Intra-growing season dry-wet spell pattern is a pivotal driver of maize yield variability in sub-Saharan Africa

The study addresses how intra-seasonal variability of dry-wet spells, beyond mean climate changes, drives interannual maize yield variability in sub-Saharan Africa (SSA). Agriculture is sensitive to weather and climate variability, and previous work suggests temperature and precipitation during the growing season can explain a substantial fraction of global yield fluctuations. In SSA—a region facing significant food security risks—there is debate over the relative roles of temperature and precipitation on maize yields, with evidence for strong temperature sensitivity but marked regional heterogeneity. Traditional assumptions of stationary climate–yield relationships are increasingly untenable under climate change. The purpose here is to quantify the influence of sub-seasonal dry-wet spell patterns across maize growth stages on yield variability and to map associated regional weather-related risks while accounting for spatial heterogeneity in crop responses.

Two complementary approaches dominate climate–yield characterization: data-driven statistical models and process-based crop models. Reported limitations in prior work include assuming time-invariant climate–yield relationships, focusing more on mean climate changes than interannual variability, analyzing single drivers or isolated extremes rather than compounded effects, and using coarse spatio-temporal indicators (seasonal means, national scales) that can mask topographic and bioclimatic diversity. Global studies indicate climate variability can explain a large share of yield variability and that extreme weather indices can be as or more informative than mean indices. For SSA maize, prior findings on temperature dominance vary by region and the direction of yield responses, underscoring the need for finer temporal resolution and spatially explicit analyses of sub-seasonal compound dry–wet and hot–dry events.

Study area and period: SSA, analyzed primarily over 1982–2009, with comparisons between 1982–1991 and 2000–2009. Regions were aggregated as Western Africa (WA), Eastern Africa (EA), and Southern Africa (SA). Crop calendars and stage lengths were region-specific. Data sources: (1) Climate: AgMERRA daily forcing (1980–2010). (2) Maize yields: GDHY v1.2+v1.3 at 0.5° (1981–2016). (3) Growing seasons: GGCMI Phase 2 (planting dates and season lengths). (4) Agroecological zones: GAEZ. Agroclimatic index: Standardized Precipitation Evapotranspiration Index (SPEI) at dekadal (10-day) timescale. Evapotranspiration computed via FAO Penman–Monteith. Aggregated effective precipitation fitted with a generalized extreme value distribution; standardized to N(0,1) and bounded to [−3,3]. Goodness-of-fit checked with Anderson–Darling and Kolmogorov–Smirnov tests. SPEI classification: Objective quantile thresholds defined six classes: very dry (A, <0.1), dry (B, 0.1–0.3), moderately dry (C, 0.3–0.5), moderately wet (D, 0.5–0.7), wet (E, 0.7–0.9), very wet (F, >0.9). Dekadal SPEI sequences were partitioned by maize development stages (initial, development, mid-season, late), with stage lengths as fractions of the season length from GGCMI. Temperature vs precipitation contributions: A random forest classifier used 10-day mean precipitation and temperature to predict SPEI class per stage; variable importance quantified the relative contributions of temperature vs precipitation by stage and intensity. Class frequencies and their temporal trends were evaluated via Sen’s slope and the Mann–Kendall test (two-sided, 90% confidence). Sequence similarity: Original stage-wise SPEI class sequences were compared against six reference sequences (uniform class sequences A–F). Similarity used a Needleman–Wunsch penalty matrix with adjacent-class penalties increasing up to 5 for A vs F; similarity S=1−(Σ penalties)/(5n). Bayesian network (BN): A five-node discrete BN (four stages + yield) fit via the bnlearn R package using multiple combinations of stage-wise similarity measures (1,296 combinations) to compute joint/conditional probabilities of compound dry–wet spells across stages (2,400 combinations ranging from single-stage to four-stage events). For impact assessment, a continuous BN variant used evidence propagation (BayesNetBP) with similarity set to 1 for target patterns to derive shifts in yield distributions. Risk mapping: A risk matrix combined (i) likelihood (probability) of specific dry–wet spell patterns and (ii) severity (yield impact) into scores 0–5 for each, multiplied and min–max normalized to create regional risk maps. Multiple linear regression (MLR): First-differences approach relating 10-year time series of yield changes to (Model 1) first differences of mean precipitation and temperature at the four stages, and (Model 2) log-transformed first differences in joint probabilities of multiple dry–wet spell patterns. Iterative feature selection based on t-test P-values yielded parsimonious models (median 6–7 predictors), assessed using adjusted R² to enable comparison across differing predictor counts. Trend analyses: Climate and dry/wet spell trends assessed in 10-year blocks using Sen’s slope and Mann–Kendall tests.

