Rising ecosystem water demand exacerbates the lengthening of tropical dry seasons

Tropical ecosystems, especially rainforests, store about half of Earth’s carbon and sequester ~1.6 ± 0.5 Pg C per year, making them critical buffers against climate change. Their vegetation dynamics and ecosystem functions are highly sensitive to seasonal rainfall, with clear transitions between dry and wet seasons. As dry seasons progress, precipitation deficits and strong evaporative losses lower soil moisture, shift conditions from radiation-limited to moisture-limited, and suppress photosynthesis. Under warming, longer and more intense dry seasons and associated risks (droughts, fires) can reduce productivity, elevate mortality, and promote forest fragmentation and savannization, feeding back to climate. Traditionally, the tropical “dry season” has been defined as periods when daily precipitation (P) is lower than the multi-year average P. Using this definition, prior work has found lengthening dry seasons in the tropics (e.g., ~6.5 d decade−1 in southern Amazonia; 6.4–10.4 d decade−1 in the Congo Basin). However, surface water availability depends on both supply (P) and demand (potential evapotranspiration, Ep) or actual evaporative loss (E). With climate change, atmospheric evaporative demand (Ep) tends to increase, accelerating soil moisture depletion and land drying, while actual evapotranspiration (E) can diverge from Ep due to constraints from soil moisture, vegetation cover, and stomatal conductance. Recent studies have begun to consider P versus Ep or E to define dry conditions, but this balance has rarely been used to assess temporal changes in dry season length and timing. Because P, Ep, and E emphasize different processes, changes in dry season metrics can vary with the chosen definition, complicating synthesis and inference. Objective: This study quantifies how tropical dry season length (DSL), arrival (DSA), end (DSE), and water deficit (WD) have changed across 23.5°S–23.5°N during 1983–2016 under three definitions: (i) P < Ep (supply vs atmospheric demand), (ii) P < E (supply vs actual ecosystem water consumption), and (iii) P < P (supply vs climatological supply). We ask whether accounting for changing atmospheric demand or actual water loss alters the inferred extent and timing of dry season changes relative to precipitation-only approaches.

Prior studies using a precipitation-only definition (P < climatological P) reported widespread tropical dry season lengthening, including ~6.5 days per decade in southern Amazonia and 6.4–10.4 days per decade over the Congo Basin, often linked to changes in large-scale circulation and deforestation feedbacks. However, land surface drying also reflects demand-side changes: rising temperature and declining humidity increase Ep, while E responds to soil moisture availability, vegetation structure, and stomatal behavior. Previous work shows that Ep and E trends can diverge, and that drought and ecological impacts are strongly influenced by atmospheric water demand (e.g., vapor pressure deficit) as well as supply. Inconsistent definitions of “dry season” across studies impede synthesis, motivating an assessment that compares P < Ep, P < E, and P < P definitions for long-term changes in dry season length, timing, and water deficit.

