Cropland abandonment between 1986 and 2018 across the United States: spatiotemporal patterns and current land uses

Knowing where and when croplands have been abandoned or otherwise removed from cultivation is fundamental to evaluating future uses of these areas, e.g. as sites for ecological restoration, recultivation, bioenergy production, or other uses. However, large uncertainties remain about the location and time of cropland abandonment and how this process and the availability of associated lands vary spatially and temporally across the United States. Here, we present a nationwide, 30 m resolution map of croplands abandoned throughout the period of 1986–2018 for the conterminous United States (CONUS). We mapped the location and time of abandonment from annual cropland layers we created in Google Earth Engine from 30 m resolution Landsat imagery using an automated classification method and training data from the U.S. Department of Agriculture Cropland Data Layer. Our abandonment map has overall accuracies of 0.91 and 0.65 for the location and time of abandonment, respectively. From 1986 to 2018, 12.3 (±2.87) million hectares (Mha) of croplands were abandoned across CONUS, with areas of greatest change over the Ogallala Aquifer, the southern Mississippi Alluvial Plain, the Atlantic Coast, North Dakota, northern Montana, and eastern Washington state. The average annual nationwide abandoned area across our study period was 0.51 Mha per year. Annual abandonment peaked between 1997 and 1999 at a rate of 0.63 Mha year−1, followed by a continuous decrease to 0.41 Mha year−1 in 2009–2011. Among the abandoned croplands, 53% (6.5 Mha) changed to grassland and pasture, 18.6% (2.28 Mha) to shrubland and forest, 8.4% (1.03 Mha) to wetlands, and 4.6% (0.56 Mha) to non-vegetated lands. Of the areas that we mapped as abandoned, 19.6% (2.41 Mha) were enrolled in the Conservation Reserve Program as of 2020. Our new map highlights the long-term dynamic nature of agricultural land use and its relation to various competitive pressures and land use policies in the United States.


Introduction
Cropland abandonment is a widespread land use change process whereby croplands are no longer farmed due to socioeconomic changes or environmental conditions [1][2][3][4].While often economically or biophysically marginal for food production [5], abandoned croplands may nevertheless provide other ecosystem services like carbon sequestration and climate regulation [6,7], water and soil protection [8], and ecological conservation [9][10][11].Abandoned croplands are also frequently considered as potential sites to produce cellulosic bioenergy crops [12], which could help reduce pressure to expand energy crop production to active farmland [13][14][15].All these functions, however, depend on the specific location of the abandoned fields, duration of abandonment, and land use history [16].Knowing exactly where and when cropland abandonment occurs is vital to fully assess its environmental effects, future production potential [17], and the tradeoffs associated with alternative uses of these formerly cropped lands.
Getting field-level information on the location and time of cropland abandonment across broad areas (e.g.countries or even globally) is challenging because it requires long-term, frequent, and reliable high-resolution cropland maps, which are commonly lacking.As a result, nationwide fine-scale maps of cropland abandonment over periods of more than roughly 10 years remain unavailable even for countries such as the U.S. Accordingly, estimates of cropland abandonment for the U.S. differ widely among studies because they mostly rely on maps with coarse spatial resolution (from kilometers to hundreds of kilometers) or with limited temporal frequency and coverage (single [or a few] years to decades) [13,[18][19][20].Recent finer spatial resolution cropland maps, such as Global Food Security-Support Analysis Data [21] and Landsat-based maps [22], have become available but lack temporal frequency.This shortcoming, together with discrepancies in cropland definitions among maps, makes map fusion for field-level abandonment mapping challenging.
Compared to global maps, US-specific land cover maps like the U.S. Department of Agriculture (USDA) Cropland Data Layer (CDL), the USGS National Land Cover Dataset, and the Land Change Monitoring, Assessment, and Projection have greater spatial and temporal resolution and higher accuracy.The CDL product is particularly well suited for mapping cropland abandonment processes due to its cropland-specific nature, annual update frequency, field-scale resolution, and high accuracy due to its exclusive access to extensive high-quality training data from several USDA agencies [23,24].However, CDL coverage for the entire U.S. is available only since 2008.This limited temporal extent precludes mapping applications that require cropland cultivation status prior to 2008, such as identification of longterm cropland abandonment and expansion, croppasture rotation patterns, or trends in cultivation and fallowing over time.
There exist several global studies that use medium resolution satellite imageries (e.g.Landsat time series) to track the spatial and temporal dynamics of cropland abandonment starting in the 1990s.However, most of these are case studies with limited spatial coverage, such as select areas in Kazakhstan [25], the Caucasus region [26], and some other representative countries [27].Several factors have constrained the use of Landsat imagery for broad scale abandonment mapping; one is the lack of ground reference data to train machine learning classifiers.While the CDL's capability of directly mapping abandonment is limited before 2008, it has substantial potential to provide consistent and representative training samples covering the entire U.S [28].This opportunity makes frequent nationwide cropland mapping more practical and could enable efficient mapping of abandonment across the U.S.
Our goal here is to derive a nationwide, 30 m resolution map of cropland abandonment between 1986 and 2018 for the conterminous United States (CONUS).We generated annual cropland maps using an automated classification with CDL-derived training samples and the entire Landsat image archive.Based on these maps, we identified spatial and temporal concentrations of abandonment over the past three decades as well as the current uses of abandoned croplands.Note that our examination here focused solely on cropland loss resulting from abandonment, rather than on overall net dynamics.

