Exploring the Regional Dynamics of U.S. Irrigated Agriculture from 2002 to 2017

: The United States has a geographically mature and stable land use and land cover system including land used as irrigated cropland; however, changes in irrigation land use frequently occur related to various drivers. We applied a consistent methodology at a 250 m spatial resolution across the lower 48 states to map and estimate irrigation dynamics for four map eras (2002, 2007, 2012, and 2017) and over four 5-year mapping intervals. The resulting geospatial maps (called the Moderate Resolution Imaging Spectroradiometer (MODIS) Irrigated Agriculture Dataset or MIrAD-US) involved inputs from county-level irrigated statistics from the U.S. Department of Agriculture, National Agricultural Statistics Service, agricultural land cover from the U.S. Geological Survey National Land Cover Database, and an annual peak vegetation index derived from expedited MODIS satellite imagery. This study investigated regional and periodic patterns in the amount of change in irrigated agriculture and linked gains and losses to proximal causes and consequences. While there was a 7% overall increase in irrigated area from 2002 to 2017, we found surprising variability by region and by 5-year map interval. Irrigation land use dynamics affect the environment, water use, and crop yields. Regionally, we found that the watersheds with the largest irrigation gains (based on percent of area) included the Missouri, Upper Mississippi, and Lower Mississippi watersheds. Conversely, the California and the Texas–Gulf watersheds experienced fairly consistent irrigation losses during these mapping intervals. Various drivers for irrigation dynamics included regional climate ﬂuctuations and drought events, demand for certain crops, government land or water policies, and economic incentives like crop pricing and land values. The MIrAD-US (Version 4) was assessed for accuracy using a variety of existing regionally based reference data. Accuracy ranged between 70% and 95%, depending on the region.


Introduction
Knowledge of the spatial and temporal dimensions of irrigated lands is useful for research, land/water management, and policy making. Understanding spatial and temporal trends of irrigated land can relate to the best agricultural practices and hence to better yields and better food supply. In short, consistent mapping of irrigated lands and understanding their dynamics can help in planning how to expand the food supply to the growing population, projected to reach 9.7 billion by 2050 [1].
Whereas croplands are dynamic and impractical to map at ground level because ground surveys are time consuming and costly, using remotely sensed imagery at various spatial and temporal resolutions to map and monitor rainfed or irrigated croplands is efficient and effective [2]. Various sources of satellite remote sensing contribute to mapping irrigated or rainfed crops. Moderate Resolution Imaging Spectroradiometer (MODIS) imagery with a temporal resolution of 1 day and spatial resolution of 250 m has been widely

Background
We utilized a geospatial modeling strategy to create a multi-period CONUS-wide dataset for four eras (2002, 2007, 2012, and 2017) at a 250 m spatial resolution. In 2010, at the first publication of the 2002 map [8], existing studies were limited to 250 m resolution using MODIS, and repeat mapping did not occur. However, since that time technology has advanced, and new methods have been devised so that collection of CONUS-wide data at finer spatial resolution with a higher repeat cycle is possible. Various studies applied different approaches and methods to map irrigated land at finer resolution at larger temporal scales. Though limited to smaller regions, for example, the HPA region [15] or the Western United States [16], these studies were extended to a three-decade time period, whereas the study by Xie et al. [17] covered a larger region at a finer resolution but was limited to a single year. MIrAD-US is a coarser resolution CONUS-wide dataset for four periods at intervals of 5 years.
