Integrating Remote Sensing and Machine Learning for Groundwater Withdrawal Estimation in Arizona

,

Groundwater is the largest source of Earth's liquid freshwater and plays a critical role in global food security.Hence, overuse of groundwater resources is a major concern.
It is hard to estimate groundwater use or storage at local scales.Existing satellite methods for estimating groundwater storage change involve using GRACE/GRACE-FO (https://en.wikipedia.org/wiki/GRACE_and_GRACE-FO)data at a coarse resolution (~ 400 km).
In this study, we combine publicly available datasets into a machine learning framework for estimating groundwater withdrawals (which are related to change in groundwater storage) at very high resolution (5 km) over the state of Arizona.Here, we use data from various sensors that measure different components of the water balance for monitoring groundwater withdrawal (SSEBop (https://earlywarning.usgs.gov/ssebop/modis),PRISM (https://prism.oregonstate.edu/),USDA-NASS CDL (https://www.nass.usda.gov/Research_and_Science/Cropland/SARS1a.php)).

REMOTE SENSING AND MACHINE LEARNING | WORKFLOW
This extends a previous study (Majumdar et al., 2020) in which we estimated groundwater withdrawals in Kansas, where the climatic conditions and aquifer characteristics are significantly different.The predictors for the RF model include (Smith & Majumdar, 2020), WS_PA (WS calculated using averaged P) and WS_PA_EA (WS calculated using averaged P adjusted with averaged ET), with the response variable being groundwater (GW) withdrawal.Accordingly, URBAN is the most important predictor followed by AGRI, ET, and SW.

GROUNDWATER WITHDRAWALS | STUDY AREA, RESULTS, AND ANALYSIS
For the RF model, we kept the number of trees as 500, and maximum number of featrues as 8.The predicted groundwater withdrawals at 5 km spatial resolution for both Arizona and Kansas show good accuracy.Most of the land subsidence is occurring in southern and south-eastern Arizona.

GROUNDWATER WITHDRAWALS | MORE ANALYSIS, LAND SUBSIDENCE
All subsiding areas have high or moderate predicted groundwater pumping.
Groundwater withdrawals are slightly correlated with land subsidence.Subsidence is a function of withdrawals, clay content and aquifer confinement.The mismatch between predicted pumping and subsidence can provide clues to these properties.

CONCLUSION AND FUTURE WORK
Our machine learning model shows promising results in sub-humid and semi-arid (Kansas) and arid regions (Arizona) at very high resolution (5 km), which proves the robustness and extensibility of our integrated approach combining remote sensing and machine learning into a holistic, automated, and fully-reproducible workflow.
The success of this method indicates that it could be extended to areas with more limited groundwater withdrawal data under different climatic conditions and aquifer properties.

Figure 3 :
Figure 3: Groundwater basins and sub-basins in Arizona highlighting the AMA/INA regions.This map has been downloaded from the Arizona Department of Water Resources (ADWR) (https://new.azwater.gov/sites/default/files/GWBasin_ShowingCAP_0.pdf)portal.

Figure 5 :
Figure 5: The feature or variable importances are values (sum up to 1) signifying the impact of each variable (the higher the value, the more importantthe feature).

Figure 6 :
Figure 6: Mean groundwater (GW) withdrawals for (a) the AMA/INA region in Arizona and (b) the entire state of Kansas (Majumdar et al., 2020).

Figure 7 :
Figure 7: Mean predicted groundwater withdrawals and mean land subsidence maps for the 2010-2018 period.

Figure 8 :
Figure 8: Proposed deep learning framework as part of future work