Methods
We mapped the location and year of cropland abandonment across CONUS at 30 m resolution, which provides sufficient detail to delineate individual fields and growing seasons in most parts of the U.S. First, we generated a training sample pool from CDL and Landsat imagery.Second, we combined the training samples with Landsat composites used to generate state-and county-stratified classifiers each year.These classifiers were later used to map cropland annually from 1986 to 2018.Third, we identified the location and time of abandonment using a moving temporal window for each pixel (figure 1).Fourth, we estimated the accuracy of the location and time of abandonment based on visually interpreted sample locations and publicly available datasets.Each of these steps is detailed below.

Creating annual maps of cropland probability
We defined active croplands as any area planted to cultivated row, closely grown, or horticultural crops [18].Note that this definition does not include fields with grasses grown for pasture and hay.The Landsat imagery was used to track cropland distribution back to 1980s due to its long data archive and 30 m spatial resolution that enables us to characterize U.S. croplands at the field level.
To address the challenge of collecting nationwide training samples, we used CDL in years 2009, 2011, 2012, and 2015 as training data.We selected these four years to include a range of environmental conditions (e.g. both drought and excess moisture) and achieve a more robust universal classifier while minimizing the number of inputs [28].We relabeled CDL of these four years as 'crop' (seven subclasses: corn and soybean combined, wheat and barley combined, potato, alfalfa, cotton, rice, and other-crops) and 'non-crop' (eight subclasses: water, developed, barren, forest, shrub, herbaceous, wetland, and fallow/idle cropland) based on [29] and [18].Training samples of each subclass were extracted separately in each year by using the subclass's area as weight (one sample per 5 km 2 ).These samples were then aggregated to 'crop' and 'non-crop' for each corresponding year (i.e.2009, 2011, 2012, and 2015).
For each target year, we used all Landsat images within the year and those before and after the year to calculate input features for cropland classification.We included Landsat images before and after the target year to overcome limited availability of clear observations in locations where cloud and cloud shadows are frequent.Input features were then calculated as 10, 25, 50, 75, and 90 percentiles of red, green, blue, near infrared, and shortwave infrared bands, and normalized difference vegetation (NDVI), water (NDWI), and bareness (NDBI) index.We used percentile composites as inputs due to their proved effectiveness of aggregating possible irregular time series (across space) into consistent features and good performance in cropland classification [27,30].
These four years of training samples and associated input features were then combined to train a generalized random forest classifier for each county (state-wise classifier for counties without enough training samples) (see details in [30]).The trained rules were applied to annual feature stacks from 1986 to 2018 to create preliminary annual cropland maps.When assessed based on the CDL 2008-2018, the preliminary cropland maps had an average overall accuracy of 90% [30].Based on these preliminary maps, stable crop and non-crop areas were identified as those that remained unchanged though time.The stable areas were then used to generate final training samples, train county-and state-stratified classifer per year, and create annual 30 m maps of cropland probability for each county across CONUS.

