A Deep Learning Approach to Identify Center Pivot Irrigation

Agriculture in semi-arid regions, such as the U.S. Great Plains, is fraught with perils due to the vulnerabilities associated with infrequent, seasonal, and highly variable precipitation; cyclic drought; and land degradation due to misuse of scarce resources in these fragile ecosystems. One method employed throughout the world’s drylands to decrease vulnerability while increasing agriculture productivity is the use of irrigation (Wenger et al. 2017). Indeed, agriculture accounts for ~70% of human freshwater consumption worldwide (World Bank ND), a value expected to increase to keep pace with a rapidly expanding world population and changing climate.

The precipitous increase in irrigated farmlands in the Great Plains region initially began after the “Little Dust Bowl” of the 1950s with the introduction of new technologies, such as center pivot irrigation (CPI), which was thought to decrease vulnerability to drought. Further, plentiful groundwater, increasing competition from corporatization, and cheap fuel led to its rapid expansion in the 1980s (Opie 1993; Vadjunec et al 2019). While such technologies helped “drought proof’” agriculture operations and permitted the production of crops once limited to wetter climates (e.g. corn), numerous adverse ecological effects accompanied these technological and subsequent land use/land cover changes, among them groundwater drawdown, soil salinization, changes in surface water flow, waterlogging, changes in biogeochemical cycles, and even alterations in local and regional climate (McDermid et al. 2023). Understanding dryland land use/land cover dynamics is, therefore, paramount to fostering socioecological resilience.

The High Plains Aquifer (HPA) system lies in semi-arid region that has seen the prolific growth of irrigated agriculture over the past 75 years. The HPA is the largest aquifer system in the U.S., underlying 174,000 mi2 (450,658 km2), bisecting 237 counties across eight states (USGS 2017). The region is among the most productive cropland areas in the U.S. and supports an estimated 200,000 irrigation wells (Diffendal 2011). Concurrent with this rapid expansion of irrigated agriculture has been a steep decline in groundwater levels (McGuire 2017). Ironically, the same technologies that helped create an agriculture and economic powerhouse (Anderson 2018), now threatens the very livelihoods of those producers who have grown dependent on the HPA to support their operations, leading to what the New York Times (Searcey and Erdenesanaa 2023) called a “nationwide groundwater crisis.”

Given the geographic extent, past attempts to estimate CPI coverage within the High Plains Aquifer system have frequently involved manually digitizing irrigated fields in a small  portion of HPA over relatively coarse (e.g. decadal) temporal scales (see Wenger et al. 2017; Hassani et al. 2021; Vadjunec et al. forthcoming). However, recent developments in GeoAI, in particular deep learning, may improve the efficacy of estimating changes in CPI coverage throughout the HPA over long time periods. In this study, we demonstrate the feasibility of combining Landsat 5 and Landsat 8 imagery acquired from Google Earth Engine (GEE) with Esri’s deep learning libraries to estimate CPI coverage in the HPA over time.

The generalized deep learning process for object detection involves preparing training samples, training the model, detecting objects using the trained model (model inference), and reviewing the results.

As part of a 15-year, ongoing mixed methods study on socioecological systems (SES) resilience in the Southern Great Plains (e.g. Vadjunec and Sheehan 2010; Fagin et al. 2016; Wenger et al 2017; Vadjunec et al. 2022; Carrasco et al. 2024; Fagin et al. 2024), we manually digitized CPI fields in a tri-county, tri-state region from National Agriculture Imagery Program (NAIP) data for 2000, 2005, 2010, and 2020 (also see Vadjunect et al. forthcoming). We used these datasets for training the model.

We wrote Google Earth Engine (GEE) scripts to mask and export Landsat 5 and Landsat 8 data for every other year from 2001 to 2023. Additionally, because the HPA crosses two UTM zones (13 and 14) and given the geographic extent of the HPA, we divided the HPA into quadrants based on UTM zones for export. Despite this, GEE still exported images in tiles, necessitating mosaicking images.

Prior to using the ArcGIS Pro deep learning tools, we updated the attributes of the training data discussed above to ensure compatibility with the MaskRCNN (object detection) architecture. This requires a minimum of two classes and the classes must begin with 1 (e.g. a class of 0 is not acceptable).

Once updated, we used the Label Objects for Deep Learning tool in conjunction with the GEE extracted Landsat 8 imagery for 2021 (we later repeated the process with Landsat 5 data from 2013). This creates a series of 256 x 256 image chips used to train the model.

Lastly, we used the ArcGIS Pro Train Deep Learning Model tool and the image chips created in the last step to train the deep learning model (which we later repeated for Landsat 5 imagery).