Study domain and period: Global tropical land (23.5°S–23.5°N), 1983–2016. Data: Precipitation (P) from eight daily datasets: CHIRPS, GPCC, CPC Unified (CPC-U), PERSIANN-CDR, ERA-5, MERRA-2, PGF, and MSWEP v2.8. Potential evapotranspiration (Ep) computed via FAO Penman–Monteith using daily meteorology (net radiation, air temperature, specific humidity/relative humidity, wind speed, pressure) from three reanalyses: ERA-5, MERRA-2, GLDAS-2.0 (GLDAS forced by PGF where needed). Actual evapotranspiration (E) from GLEAM v3.3a (daily), which assimilates satellite soil moisture and vegetation optical depth (VOD) and uses multi-source P and reanalysis temperature/radiation. Land cover from MODIS MCD12C1 (IGBP) to mask water-dominated and sparsely vegetated areas; VOD from VODCA Ku-band (1987–2016). CMIP6: daily outputs (1983–2014 historical) from 34 coupled models for P, latent heat (converted to E), and meteorological variables for Ep. Seasonality diagnostics: Applied Fourier harmonic analysis to daily time series to determine uni- vs bi-modal seasonality (ratio of amplitudes at two vs one cycle per year; threshold 0.75). For each grid and each definition, smoothed daily P, Ep, and E with a 30-day running window and computed cumulative anomalies: A(d)=Σ(Pj−Epj), B(d)=Σ(Pj−Ej), C(d)=Σ(Pj−Pj). Defined dry season arrival (DSA) as day of maximum cumulative value and dry season end (DSE) as day of minimum. Ensured continuity and treated biannual regimes separately by identifying the two longest dry seasons (typically DJF and JJA). Computed dry season length (DSL) as DSE−DSA (sum across two seasons where applicable). Water deficit (WD) was the cumulative P−Ep, P−E, or P−P between DSA and DSE. Trend analysis: Computed annual DSA, DSE, DSL, and WD for 1983–2016 using windows DSA−60 to DSE+60 (45-day buffers for biannual regions) to capture season-spanning periods. Estimated linear trends via ordinary least squares (slope and two-tailed t-test), with additional Mann–Kendall tests for robustness. Aggregated regional time series (area-weighted) for southern Amazonia, central Africa (northern/southern), and southwestern Africa. Consistency classification across datasets: For each definition, classified grid trends as “very likely” (same significant sign in 6–8 P datasets; others non-significant), “likely” (4–5), “probably” (1–3), “uncertain” (conflicting significant signs), or “no change” (no significant trends). Excluded arid and humid regions from P < Ep and P < E trend area fractions where no climatological dry/wet season exists by these definitions. Driver analysis: Decomposed Ep trend contributions from meteorological drivers (air temperature T, relative humidity RH, wind speed u2, net radiation Rn) by recalculating Ep with each variable fixed at daily climatology to isolate its effect. For E, examined GLEAM evaporative stress factor S, soil moisture, and VOD trends to diagnose constraints from water availability and vegetation water content. Compared observational findings with CMIP6 model outputs using the same diagnostics. Uncertainty assessment: Discussed gauge sparsity, gridding/interpolation differences, and reanalysis biases, especially over Amazonia and central Africa; noted reliance on a single daily E dataset (GLEAM) and potential unrepresented processes (CO2 physiology, nitrogen deposition, land cover change) affecting E.

• Seasonality and baseline patterns: 87.4% of tropical land exhibits one distinct dry season per year; bimodal regimes occur mainly near the equator (e.g., Congo Basin, East Africa, parts of Amazonia and Southeast Asia). DSL differs markedly by definition: P < Ep and P < E show strong latitudinal gradients (0 to >200 days), while P < P yields relatively homogeneous DSL with >86% of tropical areas between 150–240 days, even in humid rainforests where other definitions show short or no dry season.
• Inter-metric differences by climate regime: In arid regions (P << Ep), DSL is longest under P < Ep; P < P and P < E are similar. The P < Ep minus P < E DSL difference increases from ~30–40 days at MAP ~1000 mm/yr to ~170–190 days at MAP ~200 mm/yr. In humid regions (P >> Ep), DSL is longer under P < P; P < Ep and P < E are similar, with differences >120 days when MAP >2500 mm/yr. About 39.2% of tropical grids (MAP 1000–1500 mm/yr) show little difference among definitions.
• Extent of dry season lengthening (1983–2016): All definitions indicate more widespread drying than wetting. Fraction of tropical land probably experiencing longer dry seasons: P < E 48.7% (largest), P < Ep 43.1%, P < P 33.7%. Fractions with shorter dry seasons: ~20.4% (P < Ep), 17.1% (P < E), 19.8% (P < P). Regions with no decisive change or uncertainty: 35.8% (P < Ep), 33.4% (P < E), 40.5% (P < P).
• Water deficit trends during dry season: Increasing WD over ~55.7% (P < Ep), 51.7% (P < E), and 38.0% (P < P) of tropical land; decreasing WD over ~16.8%, 17.5%, and 19.2%, respectively. Independent Ep products give similar drying fractions (MERRA-2 lengthening 41.5%, GLDAS 42.9%; WD increase 49.9% and 52.0%).
• Hotspots and magnitudes: Southern Amazonia and central Africa show robust DSL lengthening across definitions. Southern Amazonia DSL increase: 4.81–10.84 d/dec (P < Ep), 3.65–11.96 d/dec (P < E), 5.95–10.53 d/dec (P < P), mainly due to delayed DSE. Central Africa (June–August dry season) DSL increase: 7.62–11.61 (P < Ep), 6.62–11.65 (P < E), 8.19–11.93 d/dec (P < P), due to both earlier DSA and delayed DSE.
• Regional contrasts and decoupling: Southwestern Africa shows contradictory trends: P < E indicates lengthening (5.30–16.51 d/dec) with increasing E (+0.14 mm d−1 dec−1), while P < Ep shows shortening (−5.66 to −9.24 d/dec) as Ep decreases (−0.17 mm d−1 dec−1); P exhibits little trend. Increased wet-season precipitation and storage can sustain higher E into the dry season, yielding P < E drying even without dry-season P increases.
• Drivers: Rising atmospheric water demand during dry seasons is driven primarily by decreasing RH (especially late dry season), with additional contributions from higher temperature and wind speed; radiation changes are negligible. E increases are moderated or reversed where soil moisture declines and vegetation water content decreases (stomatal closure), producing smaller P < E drying than P < Ep in some regions (e.g., southern Amazonia).
• Models vs observations: CMIP6 generally shows more widespread and robust drying under P < Ep but underestimates areas with P < E lengthening (13.4% vs 48.7% observed) and does not capture observed P < E lengthening over southern Africa and Sahel, likely due to representation of CO2-physiology effects on E in models absent from observation-based E estimates.