Identifying cropland abandonment
We adopted a moving window strategy to identify the location and time of cropland abandonment [25,26].The time of abandonment was identified as the first year when: (1) average cultivation probability was >50% for the 5 year window before abandonment and <50% for the following 5 years (including the abandonment year), (2) there was at most 1 individual year with <50% cultivation probability during the 5 years prior to abandonment and at most 1 individual year with >50% cultivation probability during the 5 years after, (3) the pixel remained uncultivated once abandoned (i.e.probability ⩽50% for all five-year windows after the last abandonment time), and (4) not urban in 2014 as reflected by CDL.We used a five-year window in accordance to Food and Agriculture Organization's definition of temporarily fallow and because short-term fallow periods are common in CONUS and using a narrower window would incorrectly map those areas as abandoned [25][26][27]31].For locations where multiple abandonment events were detected, we only recorded the most recent date of abandonment.As a result, the first and last year of abandonment detected in this study was 1991 and 2014, respectively.We excluded cropland urbanization as abandonment given the nature of urban land uses-i.e.once developed, they are very unlikely to be recultivated [13,15,32,33].Lastly, a minimum mapping unit of 23 Landsat pixels (∼5 acres) was applied to remove possible incorrectly classified individual pixels [18].

Accuracy assessment
We first evaluated the spatial accuracy of the areas we mapped as abandoned.For this, we randomly selected 400 pixels from our map (200 for abandonment and 200 for stable croplands, including cultivated fallow).Second, we evaluated the accuracy of the time of abandonment using another, separate set of samples.To do so, we aggregated abandonment years into 3 year intervals (i.e.abandonment that occurred from 1991 to 1993 was aggregated to the class '1991-1993') [27] and randomly selected 60 pixels for each aggregated abandonment class (in total 480 samples).We labeled these pixels manually based on visual interpretation of time-series of very high-resolution satellite images on Google Earth and Landsat imagery on Google Earth Engine.Based on these samples, we calculated producer's accuracy (PA), user's accuracy (UA), overall accuracy (OA), and the F1 score (F1 = 2 × UA × PA/(UA + PA)) [34].
We also compared our estimates with others published in [9,13,15,18,20,33] at the nationwide scale.Given inconsistency of analysis periods across studies, we calculated equivalent areas based on our map.To do so, for a given period (e.g.1980-2012), we used our map's annual estimates for overlapping years (i.e.any year from 1991 to 2014) and our map's minimum and average yearly abandonment for the years beyond 1991-2014.Yearly estimates were finally aggregated and compared with published estimates.

