GroundHog: Revolutionizing GLDAS Groundwater Storage Downscaling for Enhanced Recharge Estimation in Bangladesh

Saleh Sakib Ahmed, Rashed Uz Zzaman, Saifur Rahman Jony, Faizur Rahman Himel, Afroza Sharmin, A.H.M. Khalequr Rahman, M. Sohel Rahman, Sara Nowreen, ∗ Corresponding authors. Emails: [email protected], [email protected] · 2025

Read GroundHog: Revolutionizing GLDAS Groundwater Storage Downscaling for Enhanced Recharge Estimation in Bangladesh by Saleh Sakib Ahmed, Rashed Uz Zzaman, Saifur Rahman Jony, Faizur Rahman Himel, Afroza Sharmin, A.H.M. Khalequr Rahman, M. Sohel Rahman, Sara Nowreen, ∗ Corresponding authors. Emails: [email protected], [email protected] (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-22771.txt.

Introduction

Eight authors led by Saleh Sakib Ahmed sharpen coarse satellite ground water maps of Bangladesh down to street scale. Well reads are sparse, yet water plans need fine maps that span years. Their first model fills gaps in well peaks and dips with strong fit scores. Those fills plus true well reads then ground a second model that upsamples satellite grids.

The input grid cells span twenty-five kilometres; the output cells span two. Fit on the fine maps hit 0.96 in the check runs of the work. Maps for 2003 to 2024 fall out of the stack in one pass. Recharge sums and drift tales for plan teams close the work. Any lone point can be read with no state well list.

Catalog record

CREATORSaleh Sakib Ahmed, Rashed Uz Zzaman, Saifur Rahman Jony, Faizur Rahman Himel, Afroza Sharmin, A.H.M. Khalequr Rahman, M. Sohel Rahman, Sara Nowreen, ∗ Corresponding authors. Emails: [email protected], [email protected]
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE48 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-22771