- Mean climate trends (1982–2009): Temperature increased significantly across most SSA during the growing season: WA ~0.15 °C/decade, SA ~0.21 °C/decade, EA ~0.27 °C/decade. Rainfall trends were spatially heterogeneous: positive in much of SA (e.g., +44 mm/decade on average, significant in Angola and N. Namibia), positive in the Sahel, negative in parts of the Gulf of Guinea and Sierra Leone, and mixed across EA (negative early season, positive mid/late season; significant declines >10% in the Horn from 1982–1991 to 2000–2009). - Stage climatology: Initial stage was warmest; temperature decreased through the season (mean seasonal decline 0.4 °C in SA to 1.2 °C in WA). Average rainfall contributions by stage: initial 12–22%, development 27–29%, mid-season 30–36%, late 21–25%. - Dry spell dynamics: Relative temperature contribution to dry-spell onset peaked at the initial stage and rose with event intensity. For extremely dry spells (SPEI<0.1): initial-stage temperature contribution exceeded 50% in SA semi-arid and sub-humid zones, 45–50% in EA, and 34–36% in WA; at mid-season it dropped to 41–45% (SA), 39–43% (EA), and 27–33% (WA). For moderately dry events, temperature’s contribution was 10–20 percentage points lower than for extreme dry events. - Dry spell trends: Overall declines in dry-spell frequency over 1982–2009: EA decreased ~11 events/decade (significant in Somalia and E. Kenya), WA ~7 events/decade, SA ~10 events/decade; localized increases (e.g., W. EA including South Sudan and Uganda) and significance in NW WA and parts of SA (N. Namibia, Botswana, Zimbabwe, S. Mozambique). - Yield characterization and trends: Mean rainfed maize yields remained low and zone-dependent: semi-arid 0.9–1.4 t/ha, sub-humid 1.4–1.5 t/ha, humid 1.5–1.8 t/ha. Regional yield trends: positive in western EA (+10–30 kg/ha/yr, significant 90%), mixed in EA (negative but often non-significant in some Somalia/Kenya areas), positive in WA (+20 kg/ha/yr average; up to +60 kg/ha/yr in Côte d’Ivoire), positive in much of SA (+10–30 kg/ha/yr in Namibia/Mozambique; >+100 kg/ha/yr in E. South Africa), negative in Zimbabwe (~−20 kg/ha/yr) and N. Botswana. - Explained variance in maize yields: Using mean precipitation and temperature by stage explained ~30–35% of interannual yield variance regionally. Using changes in stage-wise dry–wet spell patterns increased explained variance to ~50–60% regionally, with strong spatial variability (up to ~75% in the Horn of Africa; ~30% near Lake Victoria). - Dominant yield-driving patterns: Four sub-seasonal patterns dominated yield variability: (i) dry-dry-dry-dry, (ii) wet-wet-wet-wet, (iii) dry-wet-wet-wet, and (iv) wet-dry-dry-dry. - Risk changes tied to patterns: From 1982–1991 to 2000–2009, dry-dry-dry-dry became 2–4× less likely (risk generally declined). Wet-wet-wet-wet became 1.5–2.5× more likely (risk changes spatially heterogeneous: decreased in Sierra Leone/Guinea; increased in E. EA and N. Botswana). Dry-wet-wet-wet increased 1.2–1.7× (risk rose in the Horn; decreased in rainiest WA and parts of SA; slight increase in N. Namibia). Wet-dry-dry-dry became 1.5–2.5× less likely (risk declined in EA, WA—especially Burkina Faso, Nigeria—and most of SA except E. SA, e.g., Mozambique, Zimbabwe). - Production and food energy impacts: Estimated mean yield decrease under dry-dry-dry-dry was 0.3–0.6 t/ha, affecting 61% (EA), 81% (WA), and 90% (SA). Annual net production losses from this pattern equate to the yearly energy needs of ~31 million people (and ~18 million for wet-dry-dry-dry), assuming 34 Mha harvested (2021–22), 3,560 kcal/kg, and 2,100 kcal/person/day. Conversely, wet-wet-wet-wet and dry-wet-wet-wet led to net production increases equivalent to feeding ~34 and ~21 million people per year, respectively.