Spatial and temporal patterns of cropland abandonment
Our map provides 30 m resolution information of the location and time of croplands abandoned during the period of 1986-2018 (figure 2).Collectively, CONUS experienced 12.3 (±2.87) million hectares (Mha) of gross cropland abandoned during the past three decades, or 7.3% of the total cropland area present in 1986.Abandonment was especially widespread in the central and southern Ogallala Aquifer region, southern Mississippi Alluvial Plain, and the Northern Great Plains.At the state level, Texas had the largest area of abandonment of 1.88 ± 0.44 Mha, followed by North Dakota (0.86 ± 0.21 Mha), Kansas (0.71 ± 0.17 Mha), Montana (0.68 ± 0.16 Mha), South Dakota (0.62 ± 0.15 Mha), and Oklahoma (0.59 ± 0.14 Mha) (figures 3(a) and table A1).The distribution of abandoned croplands also varied greatly within states and regions (figure 3(b)), with counties of over 20 thousand abandoned hectares located across northern and southern Texas, western Oklahoma, eastern Louisiana, northeastern Colorado, southeastern Washington, northern Montana and South Dakota, and throughout North Dakota.
Among 3108 counties (or county equivalents), 30% (i.e.941 counties) experienced abandonment of more than 10% of their total cropland area since 1990 and 166 of them had >20% abandoned (figure 4).Counties with high proportion of abandonment occurred most prominently throughout the South and Southeast.In more intensively cultivated areas of the U.S. such as the Midwest, northern Colorado, southeastern Washington, and northern Montana, most counties saw less than 10% of their total croplands abandoned.
Temporally, we found an average rate of abandonment of 0.51 Mha yr −1 during the study period, but this varied greatly among years (figure 5).At the nationwide level, the first six reporting years (1991-1996) experienced a deceleration in annual abandonment, followed by a rapid increase from 0.38 Mha yr −1 in 1994-1996 to 0.63 Mha yr −1 in 1997-1999 (0.25 Mha yr −1 ).Since then, annual cropland abandonment decreased to 0.41 Mha yr −1 in 2009-2011.However, annual abandonment area increased in more recent years (2012-2014) to 0.52 Mha yr −1 .Nationwide, abandonment proportions (i.e. the percent of abandonment area relative to total crop extent) peaked between 1997 and 2002 with 1.3% of total croplands abandoned annually (figure 6).Consistent with the nationwide trend of abandonment area, the relative rates of abandonment were lowest during 1994-1996 (0.8%) and remained decreasing from 2001 to 2011.However, these trends varied across states (figures 5 and 6).

Current land cover/uses of abandoned croplands
We used CDL 2020 as a current land use reference and calculated the area of forest, shrubland, grassland/pasture, wetlands, non-vegetated lands, and urban that were originated from croplands.In addition, we calculated the proportion of abandoned land that was enrolled in a Conservation Reserve Program (CRP) contract in 2020 [35].The most common fate of abandoned croplands during the study period  was conversion to grassland/pasture (including nonalfalfa hay, herbs, and switchgrass) in 2020, accounting for 53% of the total (i.e.6.5 Mha) (figures 7 and 8).An additional 12.2% of abandoned croplands were converted to shrubland (1.5 Mha) and 4.6% to nonvegetated lands (0.56 Mha, including barren lands and fallow/idle croplands in CDL).Wetlands, which include woody and herbaceous vegetation, occupied another 8.4% of the total abandoned area.
We also found that 19.6% (2.41 Mha) of the abandoned croplands were enrolled in CRP contracts in 2020, 56% of which classified as grassland/pasture (table 1).Given a total CRP area of 8.2 Mha in 2020, our map suggests that only 29.4% of 2020 CRP lands met our criteria for permanently abandoned lands-i.e. were persistently, intensively cultivated during the full duration of our study period prior to their enrollment.Note that this low proportion likely stems in part from our cropland definition (i.e.excluding planted pasture and non-alfalfa hay) and our stringent definition of abandonment, especially the minimum mapping unit of 5 acres and our restricting of abandonment to only locations with a one-time, one-directional change from cultivation to non-cultivation, thus excluding marginal areas that frequently move into and out of production.These rules thus excluded most CRP lands, which may have been intermittently cultivated or in planted pasture or non-alfalfa hay before program enrollment.As such, our results instead identify only abandoned areas with more enduring shifts in land use greater than 5 acres.
Spatially, former croplands largely returned to grassland/pasture in the Midwest and Great Plains, with shrubland cover more common on abandoned croplands further west (figure 7).Cropland conversion to forest was exclusively found in the East, concentrated in the Atlantic Coastal region.Conversion of croplands to wetlands was predominantly located in the Mississippi Alluvial (Arkansas, Mississippi, and Louisiana) and the Upper Midwest (Minnesota, North Dakota, and Wisconsin) regions.Cropland conversion to non-vegetated lands were present mostly in the West, the Upper Midwest, and the Mississippi Alluvial Plain.Normalizing the area of each land use by total abandonment helps further resolve these trends (figure 8).In summary, most abandoned croplands in the U.S. transition to grassland/pasture, except for the Atlantic Coast, the Mississippi Alluvial Plain and the Upper Midwest, and the arid Southwest, where abandoned croplands most often transition to forest, wetlands, and shrubland, respectively.
Urban development is another important cause of cropland loss (figure 9).Unlike other land uses, cropland losses due to urban development were widely spread across the country, especially for metropolitan areas that are surrounded by intensively cultivated croplands, such as Chicago-Milwaukee-Madison, IL-WI; Indianapolis, IN; Minneapolis, MN; Detroit, MI; St. Louis, MO-IL; Columbus, OH; Atlanta, GA; Denver-Fort Collins, CO; Salt Lake City-Logan, UT; Phoenix, AZ; and Dallas-Fort Worth, TX.

