Late Ming Dynasty weak monsoon induced a harmonized megadrought across north-to-south China

The evolution of the Asian monsoon has strongly influenced Chinese civilizations through its impacts on agriculture, economy, and social stability, with weak monsoon periods linked to peasant uprisings, epidemics, and dynastic transitions. The late Ming dynasty megadrought (LMDMD) is widely regarded as a key climatic event contributing to the Ming collapse. To clarify the spatiotemporal structure and societal implications of this event, the authors developed a tree-ring width chronology from Keteleeria davidiana in the Daba Mountains (middle Yangtze). They reconstructed April–November scPDSI over the past 464 years and posed three core research questions: (1) Was there a consistent spatial structure of weak monsoon events during the late Ming period across the Asian monsoon region in China? (2) What were the economic and societal impacts of long-term changes in water balance? (3) What were the driving mechanisms of the regional drought dynamics?

Previous studies reconstructed the late Ming weak monsoon using diverse proxies, including stalagmites, lake sediments, historical documents, and tree rings. These works used mutual validation and climate model sensitivity experiments to attribute the LMDMD to external forcings (e.g., temperature anomalies, weakened solar activity), internal land–sea heat exchange variability, and ocean–atmosphere modes. The Monsoon Asia Drought Atlas (MADA) reflects monsoon variability across Asia but tree-ring coverage in humid southern China remains sparse. Many non-tree proxies suffer from limitations such as dating uncertainties and ambiguous physiological significance, motivating the development of robust tree-ring datasets in southern China to quantify hydroclimatic changes and link them to socioeconomic outcomes.

Study area and species: Sampling was conducted in the Daba Mountains, the southern transition zone of China, south of the Qinling–Huai line and north of the middle Yangtze, in a subtropical monsoon climate. The species Keteleeria davidiana is a tall, scattered conifer common in southern mountainous/hilly regions and forms mixed forests with Pinus massoniana, Liquidambar formosana, and Cunninghamia lanceolata. Sampling and measurement: Using 80 cm increment borers, 1–3 cores were extracted at ~1 m height from each tree across 31.1°–32.6° N, 107.7°–111.0° E (mean elevation 788 m). In total, 95 cores from 43 trees were collected. Cores were air-dried, mounted, polished, and scanned at 2400 dpi (Epson Expression 12000XL). Ring widths were measured with CooRecorder 9.4 (precision ~0.001 mm). Crossdating and chronology development: Dating accuracy was checked with COFECHA. Growth trends were removed using the Friedman super smoother (alpha=7), and series were combined into a regional standard chronology (STD) via a double-weighted robust mean in ARSTAN. Chronology variance was stabilized following Osborn et al. Robustness was assessed using EPS and Rbar computed in 25-year windows with 50-year overlap. The chronology was truncated at 1560 following criteria EPS ≥ 0.85 and minimum depth ≥ 6 cores (≥ 3 trees). Climate data and growth–climate analysis: Instrumental climate data were from nearby stations (XS, ZP, LG, ZB, WY) and CRU TS 4.0 gridded products. Variables included monthly mean temperature (T), total precipitation (P), and scPDSI. Pearson correlations between the ring-width chronology (RC) and climate were computed from June of the previous year to December of the current year. RC correlated negatively with previous Sep and Dec T, and current Jul and Dec T; positively with previous Sep P and current Jan, Apr, Jun, Aug–Oct P. RC showed persistent positive correlation with scPDSI from previous Sep to current Dec, peaking for Apr–Nov scPDSI (r=0.56, P<0.001, n=97). Reconstruction target and modeling: April–November scPDSI for the middle Yangtze region (near Yichang 30.7°N, 111.3°E, 587 m; Chongqing 29.5°N, 106.4°E, 351 m) during 1924–2020 served as the target. Seven regression models were trained with RC as predictor: Linear Regression (LR), Neural Network (NN), Support Vector Machine (SVM), Linear SVM (LSVM), K-Nearest Neighbor (KNN), XGBoost, and Random Forest (RF). Data were randomly split 50/50 into training/testing; each model underwent hyperparameter optimization in Python (parameters in Table S3). Performance was evaluated using R², RE1, PMT, RE2, ST, NSEC, RMSD, and KGE. Ensemble approach: Nonlinear ML models (e.g., XGBoost, RF) generally outperformed LR, though NN underperformed among ML models. To balance biases, avoid overfitting, and capture extremes and wet–dry amplitudes, an ensemble average of all models was adopted, yielding R² = 45.2% (1924–2020). Event detection and return periods: The reconstruction was smoothed with a 10-year LOWESS filter to identify multi-year events. Wet/dry/extreme years were classified via mean (M) and standard deviation (SD): dry ≤ M−1 SD, wet ≥ M+1 SD, extreme at M±2 SD. Continuous wet/dry runs were defined on the low-pass series relative to the multi-year mean. Duration and cumulative scPDSI severity were computed, and a Copula-based joint probability model estimated return periods for run duration–severity combinations. Spectral and spatial analyses: Spectral properties were assessed using MTM with red-noise testing; EEMD complemented spectral insights. Spatial correlations with observed scPDSI, precipitation, runoff, and temperature were computed for 1924–2020. Broader drivers were examined via correlations with SST (HadISST1, 1870–2023), and comparisons with reconstructions of PDO, ENSO, total solar irradiance (TSI), and Asian volcanic eruption chronologies. A 21-year sliding correlation assessed time-varying relationships. Superposed epoch analysis (SEA) with double-bootstrap evaluated scPDSI responses to extreme PDO/ENSO phases and volcanic eruptions over ±7-year windows.

