Data-driven model maps Kabul’s shrinking groundwater and redraws its well network

Beneath Kabul, one of the world’s fastest-growing capital cities, the shallow aquifer that supplies most of the city’s water is quietly running down. Groundwater levels across the Kabul Plain have been falling at an average rate of 0.8 meters per year, and in some areas the water table has dropped by as much as 60 meters over two decades. Roughly a third of the city’s wells no longer function at all. Against this backdrop, a research team led by Javad Hussainzadeh and Asadullah Farahmand has published one of the first systematic maps of where groundwater is most likely to be found in the Kabul aquifer, together with a proposal for reorganizing the city’s entire well network. The study, published in Environmental Earth Sciences on 24 October 2026, combines geographic information systems with two competing statistical models and offers a rare, quantified decision-support tool for a city whose water future is increasingly precarious.

The stakes could hardly be higher. Kabul is home to an estimated 5.3 million people, and its urban footprint expanded by 40 percent between 2000 and 2020. Groundwater is the primary source for domestic, agricultural, and industrial use across the Kabul Plain, yet abstraction has outpaced recharge for years. Monitoring data show that between September 2016 and September 2019 alone, the aquifer’s water level fell by roughly seven meters, from 1791.2 to 1784.3 meters, with the steepest drop occurring in 2018. Recent water-level declines have exceeded 1.5 to 2 meters per year. The authors argue that any meaningful reform of water abstraction limits must rest on rigorous, defensible assessments of where the remaining resource actually lies, which is precisely what their mapping exercise was designed to provide.

The study area covers approximately 329 square kilometers of the Kabul Plain, wedged between metamorphic mountain ranges and split by a central ridge into eastern and western sectors. The plain is underlain by Quaternary and Tertiary alluvial deposits: Quaternary sediments less than 80 meters thick in the valleys, resting on Tertiary deposits estimated to reach up to 800 meters beneath the city. Three hydrogeological units have been identified previously, including a shallow aquifer less than 100 meters thick composed of Quaternary deposits, an intermediate confining layer of upper Neogene deposits, and a deep aquifer extending to about 388 meters above bedrock. The research focused on the shallow aquifer, whose annual water balance is severely negative. The aquifer material consists of unconsolidated to semi-consolidated lacustrine, fluvial, and aeolian sediments, including gravel, sand, and silt with interbedded clay lenses, and its hydraulic properties vary dramatically across the plain.

That variability is central to the study’s findings. Well logs and grain-size analyses show hydraulic conductivity reaching 109 to 382 meters per day in the eastern part of the plain, where coarse gravel and sand dominate, but falling to just 0.14 to 73 meters per day in the west, where finer sediments prevail. The depth to groundwater also varies widely, with the unsaturated zone reaching 24 to 85 meters thick in the north and west, while the water table lies only 2.6 to 7.1 meters below ground in the south and southwest. To capture this complexity, the team assembled nine hydrogeological predictor layers within a GIS framework: aquifer media, aquifer thickness, unsaturated zone thickness, groundwater drawdown, hydraulic gradient, vadose zone type, aquifer recharge areas, geology, and distance from drainages. Data came from national monitoring programs, government ministries, and satellite remote sensing, including water-level records from 90 observation wells spanning 2005 to 2019 and a 12.5-meter-resolution digital elevation model.

The researchers then pitted two fundamentally different modeling philosophies against each other. The first, a Fuzzy Analytical Hierarchical Process, or FAHP, is a knowledge-based method in which experts assign relative weights to each factor through pairwise comparisons, with fuzzy logic used to soften the subjectivity of human judgment. The second, an Evidential Belief Function model grounded in Dempster-Shafer theory, is data-driven: it calculates class-specific weights directly from training data without any prior weighting, and explicitly tracks uncertainty alongside belief. To calibrate and test both models, the team identified 149 wells with hydraulic conductivity exceeding 59 meters per day, the median value among 385 surveyed wells, as positive indicators of high groundwater potential. These were split randomly into 104 training wells and 45 independent validation wells, and all spatial analyses were run at a 12.5 by 12.5 meter pixel resolution.

The results delivered a clear verdict. The EBF model achieved a validation area under the receiver operating characteristic curve, or AUC, of 62.5 percent, down only modestly from its training value of 69.1 percent, placing it in the moderate-to-good performance category. The FAHP model, by contrast, managed a validation AUC of just 51.1 percent, essentially indistinguishable from random guessing, and its training performance of 50.2 percent offered no comfort either. According to the authors, the near-random performance of the expert-weighted approach indicates that subjective weighting schemes are inadequate for capturing the complex, non-linear interactions among hydrogeological factors in this heterogeneous aquifer. The EBF model’s comparatively modest score, lower than figures reported in less complex settings such as Malaysia and India, reflects Kabul’s challenging hydrogeology, sparse monitoring coverage, and multiple aquifer units with variable connectivity.

