. Bogotá D.C., Colombia.Fernandes, K., Baethgen, W., Bernardes, S., DeFries, R., DeWitt, D.G., Goddard, L.,

Lavado, W., Lee, D.E., Padoch, C., Pinedo-Vasquez, M., Uriarte, M., 2011. NorthTropical Atlantic influence on western Amazon fire season variability. Geophys. Res.Lett. 38https://doi.org/10.1029/2011GL047392. n/a–n/a.

Fraisse, C.W., Breuer, N.E., Zierden, D., Bellow, J.G., Paz, J., Cabrera, V.E., Garcia yGarcia, A., Ingram, K.T., Hatch, U., Hoogenboom, G., Jones, J.W., O’Brien, J.J., 2006.AgClimate: a climate forecast information system for agricultural risk management inthe southeastern USA. Comput. Electron. Agric. 53, 13–27. https://doi.org/10.1016/j.compag.2006.03.002.

Funk, C., Peterson, P., Landsfeld, M., Pedreros, D., Verdin, J., Shukla, S., Husak, G.,Rowland, J., Harrison, L., Hoell, A., Michaelsen, J., 2015. The climate hazards in-frared precipitation with stations—a new environmental record for monitoring ex-tremes. Sci. Data 2, 150066. https://doi.org/10.1038/sdata.2015.66.

Gallée, H., Moufouma-Okia, W., Bechtold, P., Brasseur, O., Dupays, I., Marbaix, P.,Messager, C., Ramel, R., Lebel, T., 2004. A high-resolution simulation of a WestAfrican rainy season using a regional climate model. J. Geophys. Res. 109, D05108.https://doi.org/10.1029/2003jd004020.

Garcia-Carreras, L., Challinor, A.J., Parkes, B.J., Birch, C.E., Nicklin, K.J., Parker, D.J.,2015. The impact of parameterized convection on the simulation of crop processes. J.Appl. Meteorol. Climatol. 54, 1283–1296. https://doi.org/10.1175/JAMC-D-14-0226.1.

Georgeson, L., Maslin, M., Poessinouw, M., 2017. Global disparity in the supply ofcommercial weather and climate information services. Sci. Adv. 3, e1602632.https://doi.org/10.1126/sciadv.1602632.

Glahn, H.R., 1968. Canonical Correlation and Its Relationship to Discriminant Analysisand Multiple Regression. J. Atmos. Sci. 25, 23–31. https://doi.org/10.1175/1520-0469(1968) 025<0023:CCAIRT>2.0.CO;2.

Goddard, L., Mason, S.J., Zebiak, S.E., Ropelewski, C.F., Basher, R., Cane, M.A., 2001.Current approaches to seasonal to interannual climate predictions. Int. J. Climatol.21, 1111–1152. https://doi.org/10.1002/joc.636.

Guzman, D., Ruiz, J.F., Cadena, M., 2014. Regionalizacion de Colombia segun la esta-cionalidad de la precipitacion media mensual, a traves de analisis de componentesprincipales (ACP). Bogotá D.C, Colombia.

Hammer, G., Holzworth, D., Stone, R., 1996. The value of skill in seasonal climateforecasting to wheat crop management in a region with high climatic variability.Aust. J. Agric. Res. 47, 717. https://doi.org/10.1071/AR9960717.

Hansen, J.W., Challinor, A., Ines, A., Wheeler, T., Moron, V., 2006. Translating climateforecasts into agricultural terms: advances and challenges. Clim. Res. 33, 27–41.https://doi.org/10.3354/cr033027.

Hansen, J.W., Indeje, M., 2004. Linking dynamic seasonal climate forecasts with cropsimulation for maize yield prediction in semi-arid Kenya. Agric. For. Meteorol. 125,143–157. https://doi.org/10.1016/j.agrformet.2004.02.006.

Hansen, J.W., Mason, S.J., Sun, L., Tall, A., 2011. Review of seasonal climate forecastingfor agriculture in Sub-Saharan Africa. Exp. Agric. 47, 205–240. https://doi.org/10.1017/S0014479710000876.

Harktamp, A., White, J., Rodriguez Aguilar, A., Baenziger, M., Srinivasan, G., Granados,G., Crossa, J., 2000. Maize production environments revisited. A GIS-based approach.Mexico D.F.