Mapping accuracies
Our map has an overall accuracy of 0.91 and F1 Score of 0.9 for abandonment vs. non-abandonment (table 2), indicating good overall detection of cropland abandonment during the past three decades.The user's accuracy (1-commission error) of the abandonment (0.84) suggests that our map reliably represents abandoned croplands and does not often mistakenly label stable croplands as abandoned.In contrast, the producer's accuracy of the abandonment class reaches 0.98, suggesting that our map correctly captures nearly all actual abandonment extent on the ground.
Our map also provides reasonable estimates of the time when a cropland field was abandoned, with the overall temporal accuracy and average F1 score of 0.65 (±0.06) and 0.65, respectively (table 3).The accuracy of mapped abandonment time varies across years, with F1 Score ranging from 0.52 for '2009-2011' to 0.80 for '1991-1993' .Our map of the year of abandonment also has similar User's and Producer's accuracies, which is an indication of balanced commission and omission errors.

Spatiotemporal patterns and potential driving forces
Our map identified several local to regional concentrations of cropland abandonment during the past three decades.While identifying the specific reasons for cropland abandonment is beyond the focus       of this study, our findings point to several probable causes.For example, we found notable conversion of croplands to grassland/pasture in the central and southern Ogallala Aquifer, where previous studies suggest that groundwater depletion has contributed to localized irrigation losses for the past several decades [30,36,37].The relationship between abandoned croplands and formerly irrigated lands suggests a considerable amount of formerly irrigated cropland that has either been converted to rainfed cropping, abandoned from cultivation altogether, or converted to grassland/pasture under CRP.Abandonment in conjunction with CRP enrollment is also widespread.In consistent with CRP enrollment [38,39], a large share of our mapped abandonment spans from Texas to Montana across the Great Plains, where formerly cropped lands have been restored to grassland/pasture.Smaller concentrations are found in eastern Washington, the Mississippi Alluvial Plain, and southeastern Idaho.
Agricultural intensification and industrialization may be additional contributing factors.There is evidence indicating that more industrialized agriculture could lead to a spatial agglomeration of cropland while simultaneously reducing the overall need of cropland (for example, the overall decline in cropland area from 1986 to 2002 as observed in this study).Consequently, this dynamic contributes to the abandonment of croplands that are vulnerable to environmental and climate change, such as the Great Plains where rainfall is limited and much of the land is subject to potentially severe wind erosion [40].
We also found evidence of urbanization-induced cropland abandonment especially in the peri-to sub-urban areas of metropolises.Urban development not only converted croplands directly to residential houses, road networks, and other urban infrastructure, but it also stimulated cropland abandonment close to newly built urban areas.This peridevelopment abandonment may stem from increased land and operational costs of farming near cities, loss of agricultural viability, and/or the necessity of conserving environmental and recreational benefits of urban areas [41][42][43][44].Furthermore, it is not yet clear what proportion of peri-urban abandonment is a transitory state prior to that lands' further development versus a persistent abandonment to an alternative undeveloped use (e.g.protected open space).
The observed trends in abandonment over time are also part of a broader landscape dynamic in cropland use.In this work, we focused on the gross, one-directional accumulation of relatively permanent abandoned croplands over time, but this process is occurring amongst a backdrop of lands moving both into and out of crop production.For example, our annual cultivated land maps suggest that total cropland area primarily declined from 1986 to 2002 before beginning to rise in total area through the end of the study period.Thus, the cropland abandonment we identify here is expected to persist regardless of whether net cropland expands or contracts in any given year.

