Climate change linked to drought in Southern Madagascar

The study investigates whether the recent multi-year drought in Southern Madagascar reflects anthropogenic climate change, focusing on seasonal hydroclimatic shifts rather than only annual totals. Prior assessments (e.g., by World Weather Attribution) emphasized two-year average precipitation and found limited climate-change influence but relied on datasets that are not optimized for multi-decadal trend detection in Southern Madagascar due to sparse in-situ gauges and changing observational coverage. Given the region’s strongly seasonal rainfall and agriculture tightly aligned to rainy-season onset (October–February), delays in onset can increase crop water stress, delay maturation, and extend the lean season. Observational analyses of regional stations and related Southern African records suggest a trend toward later rainy-season onset, potentially linked to Hadley cell expansion and poleward storm-track shifts. The purpose of this study is to define a seasonal hydrologic fingerprint of anthropogenic change using climate models, test for its presence in multiple remotely sensed indicators, and quantify how these trends altered the likelihood of the 2017–2022 drought, thereby informing agricultural risk and adaptation needs.

Previous work on the 2019–2021 drought by the World Weather Attribution group concluded a limited role for climate change when considering two-year running average precipitation using ERA5 and CHIRPS datasets. However, the scarcity of rain gauges in Southern Madagascar (only three coastal, incomplete stations feeding CHIRPS) and evolving observation systems complicate trend detection. Independent station analyses (Toliary, Taolagnaro; 1950–2018) show delayed rainy-season onset, consistent with signals observed in South Africa where rainfall seasonality shifts have been linked to Hadley cell expansion and poleward shifts of storm tracks. Broader literature documents Southern Hemisphere tropical widening and associated circulation changes, and prior CMIP5 and downscaled studies project delayed wet seasons and reduced early-season precipitation across Southern Africa, including Madagascar. Additional literature highlights potential roles of land-use change (deforestation), ozone depletion, and model representation of synoptic systems (e.g., tropical temperate troughs, Mozambique Channel Trough) as factors influencing hydroclimate and its simulation.

The study integrates observations, proxy reconstructions, and climate model simulations. Observations: Monthly NDVI from GIMMS (1981–2015; 1/12°) and MODIS/Terra MOD13C2 (2000–present; 0.05°) were combined by bias-adjusting MODIS to GIMMS over their overlap, using GIMMS through 2015 and MODIS from 2016–2022 to create a 1982–2022 NDVI record. Soil moisture is from ESA CCI v08.1 (0.25°, 1979–2022), with analyses restricted to 2003–2022 to minimize sensor-homogenization issues after AMSR-E integration. Precipitation is from CHIRPS. To mitigate land-use influences, analyses focus on areas with minimal deforestation (forest fraction decrease <0.15 between 1973 and 2017), based on 30 m forest cover maps merged from national historical maps and global tree cover loss products. Climate models: CMIP6 near-surface soil moisture (mrsos) from 12 runs (historical to 2015; SSP5-8.5 to 2100) and pre-industrial control (piControl) were regridded to 0.5°. Of 12 models, 3 were excluded due to unrealistic pre-industrial seasonality, leaving 9 runs for attribution. Trends: Linear least-squares trends were computed for monthly, seasonal (SON, DJF, MAM, JJA), and annual (Sep–Aug) averages. Observational and proxy trend uncertainties used bootstrap resampling (1000 samples; 90% CIs). Model trend variability used sampling of non-overlapping trend windows (e.g., 118-year windows from piControl; 41-year from forced), every 5 years, to characterize internal versus forced variability. Proxy soil moisture: Soil moisture was reconstructed from NDVI by fitting a non-negative linear regression with a one-month NDVI lag at 0.25° resolution, using sub-grid 0.05° NDVI pixels as predictors and a single intercept per grid cell. The optimal lag (generally one month) was determined by maximizing correlation over 2003–2022. Performance was validated by rolling 4-year block cross-validation; mean squared error increased slightly from training (0.0011 cm³ cm⁻³) to testing (0.0012 cm³ cm⁻³). Fingerprint attribution: A multivariate fingerprint approach modeled seasonal structures. Four-dimensional multivariate normal (MVN) distributions of seasonal trends (SON, DJF, MAM, JJA) were fit separately to CMIP6 forced and piControl trend samples. The likelihood of observed proxy seasonal trend vectors (1982–2022) under forced versus control MVNs yielded likelihood ratios with bootstrap-derived 90% CIs. Drought attribution: For anomalies, MVNs were fit to seasonal soil moisture anomalies in a baseline (1970–1999) and a recent 30-year window centered on the drought (2005–2034) using CMIP6 simulations; observed drought-season anomalies were evaluated to obtain likelihood ratios per year and combined, with bootstrap uncertainty. A second approach updated MVN means using extrapolated observed proxy-soil-moisture trends (1982–2017) while retaining simulated covariances. Anomalies were used rather than absolute values to account for inter-model biases in soil moisture magnitude.

