The model at the heart of the study, called GLOBGM v1.1, is a two-layer, transient groundwater flow model built on a parallelized prototype of the widely used MODFLOW6 code. To make a global simulation at 30-arcsecond resolution computationally feasible, the researchers split the planet into four independent but coupled continental-scale models covering Afro-Eurasia, the Americas, Australia, and the remaining islands. Each of these is further partitioned into non-overlapping sub-models that are solved together within the MODFLOW linear solver. The two-layer design allows the model to represent unconfined, confined, and semi-confined aquifer systems: where a confining layer exists, the upper layer represents that barrier and the lower layer the pressurized aquifer beneath it. Forcing data, including runoff, recharge, and groundwater abstraction, come from the global hydrological model PCR-GLOBWB2, resampled from its native 5-arcminute grid down to the finer GLOBGM grid.

Two technical innovations distinguish this version from its predecessor. The first is a dynamic drainage elevation routine, in which the elevation at which water drains from the groundwater system adjusts according to soil saturation. As soils become wetter, the drainage level rises toward the surface; as they dry, it drops. This coupling, based on the improved ARNO scheme for identifying saturated areas, keeps simulated water levels closer to the surface in wetland regions and suppresses unrealistic head fluctuations and intermittent discharge events. The second is a statistical downscaling of groundwater recharge. Because PCR-GLOBWB2 is known to underestimate recharge, particularly in arid and semi-arid regions where recharge is low and episodic, the team used a Generalized Additive Model trained on observed recharge data to produce a kilometer-scale recharge field, then applied a multiplicative correction factor to the model’s monthly recharge inputs, capped at local precipitation to preserve numerical stability.

Calibration was a major undertaking. The team tested 162 permutations of prefactors applied to hydraulic conductivity, anisotropy, and entrance resistance, varying the conductivity adjustments by lithological class drawn from the global GLiM dataset. Candidate settings were evaluated against long-term average water table depths from 34,800 observation wells in IGRAC’s Global Groundwater Monitoring Network, deliberately excluding known groundwater abstraction hotspots so that depletion would not be confused with model error. The best-performing configuration reduced the mean bias in water table depth from minus 4.8 meters in the uncalibrated predecessor to 3.6 meters, with the depth-weighted bias falling from 34.2 to 32.5 meters. The improvements were most pronounced for shallow water tables, which matter most for human use and for sustaining ecosystems through baseflow.

Validation against independent observations showed that the model produces skillful predictions, measured by a non-parametric Kling-Gupta Efficiency skill score, in approximately 83 percent of shallow to intermediate wells with water tables between zero and 60 meters deep when compared at the monthly scale, and in about 60 percent of deep wells exceeding 60 meters. At the annual scale, 71 percent of shallow wells and 66 percent of deep wells were skillful. The depth dependence is physically sensible: deeper aquifers respond slowly to recharge through thick unsaturated zones that the model does not explicitly resolve. To check performance at scales where point observations are absent, the team compared simulated total water storage anomalies with satellite gravimetry from the GRACE and GRACE-FO missions, finding agreement in trend direction over 62.3 percent of the global land surface and agreement in both direction and magnitude over 45.7 percent, consistent with other state-of-the-art global hydrological models.

The researchers also added a machine learning bias correction. A two-stage LightGBM ensemble, one classifier predicting whether the model overestimates or underestimates groundwater heads and one regressor predicting the magnitude of the error, was trained on nearly 97,000 observation wells using model parameters, Google’s AlphaEarth geospatial embeddings, and absolute latitude as a proxy for aridity. Training and testing were separated spatially using K-means clustering to prevent autocorrelation from inflating performance. The correction correctly predicted the direction of bias in 80 percent of holdout cases and shifted median bias toward zero across all depth classes, with the largest gains in deep wells. An asymmetric ramp-factor scheme prevents the correction from pushing simulated water tables unrealistically close to the land surface, preserving the model’s physics where it matters most.

The historical simulation from 1960 to 2019, driven by the ISIMIP3a observational forcing, reproduced the world’s best-documented groundwater depletion zones with striking fidelity: the southern and central High Plains of the United States, Iran, the Arabian Peninsula, and the Indo-Gangetic Plain all show the deepening water tables that independent observations have recorded. The model also identified rising water tables in regions known to be gaining storage, including the Guarani aquifer in South America, the northern Great Plains of North America, and the Karoo Basin in South Africa. Intriguingly, it flagged rising water tables across northern latitudes and the Arctic, potentially linked to climate-driven increases in recharge, though the authors caution that these areas overlap with permafrost zones where the model is known to be less reliable.

