Assessing Differences in Groundwater Hydrology Dynamics Between In Situ Measurements and GRACE-Derived Estimates via Machine Learning: A Test Case of Consequences for Agroecological Relationships Within the Yazoo–Mississippi Delta (USA)

Abstract

In situ groundwater monitoring is critical for irrigated agroecosystems and informs land cover changes. Yet, such data can pose management challenges and confound agroecological relationships. Correspondingly, satellite-based approaches, including the GRACE-constellation, are increasing. Although in situ and GRACE-derived comparisons occur, limited research considers agroecological dependencies. Herein, we examined differences in groundwater monitoring approaches (observed [in situ, O] vs. predicted [GRACE-derived, P]) within the Yazoo–Mississippi Delta (YMD), an agroecosystem in the southeastern USA. We compared variations in modeled groundwater hydrology, land cover, and irrigation dynamics of the YMD within the upper-quartile (UQ) area of interest (AOI) (highest groundwater levels) and lower-quartile (LQ) AOI (lowest groundwater levels) every year from 2008 to 2020. Spatially, OUQ and PUQ were in northern portions of the YMD, with the OLQ and PLQ in southern portions. Groundwater levels between OUQ:PUQ and OLQ:PLQ each had correlations > 0.85. Regarding land cover, most categories varied within ±2.50% between model estimates over time. Relatedly, we documented 14 instances where correlations between land use category and groundwater level were inverted across models (OLQ:PLQ (5), OUQ:OLQ (6), PUQ:PLQ (3)). Irrigation results were not statistically different among all models. Overall, our results highlight the importance of quantifying model incongruences for groundwater and land cover management.

1. Introduction

Agricultural production necessitates water resource management and often incorporates corresponding land cover change. Globally, >90% of water use is directed toward agricultural purposes [1]. During contemporary periods, substantial irrigated agricultural expansion has occurred within already water-stressed regions [2]. Moreover, irrigation practices are dynamic, with corresponding regional effects [2]. Within the United States (US), irrigated agriculture (not including aquaculture) is responsible for the largest percentage of freshwater withdrawals (42%) [3]. In the US, 25% of the land in farms (ha) is irrigated, while 52% of cropland (ha) is irrigated [4]. Moreover, from 2008 to 2018, the US has seen a 12% increase in ha of land in farms and a 2% increase in ha irrigated [4]. Unsurprisingly, water stress occurs across several US agricultural sub-regions [5,6], with an estimated 60% of irrigated areas supplied by groundwater [5]. Among US agricultural sub-regions, the Mississippi Alluvial Plain (MAP) [7,8]—with its accompanying Mississippi River Valley Alluvial Aquifer (MRVAA)—experienced the second highest rates of groundwater extraction in 2015 (12,100 Mg/d), with 97% of withdrawals supporting irrigation [9]. In the MAP states of Arkansas, Mississippi, and Louisiana, there has been an overall decrease in ha of land used in farms (17%, 1%, and 7.5%, respectively), but a concomitant 15% increase in ha irrigated in Mississippi and Louisiana between 2008 and 2018 [4]. In Mississippi alone, 85% of irrigated ha reported only groundwater wells as source water in 2008; by 2018, that fraction of irrigated ha had dropped to 72% [4,10].

Correspondingly, groundwater monitoring efforts occur across the MAP [11,12] and are exemplified within the Yazoo–Mississippi Delta [YMD]. The YMD refers to the northwestern portions of the US state of Mississippi. The YMD features a sub-tropical and humid climate, with distinct wet (November–April) and dry seasons (May–October). Given seasonal precipitation patterns, embedded agriculture is extensively furrow (flood) irrigated and sourced from groundwater extractions [13]. Major commodities of the YMD are soybeans (Glycine max), corn (Zea mays), cotton (Gossypium hirsutum), rice (Oryza sativa), and catfish (Ictalurus punctatus) aquaculture—with ponds also sourced from groundwater [13,14,15]. Regionally, groundwater monitoring data is conducted through the Yazoo–Mississippi Delta Joint Water-Management District (hereafter, JWMD). JWMD monitors groundwater conditions within the unconfined alluvial aquifer in the YMD area twice annually (April and October). There are approximately 700 wells that have been monitored over the past 23 years (JWMD; https://www.ymd.org/ accessed on 15 May 2024). Although such monitoring is robust, spatial limitations and temporal gaps occur amidst this in situ record. Nevertheless, such data has informed numerous groundwater assessments, often highlighting “cones of depression” associated with irrigation intensity [14,16,17]. Yet, a greater understanding of the spatial and temporal dynamics of groundwater hydrology dynamics (with associated agroecological relationships) persists as a management focus across the MAP [18,19]. Increasingly, studies have begun integrating satellite-based data and other “big data” sources with existing in situ assessments [18,19,20]. Such integrations can provide the availability of high spatial and temporal hydrological products, which are crucial for management [20,21]. Satellite-based data from The Gravity Recovery and Climate Experiment (GRACE) has received recent scientific attention.