USDA NASS reported an increase in irrigated area between 2002 and 2007, a decrease in irrigated area between 2007 and 2012, and an increase in irrigated area between 2012 and 2017 across the United States [18]. This statistic is available at 5-year intervals at county level only and does not provide the spatial information of the irrigated area. Researchers who are seeking robust and comprehensive methods to produce detailed geospatial maps for irrigated areas found remote sensing to be the most successful technique for mapping irrigated land [5,8,15,17]. The use of remote sensing enables a diverse range of optimal spatial and temporal observational capabilities for a given study area. Remote sensing has the flexibility to widen and narrow the spatial and temporal scope of the study area. Though irrigated areas are mapped at different scales and for various regions, the studies on change dynamics in U.S. irrigated land are still limited. Earlier studies explored the irrigation pattern at particular states and/or regions [5,8,15,17] or ecoregions, but a nationalscale watershed-level study of irrigation dynamics has not been conducted. This study also aimed to improve understanding of irrigation change dynamics across the United States at the level of major watersheds.
Watersheds provide a common geographic basis for understanding irrigated agriculture dynamics across the U.S. Watershed boundaries provide a seamless/standardized system where national hydrologic units are organized in a nested and hierarchical system using a unique code and represent the areal extent of surface water drainage to a single outlet [19]. The USGS Watershed Boundary Dataset (WBD) contains eight levels of progressive hydrological units identified by unique two-to 16-digit codes. We selected unique two-digit code hydrological units across CONUS at the most general level and collectively called them "watersheds" for this study. For further study, we selected 11 of these watersheds based on the dynamics and areal extent in irrigated areas between 2002 and 2017.
The two ways that water is supplied to cultivate crops are from direct rainfall and irrigation [20]. For irrigation, water is supplied either from surface water (via gravity flow from rivers, lakes, and reservoirs) or ground water (through springs or wells). While water availability is the most important factor for irrigation, other driving forces including types of irrigation, source of water, tariffs attached with water supply, the laws governing water-control [21], geographical location, land quality and physical attributes of the soil, and environmental parameters such as precipitation, temperature, and drought determine irrigation needs. Water supply for irrigation may be controlled at both local and regional levels by a state or an irrigation district [21]; therefore, it is crucial to understand the availability effects of these driving forces on irrigation. There are several studies that focus on evaluation of irrigation dynamics at local, state, and aquifer scales. While studies at watershed scale could help understand overall trend at a large scale providing more information on water availability (mostly surface water) surpassing the impacts of local variables, we found no studies at this scale. The aim of this study was to illuminate irrigation dynamics at a generalized watershed scale, which may be useful for understanding broader geographic patterns and developing regional water management strategies and hence combatting future food security threats.

Input Data/Geospatial Model
In developing an updated version of MIrAD-US (Version 4, [22]), we implemented the geospatial model developed by Pervez and Brown, using similar input sources but updating them to reflect current data availability [8] in order to map irrigated lands for four eras (2002, 2007, 2012, and 2017). Each MIrAD-US product corresponds with the year when USDA released their county-level irrigation area statistics data [18]. The other input data (eMODIS annual peak NDVI and NLCD) were selected from the same year or closest year, respectively. The guiding concept states (1) the irrigated crops show higher NDVI values than non-irrigated crops in the same local area; (2) both the irrigated and non-irrigated crops may exhibit similar NDVI values during optimum precipitation period (however, during drought the difference in NDVI between irrigated and non-irrigated crops may be maximized, and therefore, for this model to operate smoothly, ideally the annual time period with severe drought condition should be selected); and (3) the growing season peak NDVI, at any time it occurs, will vary for each crop and for each geographic region of the United States. Because the annual peak NDVI varies by crop, the model could be biased toward some of the peak NDVI values [8] as some crops tend to reach higher NDVI peaks than others. The best approaches to distinguish irrigated cells from the nonirrigated cells are either to compare with field data or compare wall-to-wall with United States crop maps [8]. A CONUS-wide detailed cropland specific dataset is not available for 4 mapping intervals, and producing field data is outside of our scope. Additionally, Pervez and Brown [8] investigated that whether peak NDVI would be a good indicator regardless of crop type at selected locations, and found it generally well supported. Further assessment of MIrAD-US by Wardlow and Callahan [23] of Nebraska irrigated agriculture also supported this finding.
