Widespread spring phenology effects on drought recovery of Northern Hemisphere ecosystems

Extreme droughts are increasing in frequency and severity, threatening ecosystem stability and the terrestrial carbon sink. A key aspect of ecosystem resilience is the time required to recover from severe drought to predrought functional states. While recovery within a single growing season has been studied, the interactions between vegetation phenology (start and end of growing season), drought timing, and intensity remain understudied. This study focuses on mid- and high-latitudinal Northern Hemisphere ecosystems, which exhibit pronounced seasonality and high drought susceptibility. The authors aim to quantify divergent drought recovery trajectories—within a single growing season versus extending into the subsequent growing season—and to assess how vegetation phenology, particularly spring phenology, influences recovery. They hypothesize: (1) earlier spring phenology in the drought year lengthens recovery due to biophysical feedbacks when drought occurs mid-growing season; and (2) delayed spring phenology in the subsequent year postpones growth and lengthens recovery due to biological processes. Understanding these mechanisms is important for improving ecosystem resilience assessments and Earth system model representations.

Prior work shows drought recovery patterns and legacy effects across biomes, including that recovery times can vary with drought timing, intensity, and ecosystem type. Studies have indicated widespread drought legacies in forests, links between phenology and climate change, and seasonal carryover effects where spring conditions influence subsequent seasonal growth and water use. Earlier spring greening can exacerbate summer soil drying and affect productivity. Snow and winter temperature also modulate spring phenology and subsequent growth, with regionally variable impacts. However, explicit quantification of how phenological variations interact with drought timing and climate anomalies to shape recovery trajectories has been limited, motivating this study.

Study region and scope: Mid- and high-latitudinal Northern Hemisphere regions with a single distinct growing season. Vegetation growth proxies: Three remote sensing datasets were used and resampled to 0.5° resolution: (1) NDVI (GIMMS NDVI3g, biweekly, 0.083°, 1982–2015), primary proxy for analysis; (2) CSIF (all-sky daily SIF, 0.05°, 2000–2016; analyzed 2001–2015); (3) VOD (AMSR-E/AMSR2 X-band LPDR v2, daily, 0.25°, 2003–2015; analyzed 2003–2015). Climate and ancillary data: CRU TS4.02 monthly precipitation, temperature, PET (0.5°, 1982–2015); VPD derived from CRU; GlobSnow SWE (25 km, 1982–2015); SPEI (0.5°, 1982–2015) as drought metric, with main analyses using SPEI3 and sensitivity to SPEI1/6/9/12; soil sand content from HWSD v1.2; Köppen–Geiger climate zones; land cover (GLC2000) regrouped into evergreen forests, deciduous forests, shrubs, and grasses. Phenology estimation: NDVI time series filtered and interpolated to daily using HANTS-Maximum; SOS and EOS per pixel retrieved with a threshold method (minimum + 30% of seasonal amplitude) on multiyear smoothed NDVI. Pixels without clear single-season phenology were excluded (~81% of vegetated land retained). Seasons defined per pixel: early-GS from SOS to start of mid-GS, mid-GS as the two consecutive months with maximum NDVI (no earlier than April, no later than October), late-GS from end of mid-GS to EOS. Dormant season from EOS to next SOS. Drought event identification: Extreme drought events require SPEI3 < -2 and contemporaneous reduced vegetation greenness (NDVI during drought period below 1982–2015 mean) within early-, mid-, or late-GS. Events end when SPEI3 > -2. Only single extreme drought events were considered; any grid cell with another extreme climate event (drought or wetness) within ±4 years was excluded to avoid compound/legacy effects. Comparable definitions were tested for SPEI1/6/9/12. Drought response metrics: - Drought response lag: months between minimal SPEI3 anomaly and maximum NDVI suppression caused by the event. - Drought sensitivity: Pearson correlation between mean growing season NDVI and mean growing season SPEI3 (1982–2015). Drought recovery definition and trajectories: Monthly SPEI3 and NDVI were smoothed with a 3‑month forward moving window, deseasonalized, and linearly detrended. For each pixel, the long-term average baseline is the multiyear mean of detrended NDVI (1982–2015). Recovery duration is months from the deepest NDVI suppression to the month NDVI returns to within 95% of the long-term baseline. Two trajectories: (1) Recovery within the same growing season before dormancy (R_SGS; referred to as R_CS in Methods); (2) Recovery extending into the subsequent growing season (R_MGS), calculated as total recovery period minus the length of dormancy to avoid inflating duration by the dormant period. Only events fully contained in 1982–2015 were analyzed. CSIF- and VOD-based recovery were computed with the same protocol for comparison. Attribution modeling: Separate Random Forest (RF) regression models were built for each of eight groups: two recovery modes (R_SGS and R_MGS) × four vegetation types (evergreen, deciduous, shrubs, grasses). For R_SGS (n=29,614 events), predictors included: SOS_drought year (spring phenology), Precip and VPD anomalies in predrought (6 months), lag, and postdrought periods (Precip_pre, VPD_pre, Precip_lag, VPD_lag, Precip_post, VPD_post), NDVI_pre (mean NDVI anomaly in the 6 months within growing season preceding drought), long-term climate (MAP, MAT, water balance = MAP − mean annual PET), interannual variability (Precip_cv, Temp_cv), drought sensitivity, and sand fraction. For R_MGS (n=42,733 events), additional predictors: SOS_subsequent year, dormant length, SWE_dorm, Temp_dorm, and drought response lag. Predictor anomalies were computed from deseasonalized, detrended series. Model settings: 500 trees, one randomly chosen covariate per split, minimal terminal node size = 5. Multicollinearity control via pairwise correlation screening (R>0.5 led to removal of the variable with lower correlation to recovery). Variable importance assessed via permutation importance; partial dependence plots used to depict marginal effects; bootstrap used for mean predictions and 95% CIs. Performance: RF captured recovery well, with R² ≈ 0.71–0.75 for R_SGS and 0.86–0.90 for R_MGS across vegetation types. Additional analyses: Relationships examined across Köppen–Geiger climate zones and vegetation types to test spatial robustness. Sensitivity tests across SPEI aggregation timescales (1, 6, 9, 12 months).