The GRACE mission flew twin spacecraft around the Earth to study key changes in the planet’s water, ice sheets, and the solid earth. Over the past 22 years, GRACE (2002–2018) and GRACE-FO (follow-on) (2018–present) have provided valuable hydrological data from on and below the surface of the Earth monthly. Despite the coarse resolution of GRACE data (originally ~300 km), this product has been used extensively to monitor for groundwater storage changes in recent years [22,23,24,25]. To overcome the challenge of using GRACE products for local-scale studies, researchers used different methods to downscale GRACE data [23,24,25,26,27,28]. These methods are either dynamic or statistical methods. Dynamic methods need high-resolution images and are computationally costly. In statistical methods, the correlation between large-scale and small-scale variables is utilized to generate the final higher-resolution outcome [23,26,29]. The most used methods in downscaling GRACE images are Random Forest (RF), Artificial Neural Network (ANN), and related Multilayer Perceptron (MLP).

Although recent GRACE-derived groundwater assessments have been shown to accurately estimate in situ sub-surface behaviors (e.g., seasonal volume changes), the spatial extent of these estimates vary [27,30,31]. Regionally, such potential spatial discontinuity may reflect stakeholder activities and linked agroecological relationships (e.g., irrigation rates and crop choice). Additionally, such linkages may be subject to lag effects (e.g., market responses) [14,15,32], further highlighting the necessity of identifying and quantifying incongruencies between in situ and GRACE-derived groundwater estimates. Indeed, testing computational models and identifying associated knowledge gaps is key for examining the pattern–process relationships for both agricultural and non-agricultural categories across varied machine learning (ML) models. Model testing is particularly important in agriculture given its dynamic nature, and can help in forecasting crop yields, optimizing resource allocation, or monitoring biodiversity. These tools allow managers to anticipate and respond to changes in environmental conditions, optimizing practices for sustainability and efficiency. In non-agricultural contexts, ML can be used to track species populations, monitor habitat loss, or assess the impact of climate change. By integrating ML into management strategies, stakeholders can make data-driven decisions, improving long-term outcomes. Yet, while such tools may be useful, failures to examine model differences may lead to inappropriate management responses and exacerbate agroecological concerns.

Consequently, the primary objectives of this study were (1) to spatially and temporally quantify differences of groundwater monitoring models (observed [in situ, O] vs. predicted [GRACE-derived, P]) within the YMD from 2008 to 2020. To facilitate these comparisons, we (2) quantified variation in groundwater hydrology (mean elevation of water surface; meters above mean sea level [m.a.m.s.l]) and land cover dynamics (as measured by percentage of landscape [PLAND]) of the YMD within the upper-quartile (UQ) area of interest (AOI) (highest groundwater levels) identified for each model over time. Similarly, (3) we quantified variation in groundwater hydrology and land cover dynamics within lower-quantile (LQ) AOI (lowest groundwater levels) identified for each model over time. Land cover results were (4) correlated with groundwater hydrology dynamics to identify strengths and signs of relationships among model AOIs. Relatedly, we (5) estimated differences in total irrigation volume among model AOIs. Lastly, we (6) summarized our results and discussed potential agroecological management considerations based on these patterns and processes.

Accordingly, the fulfillment of these objectives provided novel contributions toward managing irrigated agroecosystems and understanding their accompanying land cover changes within the YMD. Our study uniquely documented an array of agroecological relationships between observed and predicted groundwater levels and associated land uses. Further, our study indicated instances where these relationships were inverted between observed and predicted models. Moreover, we identified land-cover-specific variations and latitudinal relationships. Combining these advancements may inform water management strategies to better identify sub-regional trends in the YMD, and our methodology may be applicable to wider alluvial systems.

2. Materials and Methods

2.1. Data

2.1.1. GRACE Mascon

In 2002, the GRACE twin satellites commenced their orbit, tasked with the delicate measurement of Earth’s gravitational variations over time [21]. These intricate readings, once distilled from oceanic and atmospheric influences, serve to illuminate shifts in terrestrial water reserves. Our research harnessed the GRACE mascon monthly datasets, selected for their enhanced spatial finesse relative to the traditional GRACE offerings. The mascon data, with an approximate resolution of 56 km, stands in stark contrast to the standard resolution of about 110 km. Our first task was to refine these data grids, applying statistical downscaling to bolster their spatial precision, thereby rendering the GRACE insights more relevant for localized analysis within Mississippi’s diverse landscapes.