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

GroundHog: Revolutionizing GLDAS Groundwater Storage Downscaling for Enhanced Recharge Estimation in Bangladesh Saleh Sakib Ahmed, Rashed Uz Zzaman, Saifur Rahman Jony, Faizur Rahman Himel, Afroza Sharmin, A.H.M. Khalequr Rahman, M. Sohel Rahman, Sara Nowreen, ∗ Corresponding authors. Emails: [email protected], [email protected] Abstract Long-term groundwater level (GWL) measurement is vital for effective policymaking and recharge estimation using annual maxima and minima. However, current methods prioritize short-term predictions and lack multi-year applicability, limiting their utility. Moreover, sparse in-situ measurements lead to reliance on low-resolution satellite data like GLDAS as the ground truth for Machine Learning models, further constraining accuracy. To overcome these challenges, we first develop an ML model to mitigate data gaps, achieving R 𝟐 R^{2} scores of 0.855 and 0.963 for maximum and minimum GWL predictions, respectively. Subsequently, using these predictions and well observations as ground truth, we train an Upsampling Model that uses low-resolution (25 km) GLDAS data as input to produce high-resolution (2 km) GWLs, achieving an excellent R 𝟐 R^{2} score of 0.96. Our approach successfully upscales GLDAS data for 2003-2024, allowing high-resolution recharge estimations and revealing critical trends for proactive resource management. Our method allows upsampling of groundwater storage (GWS) from GLDAS to high-resolution GWLs for any points independently of officially curated piezometer data, making it a valuable tool for decision-making. 1 Introduction Groundwater from shallow aquifers is a crucial freshwater resource, supporting domestic, agricultural, and industrial needs globally. Nearly 50% of megacities and 80% of irrigation depend on groundwater [ 1 ] , which comprises 99% of Earth’s unfrozen freshwater [ 2 ] . Excessive extraction has caused significant depletion, threatening sustainability [ 3 ] . In Bangladesh, groundwater is the main source of drinking and irrigation. It is extracted using shallow pumping/tube wells (less than 100 m deep), whose numbers have increased from 0.1 million in 1981 to more than 1.5 million in 2013 [ 4 ] , mainly due to drying out of seasonal rivers [ 5 ] . These shallow unconfined aquifers within 100 m depth from ground surface are annually replenished through a process called recharge , where infiltrated water flows underground to maintain groundwater levels [ 6 , 7 ] . Groundwater Level (GWL) , measured in meters below ground level (BGL), refers to the depth of the water table from the land surface at specific monitoring points. The Water Table Fluctuation (WTF) method uses these annual minimum and maximum groundwater levels to estimate recharge [ 8 ] . Therefore, accurate monitoring of annual maximum and minimum groundwater level fluctuations is vital for effective groundwater management, particularly in regions facing growing demands from population growth and agricultural intensification [ 9 , 10 ] . However, in-situ measurements, though accurate, are sparse and costly, especially in Bangladesh. Relevantly, satellite dataset, such as, The Global Land Data Assimilation System (GLDAS) [ 11 ] provides global groundwater storage (GWS) estimates for each 25 km × 25 km grid cell, expressed in millimeters of water equivalent. In this context, ‘millimeters’ indicate the depth of water that would uniformly cover the cell—multiplying this depth by the cell’s area yields the total volume of water stored. GWS serves as an indirect indicator for groundwater depth: increased storage correlates with shallower water tables (lower BGL), while decreased storage implies deeper water tables (higher BGL). However, the coarse 25 km resolution of GWS data fails to capture the localized variations of Bangladesh’s highly heterogeneous landscape, where changes may occur even every 2 km [ 12 ] . Furthermore, understanding recharge rates requires precise water level measurements, which GWS alone cannot provide. This calls for models capable of downscaling satellite water storage data to finer resolutions of water levels, particularly at 2 km (or even better). To overcome the scarcity of high-quality in-situ ground truth data for water level, interpolation methods, such as, Ordinary Kriging [ 13 ] and Inverse Weighted Distance (IDW) [ 14 ] are often used. Ordinary Kriging, in particular, is known to perform poorly with sparse datasets [ 15 ] . Verma et al. [ 16 ] used empirical Bayesian kriging (EBK) interpolation [ 17 ] to downscale the Gravity Recovery and Climate Experiment (GRACE) and in-situ data of the Indian State of Maharashtra to a resolution of 13 km. Then they used artificial neural network (ANN) to establish a relationship between these interpolated global context GRACE data and interpolated local context groundwater data. However, these methods struggle in heterogeneous environments influenced by factors, such as, soil type and land use [ 18 ] . EBK’s reliance on Gaussian distribution assumptions can lead to inaccuracies in predictions for non-Gaussian data, limiting its applicability in diverse hydro-geological settings [ 19 ] . Due to these shortcomings, interpolation methods often fail to perform well. Recently, machine learning approaches integrating hydro-geological factors have gained traction for groundwater prediction, albeit not without some challenges. Tao et al. [ 20 ] reviewed models tailored for short-term predictions (hourly, weekly, and monthly) and noted that very few studies focused on yearly predictions, which are crucial for the estimation of annual recharge. Similarly, Khan et al. [ 21 ] highlighted the lack of effective multi-year annual groundwater level (GWL) prediction models, a critical gap for the estimation of recharge trends and policymaking. As has been mentioned already, the sparsity of GWL data is a significant issue and interpolation methods cannot capture the inherent intricacies of the groundwater dynamics. Previous studies have attempted to downscale GLDAS data using low-resolution groundwater storage (GWS) data as the ground truth and in-situ groundwater levels for validation. However, these efforts have faced significant limitations in both accuracy, resolution and applicability, often showing negative correlations at many validation points [ 22 ] . Additionally, existing models typically rely on year-specific designs, which further reduces their generalizability. For example, Zhang et al. [ 22 ] developed separate models for each year (2004–2016) to downscale GLDAS and GRACE data across China. They used Random Forest [ 23 ] and XGBoost [ 24 ] , with low-resolution GLDAS data serving as the ground truth, to calculate model residuals. These residuals were then downscaled to a 1 km resolution using cubic spline interpolation [ 25 ] . Notably, a significant number of validation points, specifically in-situ measurements that are compared to the predicted values, showed negative correlations. Similarly, Wang et al. [ 26 ] used Random Forests (RF) to enhance the spatial resolution of GRACE-derived groundwater storage (GWS) changes from 300 km to 25 km across regions in sub-Saharan Africa, the Middle East, and South Asia from February 2003 to December 2022. They established a relationship between low-resolution GLDAS data (used as ground truth for their model) and various hydro-geological factors (used as input for training). Then, they employed high-resolution data as input to downscale the GWS estimates. However, their approach remained dependent on low-resolution GLDAS data, resulting in restricted precision and limited generalizability. Moreover, these models are constrained to their training periods, making them impractical for real-world applications. With this backdrop, in this paper we make an effort to propose an innovative approach that can bridge existing gaps, enabling accurate annual predictions of groundwater levels across years. 1.1 Our Contributions The primary objective of this research is to downscale GLDAS maximum and

Read the complete text