• Forced CMIP6 simulations exhibit a robust seasonal fingerprint of drying with consistent negative soil moisture trends during rainy-season onset (Sep–Nov). The across-model average trend magnitude during onset is approximately −2.3% soil moisture per decade (range −4.0% to −1.1%). Pre-industrial control simulations show no significant monthly soil moisture trends, implicating anthropogenic forcing.
• Observations show significant drying at rainy-season onset: precipitation declines in SON by −2.4 mm month−1 decade−1 (90% CI −4.5 to −0.6) over 1981–2022, with corresponding declines in NDVI and soil moisture during their periods of record. Since the early 2000s, NDVI and soil moisture significantly decline across all seasons.
• Strong interannual covariability: detrended annual precipitation correlates with annual-average NDVI at 0.74 (MODIS) and 0.82 (GIMMS) for 2003–2015; soil moisture and NDVI correlations are 0.79 (MODIS) and 0.85 (GIMMS). During SON, soil moisture–NDVI correlations are 0.70 (GIMMS) and 0.76 (MODIS), while precipitation–NDVI correlations are 0.29 (both NDVI products).
• Attribution of trends: The seasonal structure of proxy soil moisture trends (1982–2022) is about 100 times more likely (90% CI 5–2683) under anthropogenic forcing than under pre-industrial variability.
• Attribution of recent drought: Using model-based MVNs, the 2017–2022 conditions are 15 times more likely (90% CI 8–27) due to anthropogenic forcing; four of five years individually show significantly higher likelihoods. Likelihoods further increase when means are updated with observed trends.
• Future likelihoods: Under SSP5-8.5, droughts like 2017–2022 become increasingly likely; by 2070–2099, the recent drought conditions are >1000 times more likely than relative to 1970–1999.
• Mechanism: Forced models simulate Hadley cell expansion and a poleward shift/intensification of subtropical descent and jet positioning, delaying storm-track steering over Southern Madagascar and drying the early rainy season.

The findings indicate that anthropogenic forcing has altered the seasonal hydroclimate of Southern Madagascar by delaying rainy-season onset and reducing early-season water availability, thereby increasing the likelihood and severity of recent drought conditions. This addresses the research question by showing agreement between observed seasonal trend structures and those predicted by forced climate simulations, while pre-industrial controls do not exhibit such changes. The mechanistic link is consistent with Hadley cell expansion and poleward shifts of subtropical jets and descending branches that extend the duration of dry-season dynamics. The study reconciles prior limited-attribution conclusions that used annual metrics by demonstrating that seasonal fingerprints can be masked in annual averages due to compensating changes within the wet season. The implications are substantial for agriculture, as delayed onset increases crop water stress, shifts cropping calendars, and lengthens the lean season. Despite potential confounders (e.g., land-use change, ozone depletion, model biases in synoptic systems), multiple independent indicators and consistency across CMIP generations and downscaled studies strengthen the attribution to anthropogenic climate change. The results underscore the need for adaptation strategies in water management, agronomy, and public health to mitigate growing risks.

Multiple independent observational indicators (precipitation, soil moisture, NDVI) reveal long-term drying at the onset of the rainy season in Southern Madagascar, consistent with a robust seasonal fingerprint simulated by forced CMIP6 models and absent in pre-industrial control runs. Anthropogenic forcing has made the 2017–2022 drought far more likely, and similar events are projected to become increasingly probable under continued emissions. The study contributes a multivariate seasonal fingerprint attribution framework demonstrating that seasonal structure is critical for detection and attribution in strongly seasonal climates. Future work should improve observational homogeneity, extend high-quality soil moisture records, quantify the roles of local and remote land-use change and ozone recovery, and refine the representation of regional synoptic systems (e.g., tropical temperate troughs, Mozambique Channel Trough) in models. Adaptation research should evaluate effective water management, climate-resilient cropping strategies, and targeted interventions for vulnerable communities.

Key limitations include: sparse in-situ precipitation gauge coverage in Southern Madagascar and reliance on gridded/reanalysis products for precipitation; potential temporal inhomogeneities in merged satellite soil moisture records (sensor changes and structural breaks) that motivate restricting analysis to 2003–2022; NDVI-based proxy soil moisture may be affected by land-use change (deforestation, seasonal burning) and vegetation disturbances, with limited availability of accurate annual deforestation maps; possible influences of both local and remote deforestation on atmospheric moisture recycling and onset timing; uncertainties in model representations of synoptic features (tropical temperate troughs, Mozambique Channel Trough) and orographic precipitation; and overlapping influences of greenhouse gases and ozone depletion on Hadley cell expansion. Despite these, consistent cross-indicator and cross-model signals support the main conclusions.