The findings demonstrate that sub-seasonal dry–wet spell patterns are pivotal drivers of maize yield variability in SSA, more so than mean seasonal temperature and precipitation alone. This challenges stationary and mean-based approaches to climate–yield relationships and underscores the importance of timing and compounding conditions across growth stages. Temperature’s heightened role in initiating extreme dry spells, especially during the initial stage, aligns with the heightened risks of compound hot–dry events and their disproportionate yield impacts. Despite widespread warming, declines in dry-spell frequency across many SSA areas reduced risks associated with persistently dry seasons, though spatial heterogeneity persists. Mapping risk tied to stage-specific patterns offers actionable insights for climate-smart agriculture, enabling regional targeting of interventions (e.g., varieties, planting windows, water management) aligned with local agroclimatic regimes and the most consequential intra-seasonal patterns. The enhanced explained variance (50–60%) using pattern-based predictors indicates that integrating sub-seasonal variability and compound-event likelihoods can substantially improve yield variability attribution and risk assessment at regional scales.

This study shows that shifts in intra-seasonal dry–wet spell patterns across maize growth stages explain a substantially larger fraction of interannual yield variability (50–60%) in SSA than changes in mean seasonal climate (30–35%). A small set of patterns—uniformly dry or wet seasons and seasons with anomalous initial-stage conditions relative to subsequent stages—dominate yield responses. Temperature is particularly influential in triggering extreme dry spells early in the season. The pattern-based Bayesian risk framework, coupled with dekadal SPEI and stage-specific analysis, enables regional risk mapping under data-limited conditions and can inform climate-smart agricultural decisions in SSA. Future research should (i) evaluate additional climate datasets and teleconnections, (ii) integrate with multi-model crop simulations (e.g., AgMIP/GGCMI) under diverse climate scenarios, (iii) test nonlinear/statistical learning methods against linear models, (iv) derive adaptive, time-varying planting dates based on rainy-season onset, and (v) incorporate biotic and management drivers to refine attribution and projections.

- Data quality and availability: Agricultural statistics in SSA are often low quality and untimely, introducing uncertainty in yield estimates and model validation. Climatic forcing datasets differ most where in situ observations are sparse (notably SSA), with larger discrepancies for precipitation, distributional characteristics, and extremes than for means. - Statistical approach constraints: Results are constrained by the historical envelope of variability inherent to statistical models, potentially limiting extrapolation under future climates. - Fixed crop calendars: Use of static planting dates overlooks adaptive shifts; although a 10-day aggregation mitigates onset-shift effects around planting, optimal time-varying planting dates are recommended. - Omitted drivers: Biotic pressures (pests, weeds), inputs, management practices, and socio-political factors were not explicitly modeled; first differences reduce but do not eliminate their influence. - Regional heterogeneity and resolution: Coarse yield data (0.5°) and regional aggregation (WA/EA/SA) may mask local-scale processes and management heterogeneity affecting yields and weather interactions.