2.1.2. Precipitation

In the realm of precipitation analysis within our study, CHIRPS (Climate Hazards Group Infrared Precipitation with Stations) data emerged as an invaluable asset [33]. Since 1981, its near real-time, quasi-global precipitation metrics have provided a temporal tapestry of precipitation patterns, instrumental for our Artificial Neural Network’s (ANN) calibration. CHIRPS’s granular 5 km spatial resolution imports a high level of precision into our model, which is crucial for the meticulous downscaling of GRACE data. This precision enables a more nuanced capture of recent precipitation dynamics, a cornerstone for the enhanced granularity of our research within Mississippi’s varied landscapes [33].

2.1.3. TerraClimate

Our exploration into the hydrosphere’s nuances utilized TerraClimate’s extensive dataset, which has charted a monthly chronicle of climate and climatic water balance across terrestrial regions since 1958. With an annual update rhythm and a spatial precision of 4 km, this dataset offered insights into variables pivotal for aquifer dynamics, including evapotranspiration and surface runoff [34]. Our study synthesized longitudinal and latitudinal coordinates with a suite of hydrologically pertinent variables to enrich the granularity of our analysis, particularly focusing on the intricate water balance dynamics pertinent to our study’s scope.

2.1.4. In Situ YMD Groundwater Hydrology Measurements and Land Use/Land Cover (LULC)

In situ groundwater hydrology data from 2008 to 2020 (in the form of water-level elevation) were electronically provided from The Yazoo–Mississippi Delta Joint Water-Management District (JWMD; https://www.ymd.org/; Received 18 May 2022, [20]). Land use and land cover data were derived from the United States Department of Agriculture National Agricultural Statistic Service’s Cropland Data Layer (CDL) [35,36] (USDA NASS CDL; https://nassgeodata.gmu.edu/CropScape/, accessed on 15 June 2022) for the 19-county extent of the YMD. Regionally, those data had been previously processed by [15], with final products accessible through AgDataCommons (https://doi.org/10.15482/USDA.ADC/1529589). As such, we used their nine land cover classification schema for analysis and visualization: soybeans, corn, cotton, rice, wetlands, water, upland forest, other agriculture, and developed were the categories. Conceptually, these categories may be regarded as “agricultural” (soybeans, corn, cotton, rice, and other agriculture) and “non-agricultural” (wetlands, water, upland forest, and developed).

2.2. Methods

2.2.1. Data Preprocessing

We began by aligning the temporal resolution of the GRACE satellite data with monthly average climatic variables from the TerraClimate and CHIRPS datasets. This alignment was essential for ensuring dataset coherence. We then resampled all the inputs to match the highest spatial resolution of the inputs we have (i.e., 4 km of TerraClimate) and clipped all datasets to the geographic boundaries of the YMD, thereby enhancing the geographical relevance of our analysis. Data normalization was performed using Python’s scikit-learn ‘StandardScaler’ to standardize the data to a mean of zero and a standard deviation of one, which prepared it for integration into the Artificial Neural Network (ANN) model, thus enhancing its predictive precision. To validate the model’s accuracy, we divided the data via random sampling into training (60%) and testing (40%) sets, commensurate with other machine learning practices [37,38], to ensure a robust assessment of the ANNs model’s performance on unseen data.

2.2.2. ANN Model

We adapted and applied a feed-forward, back-propagation Artificial Neural Network (ANN) methodology to downscale GRACE mascon satellite data for groundwater level prediction in the YMD region [39]. To construct our ANN models, we employed a variety of data sources, including GRACE mascon observations, TerraClimate variables, and precipitation data from the CHIRPS dataset. The optimized model comprised 13 input nodes, a single hidden layer with 26 nodes, and one output node, demonstrating a commendable fit to the data with an R² of approximately 0.85 for initial model development. This initial phase laid a solid foundation for the subsequent application of ANNs in predicting water table levels, emphasizing the role of a balanced network structure in achieving high predictive accuracy. Building upon this groundwork, the second phase of our investigation leveraged the downscaled GRACE data from the first-phase ANN models, integrating it with observed YMD well data and various CHIRPS and TerraClimate variables. This comprehensive dataset enabled the construction of a refined ANN model aimed at forecasting groundwater levels across the YMD. The final model featured 14 input variables, including downscaled GRACE, CHIRPS, and TerraClimate data, a hidden layer of 17 nodes, and a singular output node, achieving an outstanding R² of 0.99 during training and 0.87 on validation data. This phase underscored the importance of data integration and standardization in enhancing the model’s predictive capabilities, showing the potential of ANN models in environmental applications. Through iterative training and validation, the models exhibited a high degree of accuracy, affirming their utility in groundwater level prediction and offering a valuable tool for regional water resource management (Figure 1). For additional details, please refer to the Supplementary Materials.