Hernández-Barrera, S., Rodríguez-Puebla, C., 2017. Wheat yield in Spain and associatedsolar radiation patterns. Int. J. Climatol. https://doi.org/10.1002/joc.4975.

Hewitt, C., Mason, S., Walland, D., 2012. The global framework for climate services. Nat.Clim. Chang. 2, 831–832. https://doi.org/10.1038/nclimate1745.

Hotelling, H., 1936. Relations between two sets of variates. Biometrika 28, 321. https://doi.org/10.2307/2333955.

Huang, B., Banzon, V.F., Freeman, E., Lawrimore, J., Liu, W., Peterson, T.C., Smith, T.M.,Thorne, P.W., Woodruff, S.D., Zhang, H.-M., 2015. Extended reconstructed sea sur-face temperature version 4 (ERSST.v4). Part I: upgrades and intercomparisons. J.Clim. 28, 911–930. https://doi.org/10.1175/JCLI-D-14-00006.1.

Huffman, G.J., Adler, R.F., Rudolf, B., Schneider, U., Keehn, P.R., 1995. Global pre-cipitation estimates based on a technique for combining satellite-based estimates,rain gauge analysis, and NWP model precipitation information. J. Clim. 8,1284–1295. https://doi.org/10.1175/1520-0442(1995) 008<1284:GPEBOA>2.0.CO;2.

Iizumi, T., Luo, J.-J., Challinor, A.J., Sakurai, G., Yokozawa, M., Sakuma, H., Brown, M.E.,Yamagata, T., 2014. Impacts of El Niño Southern oscillation on the global yields ofmajor crops. Nat. Commun. 5. https://doi.org/10.1038/ncomms4712.

Ingram, K.T., Roncoli, M.C., Kirshen, P.H., 2002. Opportunities and constraints forfarmers of west Africa to use seasonal precipitation forecasts with Burkina Faso as acase study. Agric. Syst. 74, 331–349. https://doi.org/10.1016/s0308-521x(02)00044-6.

Jones, J.W., Hansen, J.W., Royce, F.S., Messina, C.D., 2000. Potential benefits of climateforecasting to agriculture. Agric. Ecosyst. Environ. 82, 169–184. https://doi.org/10.1016/S0167-8809(00)00225-5.

Jourdain, N.C., Gupta, A. Sen, Taschetto, A.S., Ummenhofer, C.C., Moise, A.F., Ashok, K.,2013. The Indo-Australian monsoon and its relationship to ENSO and IOD in re-analysis data and the CMIP3/CMIP5 simulations. Clim. Dyn. 41, 3073–3102. https://doi.org/10.1007/s00382-013-1676-1.

Korecha, D., Barnston, A.G., 2007. Predictability of June–September rainfall in Ethiopia.Mon. Weather Rev. 135, 628–650. https://doi.org/10.1175/MWR3304.1.

Kripalani, R.H., Kumar, P., 2004. Northeast monsoon rainfall variability over south pe-ninsular India vis-à-vis the Indian Ocean dipole mode. Int. J. Climatol. 24,1267–1282. https://doi.org/10.1002/joc.1071.

Landman, W.A., Goddard, L., 2002. Statistical recalibration of GCM forecasts overSouthern Africa using model output statistics. J. Clim. 15, 2038–2055. https://doi.org/10.1175/1520-0442(2002) 015<2038:SROGFO>2.0.CO;2.

Liu, W., Huang, B., Thorne, P.W., Banzon, V.F., Zhang, H.-M., Freeman, E., Lawrimore, J.,Peterson, T.C., Smith, T.M., Woodruff, S.D., 2015. Extended reconstructed sea surfacetemperature version 4 (ERSST.v4): Part II. Parametric and structural uncertaintyestimations. J. Clim. 28, 931–951. https://doi.org/10.1175/JCLI-D-14-00007.1.

Llanos-Herrera, L., 2014. RClimTool.

A. Esquivel et al. Climate Services xxx (xxxx) xxx–xxx

11

Page 12: Predictability of seasonal precipitation across major crop ... · Increasing availability in wide-area climate datasets (Funk et al., 2015; Muñoz et al., 2010), open data policies

Mason, S., 2011. Seasonal forecasting using the climate predictability tool (CPT). In: 36thNOAA Annual Climate Diagnostics and Prediction Workshop. Science andTechnology Infusion Climate Bulletin, NOAA’s National Weather Service, Fort Worth,TX, USA, p. 3.