Advantages and uncertainties of mapping approach
Our abandonment map was created using a timeseries analytic method, which provides several advantages over the common previous approach of comparing a limited number of thematic maps over time [13,15].Because cropland fallow (either short-or long-term) is frequent in the U.S., abandonment maps generated by comparing a limited number of time points might falsely identify intermittently cultivated lands as abandonment; mapping accuracies of such products are also likely to be sensitive to the selection of base map and year.In contrast, our approach provides a more robust estimate of abandonment that is less dependent upon analysis start date or observational frequency.Additionally, our map was created using an automated method without the necessity of labor-intensive training sample collection [4,25,27].It is feasible and remains efficient to manually collect training samples for small study areas, but preparing a complete and consistent training dataset for an area as large as CONUS is challenging.This characteristic makes our map easy to update if more years are to be considered (e.g. after 2018).Just as importantly, this feature makes our method capable of mapping cropland expansion and further facilitate estimation of net change across CONUS (like studies in [9,18]), although this study only presented cropland abandonment.

Y Xie et al
Compared to existing maps of abandoned cropland area, ours is the first to cover the entire CONUS while also providing field-relevant resolution in spatial detail and the year of abandonment.Despite this, uncertainties remain.Firstly, our map is expected to miss abandonments of planted pasture and hay lands given that our cropland definition explicitly excluded them.Pasture and hay specific maps would be required to identify these abandonments, and this mapping is an area of ongoing and potential future work.Second, we defined cropland abandonment using a five-year window [31].However, some of the areas that were abandoned in recent years (e.g.2010-2014) may be recultivated in the near future.For example, we found 15.4% of our mapped abandoned locations were classified as 'others' on CDL 2020, including recultivated croplands (table 3) or newly planted perennial crops that might need several years to be detectable on medium resolution satellite images.As such, it might be necessary to use more recent cropland layers to remove those mapped abandoned locations that are recultivated.Cropland layers can be annually updated with USDA CDL or ones created using our developed methods.Lastly, our 30 m resolution map is capable of capturing field-scale cropland abandonment.However, given this spatial resolution and the minimum mapping unit we applied (5 acres), we likely missed sub-field abandonment events.One possible way to improve this would be to incorporate higher spatial resolution data such as aerial imagery acquired by the USDA National Agriculture Imagery Program (submeter to meters resolution) and Sentinel imagery (10 m resolution) for more recent years.