2.2.3. AOI Delineation and LULC Integration

After completing the second phase of the ANN model, we then generated a dataset containing predicted water levels for wells within designated cells spanning the years 2008 to 2020. By using this dataset, we applied the Kriging interpolation technique within ArcGIS Pro (ESRI, Redlands, CA, USA) to estimate water levels across the entire YMD region. Subsequently, employing the quantile method within raster symbology, we stratified cells based on water level values into four distinct categories. Each quantile was then vectorized to delineate upper and lower quantiles within the YMD region. With delineated AOIs established for each year, we conducted zonal statistics analysis to determine the percentage distribution of each land use/land cover (LULC) category within the defined AOIs for each year.

2.2.4. Statistical Analysis of Groundwater Hydrology, LULC Category, and Calculated Irrigation Volume Differences

To investigate the influence of the LULC category on groundwater hydrology, Pearson correlations between categories were calculated across the span of 2008–2019 (to address the inherited lag effects) in R (v 3.4.1) [40] using the cor() function [39]. Correlation heatmaps were constructed using the package corrplot (v 0.92) [41]. As irrigation (primarily sourced from groundwater) [13] is integral to agricultural productivity within the YMD and influenced by land cover change, including crop rotations [14,32,42,43], we considered the potential differences between total irrigation volumes between models (e.g., OUQ vs. PUQ, and OLQ vs. PLQ) as calculated from YMD specific surveys for key rotational crops [14,32,42]. Crop rotations in the YMD are typically rice–soybean or soybean–corn, yet variations exist [15]. Irrigation intensity decreases from rice to corn to soybeans. According to [42], regional irrigation rates from 2002 to 2013 averaged 2800 m³ha⁻¹, 3100 m³ha⁻¹, 1800 m³ha⁻¹, and 9200 m³ha⁻¹; for soybeans, corn, cotton, and rice, respectively. Additionally, ref. [42] estimated that the proportion of each crop type that experienced at least one year of irrigation to be 61.00%, 67.00%, 56.00%, and 99.00%, respectively. As such, we integrated these data with our previous results and estimated irrigation volume differences between OUQ:PUQ and OLQ:PLQ models. Irrigation volumes were generated by the following procedure: [# of cells of a LULC category = x] multiplied by [proportion of crop that is regionally irrigated = y] which was multiplied by 900 (converting cells to square meters) then divided by 10,000 (which converts square meters to hectares). The area in hectares was then multiplied by [regional average irrigation rate of the given crop = z]. The process was repeated for soybeans, corn, cotton, and rice. Regarding subsequent ANOVA, single factors were conducted in Excel Office 16 (Microsoft, Redmond, WA, USA).

3. Results

3.1. Observed Model Results

During our focal timespan, the location and extent of the OUQ AOI was highly consistent among northern counties of the YMD. However, the location of the OLQ AOI consistently occurred within southern counties of the YMD. Yet, the extent and shape of the OLQ AOI became less stable from 2016 onward (Figure 2).

• Groundwater hydrology and LULC dynamics of the OUQ and OLQ AOIs (Tables S1–S4)

3.2. Predicted Model Results

During our focal timespan, the location and extent of the PUQ AOI were similarly highly consistent among northern counties of the YMD. As expected, the location of the PLQ AOI consistently occurred within southern counties of the YMD. Interestingly, the extent and shape of the PLQ AOI featured high stability over time (Figure 3).

• Groundwater hydrology and LULC Dynamics of the PUQ and PLQ AOIs (Tables S5–S8)

3.3. Comparisons Between OUQ vs. PUQ

3.3.1. Groundwater Hydrology Dynamics

Groundwater hydrology levels between the OUQ and PUQ varied by 0.19 m.a.m.s.l. on average over time. Generally, the PUQ underestimated groundwater levels except for 2014 and 2020. Over time, there were only seven wells that were not spatially shared between the OUQ and PUQ AOIs (Figure 4).

3.3.2. LULC Dynamics

Comparing the OUQ versus the PUQ AOIs, we documented minor variations in percent land cover for each of the nine focal categories over time. Indeed, these variations never exceeded +2.5% for any given year or category. Yet, the PUQ underestimated soybeans in all years, and overestimated wetlands in all years. In terms of overall area extent, on average, these AOIs differed by ~1.50% over time (Figure 5).

3.4. Comparisons Between OLQ vs. PLQ

3.4.1. Groundwater Hydrology Dynamics

Groundwater hydrology levels between the OLQ and PLQ varied by 0.05 m.a.m.s.l. on average over time. Overall, the PLQ underestimated groundwater levels with exceptions in 2008, 2009, 2013, 2014, and 2017. Over time, there were 52 wells that were not spatially shared between the OLQ and PLQ AOIs (Figure 6).