Reconstruction characteristics: The April–November scPDSI reconstruction spans 1560–2023 (464 years) with long-term mean ~0.01 and SD ~0.8, indicating coherent regional moisture variability. After 10-year LOWESS smoothing, multi-year droughts are more frequent than pluvials but generally shorter and weaker. Identified multi-year events: Major droughts (≥15 years) occurred in 1625–1644, 1655–1669, 1684–1704, 1737–1751, 1782–1796, 1809–1827, 1832–1849. Major pluvials (≥15 years) occurred in 1560–1578, 1601–1624, 1762–1781. Late Ming megadrought (LMDMD): A megadrought from 1625–1644 lasted ~20 years with cumulative scPDSI severity −6.77, aligning with the Chongzhen drought. Its duration–severity falls within a 100–200-year return period. A strong pluvial (1560–1578) had cumulative severity 10.64 with return period >200 years. Year-type frequencies: Over 464 years, wet and dry years are comparable (76 wet, 77 dry). Extreme dry years were rarer (e.g., 1962, 1632, 1740, 1758) than extreme wet years (e.g., 1776, 1611, 1973, 1910, 1575, 1648, 1620, 1567, 1560, 1983, 1615, 1732, 1736). Notable verification points include 1835 drought (Daoguang 15th year; scPDSI −0.960) and 1877 northern drought (scPDSI −1.325); floods include 1648 Huang–Huai–Hai flood (scPDSI 2.216) and 1849 eastern China flood (scPDSI 1.207). The probability of wet/dry events in 1980–2023 (18.1%) is lower than in 1560–1979 (34.5%). Spectral signals: MTM and red-noise tests (99%) show dominant interannual cycles at ~3.3, 3.7, 4.0, 4.1 years explaining ~55.2% of variance; decadal bands (~9.8 years and below ~36 years) contribute ~34.0%. EEMD corroborates key interannual (~3–4 years) and decadal (~10–30 years) variability. Spatial coherence: Reconstructed scPDSI correlates significantly with observed scPDSI, precipitation, and runoff across the middle Yangtze and southern China during 1924–2020; correlations with temperature are negative but not significant, indicating a water-balance signal. North–south consistency: Comparison with 10 tree-ring hydroclimate series across the East Asian monsoon domain shows significant positive relationships and coherent dry/wet periods (e.g., common droughts in 1587–1591, 1626–1643, 1711–1716, 1925–1932, 1997–2011; common pluvials in 1568–1578, 1602–1606, 1670–1674, 1799–1806, 1868–1873, 1907–1914). PC1 of these records explains 37% variance and tracks this reconstruction well (interannual R=0.50; interdecadal R=0.55, 1856–1988). External/internal drivers: scPDSI correlates with SST patterns—negative in the mid–low latitude West Pacific and positive in the mid–high latitude North Pacific (1870–2023). The LMDMD coincides with a cold PDO phase, El Niño-like ENSO warmth, weakening solar activity entering the Maunder Minimum, and intense Asia-Pacific volcanism. SEA indicates: extreme warm PDO and extreme cold ENSO associate with wet scPDSI in the event year and following two years; extreme cold PDO and extreme warm ENSO associate with drought in the event year and following year; major Asian volcanic eruptions precede significant scPDSI decreases in the fourth year after the event. Global correlations are modest/unstable (RscPDSI–ENSO = −0.02; RscPDSI–PDO = 0.17), and many SEA results are not significant under 1000 Monte Carlo simulations.