Mason, S.J., Tippett, M.K., 2017. Climate Predictability Tool version 15.5.10.Montecinos, A., Díaz, A., Aceituno, P., 2000. Seasonal diagnostic and predictability of

rainfall in subtropical South America based on tropical Pacific SST. J. Clim. 13,746–758. https://doi.org/10.1175/1520-0442(2000) 013<0746:SDAPOR>2.0.CO;2.

Moron, V., Robertson, A.W., Qian, J.-H., Ghil, M., 2015. Weather types across theMaritime Continent: from the diurnal cycle to interannual variations. Front. Environ.Sci. 2. https://doi.org/10.3389/fenvs.2014.00065.

Muñoz, Á.G., Goddard, L., Mason, S.J., Robertson, A.W., 2016. Cross-time scale interac-tions and rainfall extreme events in Southeastern South America for the australsummer. Part II: Predictive Skill. J. Clim. 29, 5915–5934. https://doi.org/10.1175/JCLI-D-15-0699.1.

Muñoz, Á.G., Goddard, L., Robertson, A.W., Kushnir, Y., Baethgen, W., 2015. Cross-timescale interactions and rainfall extreme events in Southeastern South America for theaustral summer. Part I: Potential Predictors. J. Clim. 28, 7894–7913. https://doi.org/10.1175/JCLI-D-14-00693.1.

Muñoz, Á.G., López, P., Velásquez, R., Monterrey, L., León, G., Ruiz, F., Recalde, C.,Cadena, J., Mejía, R., Paredes, M., Bazo, J., Reyes, C., Carrasco, G., Castellón, Y.,Villarroel, C., Quintana, J., Urdaneta, A., 2010. An environmental watch system forthe Andean Countries: El observatorio andino. Bull. Am. Meteorol. Soc. 91,1645–1652. https://doi.org/10.1175/2010BAMS2958.1.

Ndiaye, O., Goddard, L., Ward, M.N., 2009. Using regional wind fields to improve generalcirculation model forecasts of July-September Sahel rainfall. Int. J. Climatol. 29,1262–1275. https://doi.org/10.1002/joc.1767.

Ndiaye, O., Ward, M.N., Thiaw, W.M., 2011. Predictability of seasonal sahel rainfall usingGCMs and lead-time improvements through the use of a coupled model. J. Clim. 24,1931–1949. https://doi.org/10.1175/2010JCLI3557.1.

Ndiaye, O., Zougmoré, R., Hansen, J., Diongue, A., Seck, E.M., 2012. Using probabilisticseasonal forecasting to improve farmers’ decision in Kaffrine, Senegal. In: Banaitiene,N. (Ed.), Risk Management-Current Issues and Challenges, pp. 497–504. Doi: 10.5772/2568.

Pacheco, Y., León-Aristizabal, G., 2001. Clasificación climática de la OrinoquíaColombiana a partir de los patrones de circulación atmosférica. Meteorol. Colomb. 4,117–120.

Paeth, H., Hense, A., 2003. Seasonal forecast of sub-sahelian rainfall using cross validatedmodel output statistics. Meteorol. Zeitschrift 12, 157–173. https://doi.org/10.1127/0941-2948/2003/0012-0157.

Patt, A., Gwata, C., 2002. Effective seasonal climate forecast applications: examiningconstraints for subsistence farmers in Zimbabwe. Glob. Environ. Change 12,185–195. https://doi.org/10.1016/S0959-3780(02)00013-4.

Phillips, J., Deane, D., Unganai, L., Chimeli, A., 2002. Implications of farm-level responseto seasonal climate forecasts for aggregate grain production in Zimbabwe. Agric. Syst.74, 351–369. https://doi.org/10.1016/S0308-521X(02)00045-8.

Pickering, N.B., Hansen, J.W., Jones, J.W., Wells, C.M., Chan, V.K., Godwin, D.C., 1994.WeatherMan: a utility for managing and generating daily weather data. Agron. J. 86,332. https://doi.org/10.2134/agronj1994.00021962008600020023x.