3.4.2. LULC Dynamics

Comparing the OLQ versus the PLQ AOIs, we similarly documented a mix of minor variations in percent land cover for each of the nine focal categories over time. Indeed, most of these variations never exceeded ±2.5% for any given year or category. However, there were eight instances where variation exceeded ±2.5%: soybeans (2018, 2019, 2020), wetlands (2016, 2018, 2019, 2020), and other agriculture (2019). Moreover, the PLQ overestimated rice occurrence in all years. In terms of overall area extent, on average, these AOIs differed by ~2.00% over time (Figure 7).

3.5. Correlation Analyses of Overall Model Behaviors

3.5.1. Groundwater Level Correlations (Table S9)

Holistically, groundwater levels were strongly correlated (> 0.85) across both the upper and lower quartiles. These results are reflective of the congruent patterns observed in Section 3.3.1 and Section 3.4.1. The groundwater level relationship in the lower quartiles was marginally higher than that of the upper quartiles.

3.5.2. Groundwater Correlations for LULC Sub-Estimates (Table S9)

Within the upper quartile, groundwater levels were strongly correlated for all LULC categories, with an average value of 0.97. Corn, cotton, and other agriculture featured the highest correlation values (1.00). Upland forest (0.89) and developed (0.96) featured the only below-average correlation values. Within the lower quartile, groundwater levels displayed a more variable range of correlation values, with an average value of 0.79. Rice featured the highest correlation value (1.00), with wetlands and upland forest featuring the lowest correlation values (0.25). The develop category had a correlation value of (0.72), indicating that all below-average correlation values were associated with non-agricultural classes.

3.6. Correlative Relationships for Groundwater Levels and LULC Categories Among Models

As indicated in Section 3.5, the GRACE-derived model robustly estimates overall groundwater levels within the YMD. Yet, the strength of the model varied spatially and among LULC categories. As such, we further evaluated the above correlations for directionality and potential sub-regional effects. We detected five signed correlative differences among models, yet only between the OLQ and PLQ. These results represented the initial focus of the study. Our expectations were that these differences would showcase the most divergent agroecological relationships (Figure 8).

3.6.1. OUQ vs. PUQ

Correlative relationships for groundwater levels and LULC categories between the OUQ and PUQ were analogous in terms of their signed relationships. Yet, minor differences in the strengths of correlations among LULC were detected. For the OUQ, the water category was the most positively correlated with groundwater levels. Whereas cotton was the most negatively correlated. For the PUQ, the wetlands category was the most positively correlated, and corn was the most negatively correlated.

3.6.2. OLQ vs. PLQ

Correlative relationships for groundwater levels and LULC categories between the OLQ and PLQ were analogous and positive in terms of signed relationships for rice and other agriculture. Relationships were analogous and negative for corn and cotton. All other relationships were mixed. Soybeans, other agriculture, and developed categories went from positively correlated in the OLQ to negatively correlated in the PLQ. Whereas the wetlands and water categories went from negatively correlated in the OLQ to positive in the PLQ. For the OLQ, rice was the most positively correlated, and cotton was the most negatively correlated. For the PLQ, the wetlands category was the most positively correlated, and cotton was the most negatively correlated.

3.7. Additional Correlative Relationships for Groundwater Levels and LULC Categories

We also detected nine correlative differences within models (OUQ vs. OLQ (6) and PUQ vs. PLQ (3)). These results represented an extension of our initial focus. Documenting these differences supported the broader goals of the study by highlighting model sensitivities for sub-regional YMD effects (Figure 8).

3.7.1. OUQ vs. OLQ

Correlative relationships between the OUQ and OLQ were analogous and positive in terms of signed relationships for rice and upland forest categories. Relationships were analogous and negative for corn. All other correlative relationships were mixed. Cotton, wetlands, and water categories went from positively correlated in the OUQ to negatively correlated in the OLQ. Other agriculture went from negatively correlated in the OUQ to positively correlated in the OLQ.

3.7.2. PUQ vs. PLQ

Correlative relationships between the PUQ and PLQ were analogous and positive in terms of signed relationships for rice, wetlands, and water categories. Relationships were analogous and negative for corn and developed categories. All other correlative relationships were mixed. Cotton and upland forest categories went from positively correlated in the PUQ to negatively correlated in the PLQ. Other agriculture went from negatively correlated in the PUQ to positively correlated in the PLQ.

3.8. Trends Across Models

Overall, we detected that rice was the only category that was positively correlated with groundwater levels across all models. In the opposite fashion, the corn category was negatively correlated with groundwater levels across all models. Cotton and other agriculture categories were positively correlated with groundwater levels in OUQ|PUQ pairs, yet became negatively correlated in OLQ|PLQ pairs. All other relationships contained at least one instance of a sign change.

3.9. Estimates of Irrigation Volume Differences Between Models

In these analyses, we examined soybeans, corn, cotton, and rice [42].