A leave-one-out sensitivity analysis sharpened the picture further. When the geology layer was removed from the EBF model, validation AUC dropped from 62.5 percent to 59.4 percent, the largest decline of any predictor, confirming that lithological heterogeneity is the dominant control on groundwater potential in the Kabul aquifer. Notably, the EBF analysis assigned its highest belief value, 0.69, to Neogene fine-grained deposits, while Quaternary alluvium scored 0.31 and Neogene conglomerate scored zero. Other layers, including recharge areas and unsaturated thickness, produced AUC changes of less than 2 percent when removed, suggesting they play secondary or redundant roles. The EBF model’s validation AUC remained stable between 59.4 and 64.5 percent across all removals, demonstrating robustness to input uncertainty. The FAHP model, in contrast, fluctuated erratically between 48.7 and 55.4 percent with no dominant factor, reinforcing the conclusion that data-driven probabilistic methods outperform knowledge-based weighting in data-scarce, hydrogeologically complex urban aquifers.

The two models also disagreed about where the water is. The FAHP map placed 33.8 percent of the area in the very high potential category and 25.9 percent in high potential, concentrated mainly in the southern parts of the plain. The EBF map told a different story: 27.4 percent of the aquifer fell in the moderate potential class, mostly in the central parts, while the very high and high potential zones covered 5.2 and 20.6 percent respectively, concentrated in the eastern and central aquifer. Nearly 47 percent of the EBF map was classified as very poor or poor potential, mostly in the southern and northern parts. The EBF spatial patterns were consistent with its belief calculations, which found that proximity to drainage networks, coarser aquifer media, greater aquifer thickness, and lower hydraulic gradients all correlated with higher groundwater probability.

The study’s most consequential output is a proposed redesign of the city’s well network. Using the EBF-derived potential zones together with interpolated hydraulic conductivity, the team defined grid scenarios matched to productivity classes: a 1000-meter grid for high-conductivity zones above 50 meters per day as the primary scenario, and 400-meter and 300-meter grids for moderate and low conductivity zones as secondary contingencies. Under the primary scenario, 122 of the existing 385 wells are proposed for removal, 263 are retained, and 45 new wells are added in high and very high potential zones, yielding a reconfigured network of 308 wells. A secondary expansion scenario for extreme urbanization pressures would add 264 supplementary wells, 218 on the 400-meter grid and 46 on the 300-meter grid, raising the total network to 572 wells, but only if future demand substantially exceeds current projections.

The authors are careful to frame these recommendations as preliminary screening rather than a final blueprint. The binary target variable simplifies what is in reality a continuous productivity gradient, the datasets span 2008 to 2019 with some temporal inconsistency, and the radius of influence was treated as a fixed parameter that ignores transient pumping effects and well interference. The 385 wells are also unevenly distributed, which may introduce spatial bias. Before any operational changes, the team calls for detailed pumping tests, numerical groundwater flow modeling, and socio-economic analysis. Still, the framework offers something Kabul has lacked: a transparent, independently validated map of where new drilling is most likely to succeed, where existing wells should be phased out, and how the network could scale adaptively as the city grows. The approach, the authors note, is transferable to other overexploited urban aquifers, provided the predictor layers and thresholds are recalibrated to local conditions.

Subject of Research: Groundwater potential mapping and well network optimization in the Kabul aquifer using fuzzy analytical hierarchy process and evidential belief function models

Article Title: Groundwater potential zone assessment and well network optimization under hydrogeological complexity: application of fuzzy-analytical hierarchical process and evidential belief function in the Kabul Aquifer, Afghanistan

Article References: Hussainzadeh, J., Mohammaddost, A., Brookes, J., & Farahmand, A. (2026). Groundwater potential zone assessment and well network optimization under hydrogeological complexity: application of fuzzy-analytical hierarchical process and evidential belief function in the Kabul Aquifer, Afghanistan. Environmental Earth Sciences, 85(17), Article 439. https://doi.org/10.1007/s12665-026-13175-x

Image Credits: AI Generated

DOI: 10.1007/s12665-026-13175-x

Keywords: Kabul aquifer, groundwater potential zones, evidential belief function, fuzzy analytical hierarchy process, GIS mapping, hydraulic conductivity, well network optimization, Afghanistan, water scarcity, Dempster-Shafer theory, ROC-AUC validation, urban hydrogeology