In summary, the input data for our geospatial model include USDA county irrigation statistics, eMODIS annual peak NDVI, and land cover mask from NLCD [8]. First, 250 m eMODIS peak NDVI values were masked by resampled (250 m) NLCD data for agricultural land classes, then the peak NDVI were overlaid on county spatial data provided by the census. Then, a list of peak annual NDVI values was sorted in descending order. Starting with the highest ranked peak NDVI, the area of the corresponding peak NDVI value cells was computed and compared with the USDA irrigation area statistics for that county. The same process was repeated for each peak NDVI value in descending order until the accumulated area exceeded the county area estimate from the census [18]. The corresponding pixels that contribute to the matched area were identified as irrigated for the county. In a final post-processing step, all lone (single) pixels were spatially filtered from the irrigated area map.

Accuracy Assessment
We accessed the accuracy of MIrAD-US (Version 4) at different temporal and spatial extents depending upon the availability of reference data. Table 1 lists the available reference data, area coverage, corresponding MIrAD-US year, and sources. As no new reference data collection was carried out by this effort, reference data were obtained from multiple sources including state agency irrigation, well databases, and data collected by other researchers. Some of these data were collected via survey, whereas some were derived from secondary sources using Google Earth Engine and remote sensing interpretation techniques. We refer to this collective dataset as "reference" data.
The temporal and spatial extent of reference data were not consistent with MIrAD-US; therefore, we matched the reference data year closest to MIrAD-US era and carried out accuracy assessment at state or region levels: reference data for the Eastern Snake River Plain aquifer (ESPA) (in vector format) were obtained from the Idaho Department of Water Resources (IDWR). We converted the IDWR irrigated lands vector database to raster format matching the spatial resolution of MIrAD-2002 and assigned 1 or 0 for irrigated and non-irrigated class based on the irrigation status information. Other reference data for the HPA for multiple years [13,15] and Columbia Plateau basin-fill aquifer (CPA) for 2012 [17] were developed using a semi-autonomous training approach in Google Earth Engine.
We applied the two most common and effective methods for accuracy assessment: (1) simple random and (2) stratified random. In the simple random method, the surveyed polygon fields were converted into grids using an irrigation attribute (1 for irrigated and 0 for non-irrigated) matching the spatial properties (projection, extent, and cell size) of MIrAD-US. Then, the MIrAD-US and new raster maps were overlaid with each other on a cell-by-cell basis and were made to have the same attribute values (e.g., 1 if the cell is irrigated and 0 if the cell is not irrigated). Second, a set of sample random points were created using the random point generator tool in ArcGIS [24]. These random points were constrained by the common areas from both data products, and any two points were minimally spaced at 1 km. Third, the random points were overlaid on a pixel-by-pixel basis with MIrAD-US raster and new raster, and irrigation information was extracted from the cell that contained the point. In the stratified random method, the irrigated (1) and non-irrigated (0) points were generated from the surveyed polygons instead. These points were overlaid on a pixel-by-pixel basis with MIrAD-US raster and irrigation information extracted from the cell that contained the point. Finally, an error matrix was generated using the extracted irrigation information from MIrAD-US to compute the following accuracies and errors: overall accuracy, user's and producer's accuracies, error of commission and omission, and kappa statistics. Due to availability of surveyed polygons, stratified random methods of accuracy assessment were left out in the regions (e.g., CPA region, Table 1) with no polygon data.