3.9.1. OUQ vs. PUQ (Tables S10–S12)

Numerically, the PUQ overestimated total irrigation volumes over time for soybeans (154,135,966.32 m³ha⁻¹), cotton (6,411,998.88 m³ha⁻¹), and rice (79,113,636.36 m³ha⁻¹). The PUQ underestimated irrigation volumes for corn (3,887,583.21 m³ha⁻¹) (compare Tables S10 and S11). Irrigation volumes for soybeans were overestimated for all years, like rice, except for 2017, where the estimates were equivalent. Corn featured overestimates of irrigation volumes in 2009, 2012, 2015, 2019, and 2020. Cotton featured overestimates of irrigation volumes for all years except the span of 2009–2011. All crops had overestimates in 2012, 2015, 2019, and 2020 (Table S12). Statistically, however, estimated irrigation volumes for all crops combined over time between the OUQ and PUQ showed no significant difference (ANOVA: single factor, p = 0.83). Crop-specific irrigation volume comparisons over time were also not significant for soybeans (ANOVA: single factor, p = 0.49), corn (ANOVA: single factor, p = 0.96), cotton (ANOVA: single factor, p = 0.95), and rice (ANOVA: single factor, p = 0.64).

3.9.2. OLQ vs. PLQ (Tables S13–S15)

Numerically, the PLQ overestimated total irrigation volumes over time for cotton (5,737,314.24 m³ha⁻¹). The PLQ underestimated irrigation volumes for soybeans (179,301,621.24 m³ha⁻¹), corn (57,691,458.18 m³ha⁻¹), and rice (360,179,229.96 m³ha⁻¹) (compare Tables S13 and S14). Irrigation volumes for soybeans were overestimated in 2009 and 2017. Corn featured overestimates of irrigation volumes from 2008, 2009, 2010, and 2017. Cotton featured overestimates of irrigation volumes from 2008, 2009, 2010, 2011, 2013, 2014, 2017, 2019, and 2020. Rice never featured an overestimate (Table S15). Statistically, however, estimated irrigation volumes for all crops combined over time between the OLQ and PLQ showed no significant difference (ANOVA: single factor, p = 0.58). Crop-specific irrigation volume comparisons over time were also not significant for soybeans (ANOVA: single factor, p = 0.56), corn (ANOVA: single factor, p = 0.74), cotton (ANOVA: single factor, p = 0.82), and rice (ANOVA: single factor, p = 0.23).

4. Discussion

4.1. Geospatial Considerations

During the focal timespan, the YMD experienced dynamic groundwater hydrology and LULC changes [13,14,15,20,29,32,42]. Based on in situ groundwater well measurements, northern portions of the YMD consistently featured the highest groundwater levels. Resultant OUQ AOIs reflected this pattern and had limited differences, generally along southern and eastern portions of extents year-to-year. Southern portions of the YMD consistently featured the lowest groundwater levels. Resultant OLQ AOIs reflected this pattern as well, yet the shape of the AOIs was less consistent. Indeed, by 2016, the OLQ was bifurcated, and nearly so for each subsequent year. Additionally, we observed consistent protrusions of this AOI along its western boundary. Also interestingly, OLQ AOIs from 2009 to 2011 featured an internalized sub-component. We attribute these instances as artefactually inherited from groundwater well records and did not affect LULC clipping extents.

The PUQ and PLQ also modeled the generalized latitudinal patterns of higher in situ groundwater measurements in the north and lesser amounts in the south. Indeed, overall area extent differences between observed and predicted AOI pairs were less than 2%. However, these numerical similarities belied an important spatial configurational difference between the OLQ and PLQ. The PLQ was not sensitive to the bifurcation and was also insensitive to changes along the western boundary of the OLQ.

4.2. Agroecological Management Considerations

4.2.1. Groundwater Hydrology and LULC Categories

Overall, the GRACE-derived model was remarkably well correlated to observed groundwater levels between the OUQ|PUQ (0.86) and OLQ|PLQ (0.90). On average for all LULC categories, the OUQ|PUQ was exceptionally well correlated (0.97), whereas the OLQ|PLQ was not as robust (0.79). The lower correlative relationship of the OLQ|PLQ were driven by the upland forest (0.25), wetlands (0.25), and developed (0.72) categories.

Relatedly, although most variations in LULC documented in this study were relatively minor with respect to percent coverage, such changes were correlated with groundwater hydrology relationships across models. As evidenced in Section 3.6, Section 3.7 and Section 3.8, we found 14 instances of signed changes in correlative relationships among LULC categories. These sign changes included agricultural (soybeans, cotton, and other agriculture) and non-agricultural categories (wetlands, water, upland forest, and developed).