The reconstruction demonstrates that the late Ming megadrought was a spatially extensive event affecting both northern and southern monsoon regions, though with regional differences in onset, duration, and magnitude. Cross-validation with independent proxy records (stalagmites, lake sediments, additional tree-ring and streamflow reconstructions) and documentary evidence (dryness/wetness index, locust plagues, grain harvest grades, war frequency) confirms a sustained monsoon weakening with widespread drought circa 1620s–1640s. The moisture deficit likely disrupted agricultural production, reduced yields, and, along with locust outbreaks and epidemics, exacerbated socioeconomic instability, contributing to Ming collapse. Dynamically, interannual (~3–4 years) and decadal (~10–30 years) variability linked to ENSO–PDO, solar variability, and volcanic forcing collectively shaped drought fluctuations; the LMDMD likely emerged from the combined influence of a cold PDO phase, El Niño-like conditions, diminished solar irradiance (Maunder Minimum), and multiple eruptions. Although the strength and stationarity of these relationships vary, the study refines their temporal alignment for the middle Yangtze. The findings underscore the vulnerability of the region’s water balance to sustained monsoon failures and support the need for adaptive water management strategies, particularly given projected warming and potential future monsoon weakening.

A robust regional tree-ring width chronology from scattered Keteleeria davidiana in the middle Yangtze records April–November water balance over the past 464 years. Using an ensemble of linear and nonlinear models, the study reconstructs scPDSI variability, identifies a 20-year late Ming megadrought with north–south coherence, and quantifies its rarity. Comparisons with multiple proxies confirm consistent spatiotemporal patterns across China. Interannual to decadal drought variability is influenced by PDO/ENSO, solar activity, and Asian volcanism, whose combined effects likely produced the LMDMD. These long-term insights are pertinent to the core area of China’s South-to-North Water Diversion Project. In the context of warming-induced aridification and potential monsoon failures, comprehensive, integrated water resource planning is recommended. Future work should continue expanding high-quality tree-ring networks in humid southern China and further constrain dynamical linkages and lags among ocean–atmosphere drivers and regional hydroclimate.

Relationships between reconstructed scPDSI and large-scale drivers (ENSO, PDO) show modest, non-stationary correlations, and many SEA responses are not statistically significant in Monte Carlo tests, reflecting nonlinear, lagged dynamics and proxy sensitivities. Reliable tree-ring coverage extending into the 16th century remains sparse in southern China, and although this study helps fill that gap, uncertainties persist. Machine learning models can overfit; hence, an ensemble was used, but model structural uncertainty remains. Prior non-tree proxies used in comparative context have limitations (e.g., dating accuracy, ambiguous physiological significance).