Once we successfully trained the models, we ran the model across the entirety of the HPA for every other year from 2001 to 2013 for Landsat 5 and 2015 to 2023 for Landsat 8. Because of the manner in which we downloaded the GEE imagery, this required running the model 4 times (once for each quadrant) for each year under investigation.

We then manually edited the output datasets to clean up spurious output along the image border areas (with spectral inconsistencies). We also removed any incorrectly detected CPI fields in a few large areas known to not have any CPI (e.g. the Sandhills of Nebraska). Cleaned datasets were transformed to a common coordinate system and merged into a seamless coverage for each year.

We developed an accuracy assessment protocol to test the overall accuracy of our method using one output dataset from Landsat 5 (2013) and one from Landsat 8 (2021).

We successfully created CPI datasets across the HPA for every other year between 2001 and 2023. We performed an accuracy assessment on model outputs from both Landsat 5 (2011) and Landsat 8 (2015 and 2023). For the former, the overall accuracy was 92.75% and for the latter 89.175% and 98%, respectively.

Nonetheless, given the geographic extent of the HPA, we documented instances of both errors of commission (areas classified as CPI that should not have been) and omission (areas that should have been classified as CPI, but were not). Additionally, while the models did a good job of identifying CPI, the resultant extracted features did not always have the geometric fidelity that manually digitizing fields have. Also, given the adjacency of many CPI fields to one another, we encountered many instances in which adjacent fields coalesced into a single feature—limiting the feasibility to estimate the number of pivots throughout the HPA. Despite these limitations, our models provide a means to estimate the areal extent of CPI in the HPA and to monitor change over time.

The model’s output effects of cyclical drought in the High Plains are reflected in the model’s output and subsequent changes associated with land use and agriculture. CPI within the HPA experienced steady numbers between 2001 to 2011 ranging from 3.7 million to 4.2 million hectares. In 2013, the HPA saw a 23.93% increase of CPI with a staggering 4.9 million hectares. 2015 and 2017 show a slight decline with approximately 10.37% decrease in CPI from 2013 to 2019 but are followed by increase of 11.01% in 2021 and 9.49% increase in 2023. In 2023, the final year of our dataset, 5.3 million hectares of land was used for CPI. Compared to 2001, CPI showed an increase of 40% in 2023.

An analysis of state results show varying results. The severe drought conditions in the Southern Great Plains from 2013 to 2015 are represented in the model’s output. For instance, Oklahoma and New Mexico experienced a 17.41% and 12.33% decrease respectively while the other states’ CPI decline was not as severe. Some states show a drastic decrease in CPI while others experienced the opposite in the same year. For instance, from 2001 to 2003, South Dakota had a 30.25% decrease in CPI while Nebraska saw an increase of 23.39%.

The model also identified some counties that were outliers from the dataset. From our analysis, the model consistently detected little or no CPI during the range of years. We identified these counties to be: Bent County, CO, Larimer County, CO, Grant County, NE, Saunders County, NE, De Baca County, NM, Eddy County, NM, Guadalupe County, NM, Hall County, TX, and Winkler County, TX. Other counties show no CPI detected for few years, but not with the consistency as seen in the listed counties

We sought to determine whether we could combine GEE imagery with Esri’s deep learning libraries to detect CPI within the HPA and gauge CPI changes over time. Despite several limitations (such as distinct errors of omission and commission; geometric fidelity; and coalescence), our method is accurate, sound and repeatable. 

This research represents the first step in ongoing attempts to determine groundwater usage and other hydrological fluxes in the HPA over time. This work can also expand on previous work (e.g. Vadjunec et al 2018; Wenger et al. 2017) to link changes in irrigated agriculture to other factors, such as groundwater governance, federal subsidies, biofuel adoption, other land use/land cover change dynamics, and, ultimately socioecological resilience in semi-arid regions.

This work was supported by the USDA National Institute of Food and Agriculture, Agricultural and Food Research Initiative Competitive Program, Climate and Land Use, grant #2018-68002-28109, building on previous work by the National Science Foundation (NSF) research grant (#CMMI-1266381). Any opinions, findings, conclusions, or recommendations expressed here are those of the authors and do not necessarily reflect the views of our funders. We are extremely grateful to residents in the Southern Great Plains for their time and generous spirit.

The authors would like to thank undergraduate research assistants Payton Dampier and Daniel Ziegler for assistance with data analysis and StoryMap creation and Dr. Jason Vogel for his contribution to the broader ideas of this research.

For more information and other project deliverables, please visit:  www.experiencingdrought.org  or contact project PIs Dr. Todd D. Fagin at tfagin@ou.edu or Dr. Jacqueline M. Vadjunec at jvadjunec@ou.edu.