The future projections, run with five global climate models under three combined socioeconomic and climate scenarios from ISIMIP3b, paint a nuanced picture. Globally, more regions are projected to see rising than falling water tables through 2100, consistent with an intensifying hydrological cycle, and the pace of change increases from the low-emission SSP1-RCP2.6 scenario to the high-emission SSP5-RCP8.5. Water tables are projected to rise across much of the Sahel, the Horn of Africa, North America’s Pacific Coast, the Tibetan Plateau, and the Arabian Peninsula, while declines are projected for the Mediterranean, southeastern Australia, and the African and South American tropics. Europe stands out as the sole continent where water tables are expected to deepen markedly, particularly under high emissions. Known depletion hotspots are expected to persist, and the model projects continued declines over southern Europe, the Amazon, California’s Central Valley, and the High Plains, diverging from earlier studies that omitted groundwater abstraction, a difference the authors attribute to the inclusion of human water demand.

The team is transparent about the model’s limits. Simulations are flagged as less reliable in karst aquifers, where Darcian flow assumptions break down; in permafrost regions, where frozen ground complicates recharge; and in steep mountainous terrain, where the schematization is too coarse to resolve hillslope hydrology. Quality assurance maps accompany the dataset so users can interpret results appropriately. All simulations, along with the model code, are openly accessible, and because the outputs follow the ISIMIP framework, they can be combined with other Earth system datasets to study how changing groundwater affects ecosystems, food production, and society. In a world where the United Nations has warned of an era of global water bankruptcy, a kilometer-scale, globally consistent view of where water tables are falling and rising may prove one of the most consequential datasets yet produced for planning a sustainable water future.

Subject of Research: Global hyper-resolution modeling of historical and future groundwater dynamics at approximately 1 km resolution

Article Title: Global hyper-resolution modeling of historical and future groundwater dynamics

Article References: van Jaarsveld, B., Wanders, N., Otoo, N. G., Sutanudjaja, E. H., Verkaik, J., Zamrsky, D., & Bierkens, M. F. P. (2026). Global hyper-resolution modeling of historical and future groundwater dynamics. Earth System Dynamics, 17(5), 1201-1236. https://doi.org/10.5194/esd-17-1201-2026

Image Credits: AI Generated

DOI: 10.5194/esd-17-1201-2026

Keywords: groundwater, hyper-resolution modeling, GLOBGM, water table depth, climate change, groundwater depletion, MODFLOW, ISIMIP, machine learning bias correction, GRACE satellite, SSP-RCP scenarios, water resources management

At the heart of the study is a measurement called aerosol optical depth, or AOD, which quantifies how much sunlight is removed from a beam passing through the atmosphere by particles suspended in the air. An AOD of 0.1 describes relatively clean skies, while values above 1.0 indicate conditions so hazy that the sun itself can appear dimmed and diffuse. The researchers worked with daily retrievals from the MODIS instrument aboard NASA’s Terra satellite, which has been scanning the Earth since early 2000. Raw satellite AOD products, however, carry well-documented biases, particularly over the bright surfaces and complex aerosol mixtures of West Africa. To correct these errors, the team built a machine learning framework known as Gradient Boosting Quantile Regression, trained against ground truth from fourteen AERONET sun photometer stations spread across nine West African countries.

The correction proved dramatic. At the Nigeria-Ilorin AERONET site, the machine learning adjustment shrank the raw mean bias of MODIS retrievals from minus 0.252 to just minus 0.015, a near-perfect agreement on average. More importantly, the fraction of satellite observations falling within the expected error envelope of ground measurements jumped from a dismal 23.0 percent to 95 percent. When the team tested the model’s ability to generalize to locations it had never seen, using a leave-one-site-out validation strategy, the expected error agreement remained a respectable 74.5 percent. This bias-corrected, boundary-clipped daily record gives Nigeria something it has never had before: a trustworthy, continuous, quarter-century chronicle of its atmospheric particle loading, suitable for everything from climate model evaluation to public health epidemiology.

So what does 25 years of data actually show? Perhaps surprisingly, the national annual trend is essentially flat. The linear slope across the full record is minus 0.0020 per year, with a p-value of 0.226, meaning the change cannot be distinguished from random year-to-year variability. But seasonal decomposition told a subtler story. The dry season, when Saharan dust and harmattan winds dominate, shows a suggestive decline of minus 0.0043 per year, equivalent to roughly a 16 percent reduction over the study period. This trend sits just at the edge of conventional statistical significance, with a p-value of 0.059 after accounting for autocorrelation through lag-1 pre-whitening, so the authors are careful to describe it as indicative rather than definitive. The wet season shows only a marginal, non-significant decrease of minus 0.0012 per year. In other words, Nigeria’s aerosol environment has been remarkably stable, with at most a gentle cleaning of its dustiest months.

The seasonal rhythm itself, however, is unmistakable. The monthly climatology is predominantly unimodal, peaking in February at a mean AOD of 0.865, when dry northeasterly harmattan winds carry desert dust deep into the country, and bottoming out in November at 0.370, as the monsoon transition washes the atmosphere clean. That February peak is more than double the November minimum, a swing that shapes everything from solar power output to aviation visibility to respiratory disease patterns across the region. Previous research has linked these dust episodes to meningitis outbreaks in the Sahel, and the harmattan season has been implicated in altered transmission dynamics of respiratory infections, making the timing and intensity of this annual maximum a matter of direct public health consequence.