• Recovery prevalence and spatial patterns: More than half of ecosystems in mid- and high-latitudinal Northern Hemisphere did not recover within a single growing season after extreme drought. Specifically, NDVI failed to fully recover within a single growing season in ~50% of early-GS droughts, >60% of mid-GS droughts, and >80% of late-GS droughts. Recovery times were longer in central North America, the Mediterranean, and central Eurasia; in northern latitudes (>50°N), 84% of extreme drought events recovered within less than 3 months. Mean recovery time decreased with latitude and with increasing aridity. Consistent patterns were obtained using different SPEI timescales and with CSIF and VOD proxies. - Recovery by timing and vegetation: Recovery after early-GS droughts was on average longer than after mid- or late-GS droughts. Under R_SGS, early-GS recovery was 0.4 months longer than mid-GS (P<0.05). Under R_MGS, early-GS recovery was 0.6 and 1.8 months longer than mid- and late-GS, respectively (P<0.05). Deciduous forests exhibited the longest recovery among vegetation types under both trajectories. - Phenology effects (drought year): Spring phenology significantly affected R_SGS with a bimodal response overall: both earlier and delayed SOS in the drought year could lengthen recovery. Disaggregating by drought timing showed that earlier SOS shortened recovery for early-GS droughts but lengthened recovery for mid-GS droughts. - Phenology effects (subsequent year): For R_MGS, delayed SOS in the subsequent year consistently lengthened recovery, while earlier SOS shortened it. The importance of spring phenology (subsequent year) ranged from 46% to 58% across vegetation types, exceeding or matching other factors. - Dormancy and winter conditions: Climate during dormancy influenced R_MGS but was less important than spring phenology. Negative temperature anomalies during dormancy delayed recovery. Both negative and positive SWE anomalies lengthened recovery, especially for deciduous forests and shrubs (importance ~20% and 19%, respectively). Longer dormant length was associated with shorter recovery. - Hydroclimate during recovery: Negative precipitation anomalies and positive VPD anomalies during the postdrought period delayed recovery under both R_SGS and R_MGS (importance up to ~52% and 34% for precipitation, and ~39% and 23% for VPD, respectively). Higher mean annual temperature (MAT) was associated with longer recovery in both trajectories. Predrought precipitation, long-term water balance, and interannual climate variability (Temp_cv, Precip_cv) had marginal or no effects. - Preceding growth condition: Lower NDVI_pre (negative anomalies) delayed recovery; NDVI_pre had notable importance, particularly for R_SGS (25–31%). - Soil and sensitivity metrics: Soil sand fraction and drought sensitivity had no significant effects; drought response lag and hydroclimate during the lag period were generally unimportant. - Model robustness: Relationships held across climate zones and vegetation types; consistent findings across vegetation proxies (NDVI, CSIF, VOD) and SPEI timescales.