Applications of the map for bioenergy and land-based climate mitigation estimation
Many studies consider abandoned croplands as potential sites to implement land-based climate mitigation strategies, for example via the growth of bioenergy feedstocks [13][14][15], natural vegetation restoration or regeneration [45], or both [46].Additional uses such as solar photovoltaic development, livestock forage provisioning via haying or grazing, and recultivation as cropland represent further alternatives for which abandoned croplands may be evaluated.Our spatially explicit maps of abandoned cropland locations can help researchers gain a better understanding of these lands' true characteristics, enable greater insights and further study into their availability, and allow evaluation of the tradeoffs of competing uses.
Robertson et al [46] included 41 (36)(37)(38)(39)(40)(41)(42)(43)(44)(45)(46) Mha of abandoned agricultural lands in their assessment of U.S. marginal land bioenergy potentials, based on assessments of historical county-level records [47] and remote sensing analyses [48] that suggest ∼70-100 Mha of U.S. agricultural lands have been abandoned since the 1930s.Our analysis suggests that relatively little of this abandonment (12.3 Mha, table 1) has happened since 1986.Using only the portion of this acreage classified as shrubland and grass/pasture (8 Mha) to avoid urban and carbondense forests and wetlands allows us to calculate its potential for negative carbon emissions.Assuming 3.1-4.2t CO2e ha −1 y −1 of sequestration potential for these lands depending on soil carbon pools when energy production commences [46], we estimate a theoretical annual sequestration rate of 19.2-26 Mt CO2e y −1 from abandoned lands not enrolled in the CRP.Over a 70 year period to 2100 beginning in 2030, this would total 1.3-1.8Gt CO2e.This represents ∼1.4% of the ∼110 (57-178) Gt CO2e that Robertson et al [46] identified as the potential U.S. contribution to land-based negative emissions.
Beyond refining broad-scale estimates of the climate mitigation potential or other outcomes from alternative uses of abandoned cropland, our maps can also enable a more nuanced understanding of such lands' availability, support further study of the socioeconomic feasibility of proposed uses, and help gain insights into landowners' willingness to adopt alternative uses.For example, our spatially explicit maps could be combined with property boundary data to identify former croplands which remain on farms active in agricultural production.Such locations may be the most likely to consider transitioning abandoned cropland to growing forage or cellulosic bioenergy feedstock.In contrast, abandoned croplands in locations no longer involved in agricultural production or in regions without active agricultural infrastructure could be less likely to adopt new agricultural practices, and therefore be more likely or suited to continue in non-agricultural uses like recreation, nature-protection, or solar energy development.As climate change and increasing demands for food and energy continue to add pressure on limited land resources, it will become increasingly important to fully and realistically consider the tradeoffs and potential of competing land uses.

Comparison with existing maps
Our estimate of total abandonment is relatively comparable to that from [13] and [18].Our estimate of 12.3 Mha is reasonably consistent to the 15 Mha estimated in [13], although the latter excludes conversion to pastureland and covers a longer time (34 years) compared to ours (24 years).Our estimates (15.5-17.7 Mha) are even more consistent with that reported by [13] if we assume the years since 1991-2014 have a similar annual abandonment increase as captured in our map.When aggregating croplands converted during 2008-2012 from our map, we found 2.15 Mha of total abandonment, compared e Estimate for the period 1991-2014.Note that all estimates, except for Lark et al [9,18], exclude urbanization as abandonment.
to 1.76 Mha in [18].Interestingly, using the same approach, Lark et al identified an even smaller abandonment area (19% less) for the period of 2008-2016 compared to their previous estimate for 2008-2012 [9].Their finding reflects the effect of identifying abandonment using longer-term, frequently updated cropland maps, which better exclude intermittently cultivated (fallow or rotation) and recultivated croplands.Consistent with this effect, our mapping here likely represents the most conservative and refined estimate of abandoned cropland area, due to its consideration of 33 independent years of observationmore than any other map reviewed.
Except for [13] and [18], our estimate differs significantly from some existing studies (table 4), for several reasons.First, we used a conservative 5 year window for defining abandonment, rather than inferring abandonment by comparing cropland maps at two different times.Identifying abandonment using a 2 year differencing approach is particularly sensitive to mapping errors.For example, exaggerated mapping of cropland in any particular year could possibly result in overestimations in [13,20], and [33].Similarly, errors in either the 1992 or 2015 cropland maps used in [15] could lead to substantial uncertainty of the final abandonment map.Another factor for disagreement of estimates is the period of analysis covered in different studies.Our map identifies abandonment between 1986-2018 (the actual abandonment years cover only 24 years, 1991-2014), whereas some of existing studies cover longer periods back to 1850 [20,33] or only since 2008 [9,18].Finally, diverse data sources contribute to divergent results.Unlike [13,20], and [33] which rely on fusion of survey data and remote sensing based maps, our estimate is based exclusively on 30 m resolution cropland layers consistent through time.
Spatially, our results highlight similar hotspots of abandonment in the southern CONUS (e.g. the central to southern Ogallala Aquifer, the lower Mississippi Alluvial Plain, and the Atlantic Coast), which is consistent with other studies [13,15,20,33].However, our map also reveals more regions with widespread cropland abandonment, such as the Northern Great Plains and Washington state.Just as importantly, our map provides another feature that is unique from currently available maps for CONUS-the time of abandonment, a variable that is not well characterized but often critical for assessing the environmental outcomes of cropland abandonment [6,8,10,16].