Poveda, G., 2004. La hidroclimatología de Colombia: una síntesis desde la escala inter-decadal hasta la escala diurna. Rev. Acad. Colomb. Cienc. 28, 201–222.

Poveda, G., Álvarez, D.M., Rueda, Ó.A., 2011. Hydro-climatic variability over the Andesof Colombia associated with ENSO: a review of climatic processes and their impact onone of the Earth’s most important biodiversity hotspots. Clim. Dyn. 36, 2233–2249.https://doi.org/10.1007/s00382-010-0931-y.

Poveda, G., Jaramillo, A., Gil, M.M., Quiceno, N., Mantilla, R.I., 2001. Seasonally inENSO-related precipitation, river discharges, soil moisture, and vegetation index inColombia. Water Resour. Res. 37, 2169–2178. https://doi.org/10.1029/2000WR900395.

Poveda, G., Mesa, O.J., 2000. On the existence of Lloró (the rainiest locality on Earth):Enhanced ocean-land-atmosphere interaction by a low-level jet. Geophys. Res. Lett.27, 1675–1678. https://doi.org/10.1029/1999GL006091.

Poveda, G., Mesa, O.J., 1997. Feedbacks between hydrological processes in TropicalSouth America and large-scale ocean-atmospheric phenomena. J. Clim. 10,2690–2702. https://doi.org/10.1175/1520-0442(1997) 010<2690:fbhpit>2.0.co;2.

Ramírez-Rodrigues, M.A., Alderman, P.D., Stefanova, L., Cossani, C.M., Flores, D., Asseng,S., 2016. The value of seasonal forecasts for irrigated, supplementary irrigated, andrainfed wheat cropping systems in northwest Mexico. Agric. Syst. 147, 76–86.https://doi.org/10.1016/j.agsy.2016.05.005.

Ramirez-Villegas, J., Challinor, A., 2012. Assessing relevant climate data for agriculturalapplications. Agric. For. Meteorol. 161, 26–45. https://doi.org/10.1016/j.agrformet.2012.03.015.

Ray, D.K., Gerber, J.S., MacDonald, G.K., West, P.C., 2015. Climate variation explains athird of global crop yield variability. Nat. Commun. 6, 5989. https://doi.org/10.1038/ncomms6989.

Recalde-Coronel, G.C., Barnston, A.G., Muñoz, Á.G., 2014. Predictability of december–-april rainfall in coastal and Andean Ecuador. J. Appl. Meteorol. Climatol. 53,1471–1493. https://doi.org/10.1175/JAMC-D-13-0133.1.

Roncoli, C., Ingram, K., Kirshen, P., 2002. Reading the rains: local knowledge and rainfallforecasting in Burkina Faso. Soc. Nat. Resour. 15, 409–427. https://doi.org/10.1080/08941920252866774.

Ropelewski, C.F., Halpert, M.S., 1987. Global and regional scale precipitation patternsassociated with the El Niño/Southern Oscillation. Mon. Weather Rev. 115,1606–1626. https://doi.org/10.1175/1520-0493(1987) 115<1606:garspp>2.0.co;2.

Roudier, P., Alhassane, A., Baron, C., Louvet, S., Sultan, B., 2016. Assessing the benefits ofweather and seasonal forecasts to millet growers in Niger. Agric. For. Meteorol. 223,168–180. https://doi.org/10.1016/j.agrformet.2016.04.010.

Saha, S., Moorthi, S., Wu, X., Wang, J., Nadiga, S., Tripp, P., Behringer, D., Hou, Y.-T.,Chuang, H., Iredell, M., Ek, M., Meng, J., Yang, R., Mendez, M.P., van den Dool, H.,Zhang, Q., Wang, W., Chen, M., Becker, E., 2014. The NCEP climate forecast systemversion 2. J. Clim. 27, 2185–2208. https://doi.org/10.1175/JCLI-D-12-00823.1.

Soler, C.M.T., Sentelhas, P.C., Hoogenboom, G., 2007. Application of the CSM-CERES-Maize model for planting date evaluation and yield forecasting for maize grown off-season in a subtropical environment. Eur. J. Agron. 27, 165–177. https://doi.org/10.1016/j.eja.2007.03.002.