To bring order to the daily chaos of aerosol variability, the researchers applied an unsupervised K-means clustering algorithm to the full record, letting the data sort itself into distinct atmospheric states. Three regimes emerged. Clean days, with a mean AOD of 0.399, account for 57.7 percent of the record. Transitional days, averaging 0.647, make up 31.6 percent. And then there are the Extreme days, averaging a staggering AOD of 1.163, which occupy 10.7 percent of all days in the record. These extreme events are overwhelmingly a dry season phenomenon: they occur on 18.1 percent of dry season days but only 3.2 percent of wet season days. That asymmetry confirms quantitatively what residents of Kano or Maiduguri know viscerally: the harmattan is when the air becomes a hazard.

One of the most practically useful contributions of the study is its characterization of how often extreme aerosol events recur. Using empirical return period analysis, the team calculated that days with AOD exceeding 1.0 arrive on average every 12.5 days, based on 711 exceedances across the record, with a 95 percent confidence interval of 11.6 to 13.4 days. Thresholds above 1.5 are crossed roughly every 61 days, with 145 exceedances recorded. And truly exceptional events, with AOD above 2.0, occur about once every 6.06 years, though with only four such events in 25 years the confidence interval is wide, spanning 3.03 to 24.25 years. These recurrence benchmarks give atmospheric hazard planners, health officials, and solar energy operators a climatological yardstick for what counts as a normal bad day versus a once-in-a-decade crisis.

The spatial structure of Nigeria’s aerosol burden is equally striking. A global spatial autocorrelation analysis yielded a Moran’s I of 0.949, with a p-value of 0.001 and a z-score of 163.9, confirming that aerosol loading is not randomly distributed but intensely clustered in space. Local Indicators of Spatial Association, corrected for the false discovery rate, flagged 40.1 percent of grid cells as statistically significant hotspots or coldspots. The pattern reveals a stable High-High hotspot belt across Southern and Southwestern Nigeria, where dense populations, urban pollution, and biomass burning combine, and persistent Low-Low spots in the Northeast. This geographic fingerprint has remained stable across the record, suggesting that the underlying drivers, from dust transport corridors to emission sources, are entrenched features of the regional environment rather than transient anomalies.

The study also probed the atmosphere’s memory, asking whether today’s aerosol conditions predict tomorrow’s. The answer is a strong yes, at least over short horizons. The lag-1 daily autocorrelation is 0.75, decaying to 0.21 by lag 30, indicating that aerosol states persist for days to weeks before the signal fades. An ARIMA (2,0,2) time series model successfully captured these synoptic dynamics, achieving a pseudo R-squared of 0.58. This short-range predictability has real-world value: it means that once a major dust outbreak begins, forecasters can anticipate with reasonable confidence that hazy conditions will persist for several days, giving hospitals, airports, and power grid operators time to prepare.

Taken together, these findings establish a rigorous long-term atmospheric baseline for a country whose air quality has historically been monitored far less intensively than its population warrants. The stability of the national trend, punctuated by a possible dry season improvement, offers a measure of reassurance, but the sheer frequency of extreme events, one in every nine days nationally and nearly one in five during the dry season, underscores the scale of the ongoing exposure. As West Africa’s combustion emissions continue to grow and climate change reshapes dust mobilization across the Sahara, records like this one will serve as the indispensable reference point against which future change is measured. For Nigeria’s policymakers, the message is clear: the dustiest skies are predictable, clustered, and recurring on a known clock, and that predictability is itself a tool for protecting public health.

Subject of Research: Long-term satellite analysis of aerosol optical depth dynamics, seasonal regimes, and extreme dust events over Nigeria from 2000 to 2024

Article Title: A 25-year analysis of aerosol optical depth dynamics over Nigeria (2000–2024): spatial-temporal dynamics, seasonal regimes, and extreme event characterization

Article References: Ekpe, E. E., Shi, C., Yin, S., Wang, H., Zhao, Y., Letu, H., & Raihan, M. Z. U. (2026). A 25-year analysis of aerosol optical depth dynamics over Nigeria (2000–2024): spatial-temporal dynamics, seasonal regimes, and extreme event characterization. Air Quality, Atmosphere & Health, 19(10), Article 224. https://doi.org/10.1007/s11869-026-02111-4

Image Credits: AI Generated

DOI: 10.1007/s11869-026-02111-4

Keywords: aerosol optical depth, Nigeria, MODIS Terra, machine learning bias correction, AERONET, Saharan dust, harmattan, air quality, spatial autocorrelation, extreme events, seasonal climatology, West Africa