The MIrAD-US Dataset
The MIrAD-US is a source of geospatial raster data for irrigated land across the U.S. for four years at 250 m and 1 km resolution. All the products are projected into Lambert azimuthal equal area, saved in ArcGIS-friendly Geotiff format, and found in ScienceBasean online portal maintained by USGS to store open data, repository services, and large datasets [22]. The MIrAD-US dataset has two classes: "irrigated" and "non-irrigated." Similar to the Pervez and Brown [8] (Figure 1a,c) show the majority of the lands (approx. 40%) were irrigated for 1 era followed by lands irrigated for 4 eras (approx. 26%). The South Atlantic-Gulf, Upper Mississippi, Great Lakes, Texas-Gulf, and a few parts of Missouri, Lower Mississippi, and Arkansas-White-Red watersheds were mostly irrigated for 1 era, suggesting unstable irrigation, whereas most of the lands in the Lower Mississippi, California, Missouri, and Pacific Northwest watersheds were irrigated for 4 eras, indicating a strong level of irrigation stability (Figure 1a). Lower precipitation rate and availability of water for irrigation can be associated with irrigation stability in these watersheds [25,26].   (Table 2). Between 2012 and 2017, irrigated area increased by 5.59% recording the highest gain in new irrigated areas among all mapping intervals. During this latest interval, approximately 67.38% of the irrigated areas remained unchanged, 32.62% was newly identified in 2017, and 33.37% was identified as lost from 2012 (Table 2).      California suffered from a prolonged drought and decreases in surface/ground water availability for irrigation throughout the study period leading to decreased irrigation, whereas the Missouri and Lower Mississippi watersheds still receive adequate ground water for irrigation. The increased number of ground water wells in Eastern Nebraska between 2002 and 2005 [8,15] and increased amount of rainfall in Lower Mississippi watershed from 1969 to 2017 [26] are driving factors for increased irrigation for these watersheds. Widespread drought in 2012 affected multiple watersheds like Missouri and California watersheds and resulted in an overall decrease in irrigation [27]. Between 2012 and 2017, increases in irrigated area in Missouri, Pacific Northwest, and Lower Mississippi watersheds offset the decrease in irrigation in the California watershed (Figure 2b).

MIrAD-US Product Accuracy
We collected reference data for accuracy assessment from multiple sources including state agencies, well databases, and data collected by other researchers (Table 1). Because these reference data were associated with different watersheds and time periods, the accuracy assessment task varied accordingly.
In California, we carried out the accuracy assessment for four eras and obtained overall accuracies above 95% ( Table 4). The overall accuracies above 95% indicate that there was good agreement between DWR surveyed polygons and irrigated areas estimated by MIrAD-US. However, a relatively high commission error of roughly above 30% for the irrigated class and low omission error of roughly 12% indicate that, in this region where ground data were collected, MIrAD-US may have mapped more irrigated fields. For 2012, both random and stratified methods showed overall accuracies above 95% and omission errors below 17%, suggesting that there was a good agreement between surveyed data, and the irrigated areas were not omitted in MIrAD-2012. In the HPA region (Table 4), the MIrAD-US agreed reasonably well with ground surveyed irrigated area information with an overall accuracy of 81.97%, 82.08%, and 80.15% at lower HPA (2008 and 2012) and Nebraska (2005). There was comparatively less agreement between MIrAD-US and ground surveyed irrigated area information with an overall accuracy of 62.02% in Kansas (2002) and 70.82% in Colorado (2005). The relatively high errors of omission (49%) and low errors of commission (16%) in HPA-CO indicate that there may have been fewer irrigated fields mapped by MIrAD-US. The differences in spatial detail between MIrAD-US and HPA-CO and the discrepancy between the MIrAD-US date (2016) and field survey dates (2017) likely resulted in high errors of omission and low errors of commission in this area. In the Western United States, the overall accuracy remained relatively lower in the ESPA region than CPA region. Using IDWR reference data, the ESPA assessment showed good agreement between the MIrAD-2002 and the reference area map with overall accuracies of about 79% and 66% using simple random and stratified random methods, respectively. In the CPA, the overall accuracy of 81% for MIrAD-2012 indicates that there was good agreement with Xie et al. [17] derived reference data. However, relatively high omission error of roughly 14% for the irrigated class indicates probable omission of irrigated areas in MIrAD-2012.