Stainforth, D.A., Aina, T., Christensen, C., Collins, M., Faull, N., Frame, D.J.,Kettleborough, J.A., Knight, S., Martin, A., Murphy, J.M., Piani, C., Sexton, D., Smith,L.A., Spicer, R.A., Thorpe, A.J., Allen, M.R., 2005. Uncertainty in predictions of theclimate response to rising levels of greenhouse gases. Nature 433, 403–406. https://doi.org/10.1038/nature03301.

Steduto, P., Hsiao, T.C., Raes, D., Fereres, E., 2009. Aquacrop—the Fao crop model tosimulate yield response to water: I. Concepts and underlying principles. Agron. J.101, 426–437. https://doi.org/10.2134/agronj2008.0139s.

Tall, A., Coulibaly, J.Y., Diop, M., 2018. Do climate services make a difference? A reviewof evaluation methodologies and practices to assess the value of climate informationservices for farmers: Implications for Africa. Clim. Serv. https://doi.org/10.1016/j.cliser.2018.06.001.

Tootle, G.A., Piechota, T.C., Gutiérrez, F., 2008. The relationships between Pacific andAtlantic Ocean sea surface temperatures and Colombian streamflow variability. J.Hydrol. 349, 268–276. https://doi.org/10.1016/j.jhydrol.2007.10.058.

Torrence, C., Webster, P.J., 1998. The annual cycle of persistence in the El Nño/SouthernOscillation. Q. J. R. Meteorol. Soc. 124, 1985–2004. https://doi.org/10.1002/qj.49712455010.

Tsidu, G.M., 2012. High-resolution monthly rainfall database for ethiopia: homogeniza-tion, reconstruction, and gridding. J. Clim. 25, 8422–8443. https://doi.org/10.1175/JCLI-D-12-00027.1.

Uvo, C.B., Repelli, C.A., Zebiak, S.E., Kushnir, Y., 1998. The relationships between tro-pical pacific and atlantic SST and Northeast Brazil monthly precipitation. J. Clim. 11,551–562. https://doi.org/10.1175/1520-0442(1998) 011<0551:TRBTPA>2.0.CO;2.

Van Wart, J., Grassini, P., Yang, H., Claessens, L., Jarvis, A., Cassman, K.G., 2015.Creating long-term weather data from thin air for crop simulation modeling. Agric.For. Meteorol. 209–210, 49–58. https://doi.org/10.1016/j.agrformet.2015.02.020.

Vaughan, C., Buja, L., Kruczkiewicz, A., Goddard, L., 2016. Identifying research prioritiesto advance climate services. Clim. Serv. 4, 65–74. https://doi.org/10.1016/j.cliser.2016.11.004.

Vaughan, C., Dessai, S., 2014. Climate services for society: origins, institutional ar-rangements, and design elements for an evaluation framework. Wiley Interdiscip.Rev. Clim. Chang. 5, 587–603. https://doi.org/10.1002/wcc.290.

Weng, H., Ashok, K., Behera, S.K., Rao, S.A., Yamagata, T., 2007. Impacts of recent ElNiño Modoki on dry/wet conditions in the Pacific Rim during boreal summer. Clim.Dyn. 29, 113–129. https://doi.org/10.1007/s00382-007-0234-0.

Yang, Y., Cui, Y., Luo, Y., Lyu, X., Traore, S., Khan, S., Wang, W., 2016. Short-termforecasting of daily reference evapotranspiration using the Penman-Monteith modeland public weather forecasts. Agric. Water Manage. 177, 329–339. https://doi.org/10.1016/j.agwat.2016.08.020.

Yoon, J.-H., Zeng, N., 2010. An Atlantic influence on Amazon rainfall. Clim. Dyn. 2–3,249–264.

Zebiak, S.E., Orlove, B., Muñoz, Á.G., Vaughan, C., Hansen, J., Troy, T., Thomson, M.C.,Lustig, A., Garvin, S., 2015. Investigating El Niño-Southern oscillation and societyrelationships. Wiley Interdiscip. Rev. Clim. Chang. 6, 17–34. https://doi.org/10.1002/wcc.294.

Ziervogel, G., Calder, R., 2003. Climate variability and rural livelihoods: assessing theimpact of seasonal climate forecasts in Lesotho. Area 35, 403–417. https://doi.org/10.1111/j.0004-0894.2003.00190.x.

A. Esquivel et al. Climate Services xxx (xxxx) xxx–xxx

12


Recommended