The study demonstrates that vegetation phenology is a key determinant of drought recovery, modulated by drought timing and vegetation type. Mechanistically, earlier spring phenology in the drought year can extend the growing season and enhance spring growth, increasing evapotranspiration and drawing down soil moisture. This biophysical pathway intensifies progressive water stress, particularly affecting summer and autumn when growth is most water-limited, thereby lengthening recovery for mid- and late-GS droughts. Earlier spring can also increase frost risk and deplete carbohydrate reserves, further constraining recovery. Conversely, for early-GS droughts, earlier spring phenology can shorten recovery because positive biological carryover (enhanced photosynthesis, carbon and nutrient allocation, improved physiological status) may outweigh biophysical soil moisture depletion. When recovery extends into the subsequent growing season (R_MGS), delayed spring phenology in the subsequent year slows vegetation growth and extends recovery across all vegetation types and drought timings; this emerged as a dominant control relative to direct climate anomalies during the drought year. Winter conditions modulate recovery through effects on spring phenology and soil moisture dynamics: higher SWE and lower winter temperature can delay snowmelt and phenology, while insufficient snow or frozen soils can reduce infiltration and spring/summer soil moisture, prolonging recovery. Preceding growth conditions matter, likely via enhanced drought resistance, carbon reserves (non-structural carbohydrates), and root development enabling deeper water access; effects were more pronounced in woody vegetation for R_MGS. Elevated mean annual temperatures exacerbated phenology-related impacts, generally slowing recovery due to reduced water availability. Contrary to some expectations, interannual climate variability, drought sensitivity, and response lags were not major determinants of recovery duration in this analysis. Collectively, the findings highlight that phenology–drought interactions are central to ecosystem resilience assessments and should be incorporated into Earth system models.

This study identifies two distinct drought recovery trajectories—within the same growing season and extending into the subsequent growing season—and shows that more than half of extreme drought events in mid- and high-latitudinal Northern Hemisphere do not recover within a single season. Spring phenology exerts pervasive and often dominant control on recovery: earlier SOS in the drought year can either shorten or lengthen recovery depending on drought timing, while delayed SOS in the subsequent year consistently prolongs recovery across vegetation types. Hydroclimate during the postdrought period (precipitation and VPD), winter conditions (temperature and SWE), and mean annual temperature further modulate recovery, but many long-term or predrought metrics play smaller roles. The results underscore the need to explicitly represent phenology–drought interactions in Earth system models to improve predictions of ecosystem resilience under a warming climate with more frequent and severe droughts. Future research should: (1) elucidate physiological mechanisms underpinning divergent phenology effects across vegetation types; (2) expand analyses to regions with multiple growing seasons and to compound/extreme event sequences; (3) integrate additional structural/physiological observations (e.g., carbon reserves, hydraulic traits) with remote sensing; and (4) test and improve model phenology and soil moisture feedback representations under changing drought seasonality.

• Event selection: Only single extreme drought events were analyzed; events with other extremes within ±4 years were excluded, which may limit generalizability to compound or closely spaced events. - Region/pixel screening: Analyses were restricted to pixels with a single, clear growing season and to events fully contained within 1982–2015, potentially excluding important ecosystems (e.g., multiple growing seasons). - Proxy reliance: Main analyses used NDVI; although validated with CSIF and VOD, each proxy has sensitivities and uncertainties (e.g., NDVI saturation, VOD retrieval assumptions). - Phenology retrieval: Threshold-based SOS/EOS estimation and resampling may introduce uncertainties, especially in heterogeneous landscapes. - Late-GS R_SGS sampling: Late-season droughts under R_SGS were not modeled in some analyses due to insufficient samples, potentially biasing comparisons. - Attribution limits: RF models infer associations and variable importance but not causality; multicollinearity control may remove informative variables. - Soil characterization: Using sand fraction as a single soil descriptor may not capture full soil hydraulic variability. - Physiological mechanisms: Divergent effects of preceding growth conditions among vegetation types are hypothesized but not directly measured; physiological bases remain uncertain.