Discussion
Lack of consistent reference data across all of CONUS and during the same mapping eras for assessing the accuracy of the MIrAD-US maps limited the possibility of carrying out a single accuracy assessment across the entire CONUS; however, results of this study show MIrAD-US classifies irrigated area with a moderately high degree of accuracy. The overall accuracies across the assessed regions ranged from about 62% to 97%.
We assume that recent irrigation dynamics are mainly determined by the amount of available water supply for irrigation and crop types. In addition, other possible factors including economic incentives such as crop pricing and land values, the demand for corn related to development of biofuels, government policies related to water or land use, land availability, and regional climate fluctuations may influence irritation dynamics [5]. The trends we found in irrigation area vary across both temporal and spatial scales.  [5,28]; registration of new wells, particularly in Nebraska [5]; crop insurance and federal subsidies [29][30][31]; and shift from non-crops and/or wheat farming to corn farming due to increased corn prices as a result of high demands on ethanol production [5,18,[28][29][30][31][32][33]. Despite these drivers, the 2012 drought led to a decrease in irrigated areas, which may be associated with a scarcity of ground water limiting the number of fields that could be irrigated [13,15].
Watersheds that showed increasing irrigated areas include the Lower and Upper Mississippi watersheds. In the Lower Mississippi watershed, the irrigated area increased by 7.15% between 2002 and 2007, followed by 9.33% and 9.05% for following mapping intervals ( Table 4). The net increase in irrigated land between 2002 and 2017 was 27.75%. The increase in irrigated acres in the Lower Mississippi watershed may be associated with the types of irrigation methods, precipitation patterns, water availability, and corn prices during the study period. Most of the irrigated land in the Lower Mississippi watershed is located at the Mississippi Alluvial Plain (MAP). The MAP is widely known for row crops including rice, corn, soybeans, and fish farms [19,25,26,34]. Because rainfall events typically occur outside of the growing season, water for irrigation is supplied from groundwater or surface water [26,35,36].
A study by Yasarer et al. [26] described increasing precipitation in the Lower Mississippi watershed from 1969 to 2017, with an average increase of about 0.03 inch per decade, whereas the ground water elevation, streamflow, total cropland, and harvested cropland have declined. Despite a decrease in stream flow and total cropland area, multiple studies detected increases in irrigated cropland [26,35,37]. The increase in irrigated area is associated with adoption of efficient irrigation methods where the existing furrow irrigation system (that accounts for 75% of the irrigated area) is coupled with computerized timers, flow meters, irrigation scheduling, soil moisture sensors, precision leveling, and tail water recovery systems [26,35]. These irrigation methods reduce water use and prevent overdraft of the ground water supply [35]. In addition, adequate rainfall recharges the groundwater. In a study by Kebede et al. [35], they established a correlation between increased corn prices relating to corn-based ethanol production in the mid to late 2000s with increased irrigated cropland in the Mississippi watershed [7,38,39]. Even though this watershed seems to have enough water for irrigation at present, streamflow depletion has raised concerns for the future of available groundwater resources, which might lead to ecosystem/water stress [26,40,41].
Watersheds that showed decreasing irrigated area include the California and Texas-Gulf watersheds. The California watershed lost the highest acreage of irrigated area of 373,210 ha (10.59%) during the entire study period-the largest decrease (approx. 7.26%) occurred from 2002 to 2007 (Table 4). Much of the irrigation in California relies on surface water (i.e., gravity irrigation), and water supply originates from snowpack in the Sierra Nevada Mountains. A series of droughts from 2003 to 2016 (particularly 2002-2004, 2007-2009, and 2012-2016) coupled with unusual warm winters [1,19,34] led to decreases in snowpack in the mountains and to surface water shortage resulting in decreases in irrigated cropland [42,43]. Just as in California, depletion of groundwater levels also led to decreasing irrigated areas (−18%) in the Texas-Gulf watershed over the years [5,44].