Conclusions
We present a nationwide assessment of cropland abandonment between 1986 and 2018 for the CONUS.In total, we identified 12.3 (±2.87)Mha of croplands abandoned during the study period.Spatial hotspots of cropland abandonment during the study period were found in the central and southern Ogallala Aquifer, the Northern Great Plains, and the Mississippi Alluvial Valley.Temporally, the period of 1996-1999 experienced the most rapid increase in abandonment.Among these abandoned croplands, over half are currently grassland/pasture.Our map with 30 m spatial details may be valuable to assess environmental effects of cropping activities and explore potential uses of these lands from local to regional scales.

Figure 1 .
Figure 1.Flowchart of our approach to map cropland abandonment.

Figure 2 .
Figure 2. Spatial patterns of cropland abandonment in CONUS, 1986-2018.The overview map shows proportion of abandoned croplands within a 6 km × 6 km grid (i.e.abandonment area/36 km 2 × 100).The local insets highlight a 30 km × 30 km region and the years of abandonment within a 6 km × 6 km grid.The satellite images highlight a cropland field abandoned around 2006 in the Mississippi Alluvial Plain.As evidenced by very high-resolution images, the field was cultivated in 2005, but was grassland and shrubland in 2007 and 2012, respectively.

Figure 3 .
Figure 3. Area of abandoned croplands by state (a) and county (b).

Figure 4 .
Figure 4. Abandoned croplands in proportion to the total cropland area in 1986-1990 (a) and the county area (b).

Figure 5 .
Figure 5.Cropland abandonment trends by state.The abandonment area is aggregated into 3-year intervals and each time point represents the middle year (e.g.1992 for the years 1991-1993).

Figure 6 .
Figure 6.Trends of abandonment rates by state.The abandonment area is aggregated by 3-year intervals and each time point represents the middle year (e.g.1992 for the years 1991-1993).

Figure 7 .
Figure 7. Land cover of abandoned croplands in 2020.The numbers show the area of each land use/cover.

Figure 8 .
Figure 8.Per-county proportion of current land covers (in 2020) of a county's abandoned croplands.The numbers show the proportion of each land cover across the conterminous U.S.

Figure 9 .
Figure 9. Cropland loss to urban development.The overview map shows the proportion of cropland conversion to urban within a 6 km × 6 km grid.The local views are the Chicago (a), Dallas-Fort Worth (b), and Phoenix (c) metropolitan areas.

Table 1 .
Current land cover of abandoned croplands (in Mha) according to the USDA CDL 2020.Grassland/pasture here includes non-alfalfa hay, herbs, and switchgrass; non-vegetated lands include barren lands and fallow/idle croplands; all other resulting land uses/covers are included in 'Others' .CRP: Conservation Reserve Program.Non-vegetated lands Forest Shrubland Grass/pasture Wetlands Others Total

Table 2 .
Mapping accuracy of abandonment vs. non-abandonment (location consistency between our map and reference).

Table 3 .
Accuracy of mapped abandonment time.

Table 4 .
Different estimates of cropland abandonment for CONUS.
a Estimate for North America; all others are for the United States.b Estimate for the period 2006-2014.c Estimate by assuming annual abandonment area of 0.34 Mha (the minimum yearly increase in our map) for the years not covered by our study.d Estimate by assuming annual abandonment area of 0.51 Mha (the average yearly increase in our map) for the years not covered by our study.