4.2.2. Estimates of Irrigation Volume Differences

Variations in overall estimated irrigation volume differences over time for all focal agricultural LULC categories were non-significant for both the OUQ|PUQ and OLQ|PLQ models. However, we did detect differences in the strengths of the non-significance between the paired OUQ|PUQ and OLQ|OLQ estimates for each focal agricultural category. Relatedly, over time, the irrigation volume for soybeans was overestimated by the PUQ and underestimated by the PLQ. For corn, the PUQ underestimated irrigation volume as did the PLQ. Irrigation volume for cotton was overestimated by the PUQ and the PLQ. Rice irrigation was overestimated by the PUQ and underestimated by the PLQ.

4.3. Examination of Pattern–Process Relationships for Agricultural and Non-Agricultural Categories Between and Across Models

4.3.1. Agricultural Categories

Reviewing findings between Section 4.1 and Section 4.2 provided a mix of anticipated and unexpected pattern–process relationships within the YMD. With respect to soybeans, groundwater level correlations were negative in all models except the OLQ. Although, in that instance the strength of the relationship was the nearest to neutral (0.01) of all relationships. Thus, the general negative pattern aligns with regional agronomic trends—the bulk of soybeans are irrigated, albeit at the second lowest rate overall [43], which we expected would reduce local groundwater levels. Moreover, during contemporary periods, soybean production within the YMD has expanded [14,15,32]. Combined estimates from PUQ and PLQ indicate that overall, over time, irrigation volume for soybeans was underestimated by 25,165,654.92 m³ha⁻¹. For corn, groundwater level correlations were negative in all models. Unlike soybeans, relationships were more negative within northern portions of the YMD. Indeed, the OUQ correlation had the highest strength of all relationships detected (−0.72). Generally, this pattern also aligns with regional agronomic trends—provided the higher irrigation demand and related prevalence [42]. Combined estimates from the PUQ and PLQ indicate that the overall irrigation volume for corn was underestimated by 61,579,041.39 m³ha⁻¹.

In contrast, cotton presented positive correlations with groundwater level in northern regions of the YMD and negative correlations in the southern YMD. This relationship also aligns with regional agronomic trends, considering that cotton requires the least irrigation demand and has a regionally lower prevalence [42]. Indeed, cotton production can occur throughout the YMD; however, contemporaneous patterns are concentrated in northern portions of the YMD (Bolivar County, Coahoma County, Quitman County, and Tunica County) where sandy soil conditions are more favorable [15,44]. Thus, the pattern of positive correlations to groundwater level in northern portions of YMD may represent a function of lower irrigation rates and extents combined with higher levels of soil infiltration. The same consideration is plausible within the southern YMD. However, given the relative rarity of cotton in the OLQ and PLQ AOIs, we suspect that cotton may be imparting a disproportionate effect across models. Combined irrigation volume from the PUQ and PLQ indicated an overall overestimate for cotton of 674,674.64 m³ha⁻¹.

Among the focal crop categories, we had anticipated rice to feature the most negative correlations. Instead, we detected positive relationships across all models. As such, this pattern is perplexing, given both the irrigation demand and the overall frequency of irrigation [42]. Regionally, the spatial extent of rice production is low and has been declining. At the same time, rice production has been concentrated into more localized zones [15]. Combined estimates from PUQ and PLQ indicate that the overall irrigation volume for rice was underestimated by 281,065,593.60 m³ha⁻¹. Possible explanations for these unexpected results are: (1) rice as a LULC category is relatively rare—such components can have outsized influence on models, akin to cotton above; (2) regionally, rice production involves levees and standing water during the production season [42], yet the structure of these levees may differ [45]—which could impart differential hydrological conditions. Taken together, we have concurred with others for greater research emphasis on GRACE-derived and other remotely sensed products in rice production environments of the YMD [18,19].

Lastly, other agriculture also displayed a distinct latitudinal gradient in terms of groundwater relationships, where northern portions of the YMD were negatively correlated and southern portions positively correlated. Other agriculture as a LULC category encompasses several alternative production systems and includes “fallow/idle cropland”. The consideration of “fallow/idle cropland” may explain the positive correlations in the southern YMD because of flooding events [15]. Meanwhile, other agriculture in the northern YMD may represent irrigated specialty crop production.