While the mapping resolution of the MIrAD-US overcomes some spatial issues presented in subpixel fractional irrigated areas, it has trouble identifying irrigated fields that are smaller than a 250 m cell [8]. Fields that are smaller than 250 m are either classified as irrigated (if the majority of neighboring cells are irrigated), potentially leading to overestimation, or as non-irrigated land (if the majority of neighboring cells are non-irrigated), potentially leading to underestimation. MIrAD-US provides an estimate rather than an exact rendering of irrigated lands. Use of higher resolution images such as 30 m Landsat or 10 m Sentinel-2 might overcome the spatial issues [15][16][17]. The MIrAD approach identified irrigation status, but a change in irrigation status does not necessarily identify a change (presence or absence) in irrigation equipment [5]. Different approaches would be needed to detect irrigation equipment, which could be explored in further research.
Higher resolution (e.g., 30 m) annual irrigation mapping may be more useful in exploring the dynamics of U.S. irrigated agriculture as these higher resolutions can provide precise field locations of smaller fields with higher accuracy [15][16][17]45]; however, because the majority of irrigated fields in the United States are generally larger than a single pixel (250 m × 250 m = 62,500 m 2 ), the 250 m MIrAD-US data are still good for detailed geospatial information of irrigated areas across the conterminous United States [4,5]. In addition, because MIrAD-US provides consistent and accurate information of irrigated area across CONUS for 15 years at 5-year intervals, these datasets are important for estimating irrigated croplands [5,8].

Conclusions
This study investigated both regional and periodic increases and decreases in irrigated agriculture across the lower 48 U.S. states and more specifically at the watershed level. While there was a 7% overall increase in irrigated area across CONUS from 2002 to 2017, we found surprising variability by watershed and by 5-year map interval. Although the Missouri, California, and Lower Mississippi watersheds showed the most stable irrigation patterns via irrigation frequency maps, they depicted different trends. The California watershed showed a general decrease in irrigated area, the Lower Mississippi watershed showed increases over time, and the Missouri watershed showed an increase with a break in 2012. A series of drought events in California led to surface water shortage resulting in a decrease in irrigated cropland, whereas abundant rainfall and water availability led to increase in irrigated cropland in the Lower and Upper Mississippi watersheds.
Accuracy of the MIrAD-US (Version 4) ranged between 62% and 97%, depending on the region. The accuracies vary because of the irrigated field size, the density or homogeneity of irrigated lands within the agricultural setting, and the reference data used. In California where irrigated field sizes were fairly large and homogenous and reference data were collected by field survey, agreements as high as 97% occurred between reference data and MIrAD-US. Alternatively, in the CPA region where the reference data were developed using a semi-autonomous training approach in Google Earth Engine, the overall accuracy was comparatively lower (81%). Additionally, low numbers of irrigated samples in reference data affected the overall accuracy by stratified methods. The overall accuracies (stratified) varied at different geographical locations for the same era of data; for example, for MIrAD-2012, the overall accuracy for California is about 97%; HPA-TX, NM, OK is about 82%; and CPA is 81%. Climatic variance could have played a role in irrigation accuracy variability, but in order to perform a quantitative analysis of irrigation accuracy during different climate conditions, we would need reference data that were collected to represent larger variability in climate. In the most cases, we have reference data for only a single year. Future research could address this topic by providing access to multi-temporal reference data for irrigated agriculture while considering climatic condition as a variable for data collection.
We recognize that important advances have recently occurred in mapping and monitoring irrigated lands at higher spatial resolutions (e.g., 30 m) through time. Two studies by Deines et al. [15] and Ketchum et al. [16] are notable in that they offer an annual mapping periodicity at 30 m spatial resolution, but for smaller regions. Advancing these types of efforts to operational annual mapping across CONUS would be the next logical step. In addition, the collection of comprehensive reference data that represents high frequency and geographic coverage would also be a great benefit for robust accuracy assessment.