Accounting for atmospheric water demand (Ep) and actual water loss (E) substantially alters the inferred extent and timing of tropical dry season changes compared to precipitation-only definitions. In a warming climate, decreased relative humidity and increased temperatures raise Ep, exacerbating drying and extending dry seasons (especially delayed DSE), notably in southern Amazonia and central Africa. However, E trends can decouple from Ep due to soil moisture limitations, vegetation water content declines, and stomatal regulation, moderating P < Ep-based drying in some regions. Conversely, enhanced wet-season water storage can elevate E during the subsequent dry season, producing P < E lengthening where P and Ep trends alone might suggest the opposite (e.g., southwestern Africa). These dynamics imply that precipitation-only metrics (P < P) may underestimate dry season lengthening and ecosystem water deficits, while P < Ep is informative for demand-driven risks (heat, drought, wildfire) and P < E better reflects surface water resource dynamics. Discrepancies between observations and CMIP6, including smaller model-derived P < E drying, suggest important roles for plant physiological responses to rising CO2 and vegetation changes that complicate large-scale E estimation and should be reconciled to improve projections of tropical hydrology and ecosystem impacts.

The study demonstrates that the perceived lengthening of tropical dry seasons depends strongly on how “dry season” is defined. Incorporating changes in atmospheric water demand (P < Ep) or actual evaporation (P < E) reveals more extensive and robust dry season extensions and increased water deficits across the tropics than precipitation-only (P < P) approaches, with southern Amazonia and central Africa as consistent hotspots. Climate change alters both water supply and demand, and ecosystem constraints modulate E, so assessments of seasonal dryness and ecological risk should explicitly account for changing water demand and actual water loss. Future work should reduce uncertainties in E estimation (including daily products), incorporate plant physiological and structural responses (e.g., CO2 effects, phenology), improve observational coverage in data-sparse tropical regions, and reconcile observation–model discrepancies to enhance confidence in projections of tropical water availability and ecosystem responses.

• Data sparsity and biases: Tropical gauge networks are sparse and uneven; gridding/interpolation methods differ across products. Reanalysis datasets can exhibit regional biases (e.g., Amazonia, central Africa).
• Evapotranspiration uncertainty: Only one daily, observation-driven E dataset (GLEAM) was available; E is influenced by complex processes (soil moisture dynamics, vegetation physiology/structure, CO2 fertilization, nitrogen deposition, land cover change) not fully captured.
• Definition constraints: Arid and humid regions lack meaningful dry/wet seasons for P < Ep and P < E, reducing area coverage for trend assessment under those definitions.
• Model–observation discrepancies: CMIP6 models and observation-based diagnostics differ in the extent and loci of DSL changes, particularly for P < E; representation of CO2-physiology in models versus its absence in observation-based E estimates contributes to inconsistency.
• Regional uncertainty: A substantial fraction of tropical land shows no decisive trend due to dataset disagreements and variability, limiting generalizability at local scales.