4.3.2. Non-Agricultural Categories

Shifting to non-agricultural categories, wetlands featured positive correlations with groundwater levels in all models except the OLQ. The positive relationships were broadly anticipated, given the integral ecosystem services provided by wetlands [46] and their ability to mediate groundwater recharge [47]. Thus, the behavior of the OLQ was perplexing. We suspect that the bifurcated nature of this AOI may have influenced this relationship. Yet, this outlier pattern reinforces the assertion that sub-regional effects are important to consider in GRACE-derived models. Similar trends were documented for the water category. The negative correlation of the OLQ may also be attributed to the bifurcated AOI. Of note, aquaculture was included in this category. Aquaculture (primarily catfish) is a major economic driver in the YMD [14,48], concentrated in southern portions of the YMD [15], which also rely on groundwater inputs [13]. Accordingly, this may also partially account for the negative relationship of the OLQ and the comparatively weaker positive relationship of the PLQ. Whereas the strength of the positive correlation was highest for the OUQ, which we attribute to the longer border and associated lakes of the Mississippi River. Indeed, that correlation had the highest strength of all positive relationships detected (0.71). Similarly, the upland forest category featured positive correlations with groundwater levels across all models except the PLQ. We considered upland forests to function in a conceptually similar fashion to wetlands, which often contain forest remnants. The upland forest category was restricted to the eastern peripheries of the YMD and rare [15]. Thus, we suspect that the overall rarity may be poorly reflected in the PLQ. Lastly, the developed category featured negative correlations with groundwater level, except the OLQ. Again, we suspect that the bifurcated nature of this AOI may be driving this unexpected behavior. Nevertheless, these generally negative relationships were broadly anticipated, given the ecological effects that can be associated with wider regional urbanization [49]. Noting, however, that the relationship between development and groundwater recharge may vary [50] and that relationships within the rural YMD remain poorly documented.

4.3.3. Uncertainty Considerations, Limitations, and Future Research Directions

Provided the correlation-based study design, we acknowledge that confounding effects from both incorporated and non-incorporated variables may exist in this study. Groundwater flow patterns and geologic variations within the MRVAA are complex [51] and were not directly assayed. The MRVAA features a natural southward gradient and generally receives groundwater flows from the eastern periphery. Additionally, contemporary sub-surface flows in the “zones of depression” result in discharges to irrigation wells instead of their historical riverways [51]. Accordingly, these sub-surface patterns may be a factor that fosters insensitivity for some of our results in the southern YMD. Another source of confounding variables is related to LULC classification accuracy. Our study used data from [15], which included reclassification and filtering techniques. Those adjustments were made to reduce inherited errors from original CDL layers and to reflect crop-specific classification accuracies. Given that most of the percentage differences in LULC between the OUQ:PUQ and OLQ:PLQ pairs were minor (±2.5%), it is possible that some correlative relationships were obscured. Despite both realms of uncertainty, our approach and results remained germane to the wider applicability of remotely sensed studies for the YMD [18,19].

Related to the above uncertainties, we acknowledge the limitations of our study. Our study design only investigated correlative relationships among groundwater levels and land cover types within the realms of the OUQ|PUQ (upper quantiles) and OLQ|PLQ (lower quantiles) over time. As such, secondary and tertiary quantiles, which occurred within the western and central portions of the YMD, were excluded from our analyses. Thus, our study cannot directly comment on any correlative relationships that may have occurred therein. Other limitations of our study pertain to the role of seasonality associated with agricultural land cover categories inherited via the CDL. While typically agricultural land cover assignments are applied year-round, actual production field surfaces are dynamic. As such, our resultant relationships reflect periods of vegetative cover. Yet, associated irrigation rates across crop-production stages and specific crop varietals are beyond the scope of this analysis. Moreover, our irrigation estimates were based on data collected from 2002 to 2013. Irrigation patterns and methods may have changed during our focal period. Indeed, while these limitations have existed, they also provide a basis for future investigations. Assessments of excluded quartiles and associated correlations would be a direct extension. As noted, we recommend more GRACE-based assessments within rice production environments. Lastly, an expansion of our methodology to analogous alluvial systems within the wider MAP and internationally is envisioned.

5. Conclusions

Our assessment of groundwater hydrology dynamics, comparing in situ measurements with GRACE-derived estimates via machine learning in the YMD, revealed only minor discrepancies between datasets. Indeed, the GRACE-derived model effectively predicted in situ groundwater levels with a strong overall correlation of 0.88. Relatedly, the total surface area estimates for the areas of interest (AOIs), representing the upper (highest groundwater levels) and lower (lowest groundwater levels) quantiles, showed an average ~1.75% difference in area across all models.

Despite this numerical similarity, we observed that the GRACE-derived AOIs were less sensitive to their in situ counterparts in southern portions of the YMD. Furthermore, LULC characteristics within the AOIs exhibited a range of correlative effects on groundwater hydrology. Rice was positively correlated in all models, whereas corn was negatively correlated in all models. Cotton and other agriculture displayed latitudinal variations in their correlations. Wetlands and water categories had generally positive relationships outside of the OLQ. Soybeans and developed categories had generally negative relationships outside of the OLQ. In contrast, other agriculture had generally positive relationships outside of the PLQ. However, irrigation estimates revealed no statistically significant differences across models for the focal crops.

The observed differences underscore the importance of examining sub-regional effects within the YMD, which likely influence agroecological responses. Given the continued reliance on groundwater irrigation and prevailing agricultural practices in the YMD, accurate large-scale groundwater mapping remains essential. We believe that a GRACE-derived approach can effectively support the interests and managerial needs of stakeholders.