Introduction

Agroecosystems experience spatial and temporal land-cover heterogeneity (Goslee 2020; Jeanneret et al. 2021). Categorizing such heterogeneity is key to elucidating pattern-process relationships within agroecosystems (Tscharntke et al. 2005; Hopfenmüller et al. 2014; Jeanneret et al. 2021). However, it is unclear which aspects of landscape pattern (composition, configuration, connectivity, or context) are most modified within agroecosystems. (Composition refers to what kinds of land-cover types are present, configuration is how patches of land-cover types are arranged, connectivity assesses potential isolation of land-cover patches of a given type, and context refers to the overall setting in terms of homogeneity/heterogeneity and stability/instability [persistence] over time.) These attributes are poorly understood because the spatial properties of agroecosystems are seldom measured (Fahrig et al. 2011; Hopfenmüller et al. 2014) or are not examined over time (Corry 2019; Coffin et al. 2021). Which aspect of landscape pattern change is most pertinent to support sustainable agriculture remains uncertain (Landis 2017; Tscharntke et al. 2021).

In the US, one important driver of agricultural change is the Farm Bill, a blanket term for federal legislation enacted every 4–6 years (most recently in 2008, 2014, and 2018) that provides financial support to produce certain crops, which serves to stabilize commodity prices. Additionally, Farm Bills support programs covering nutrition assistance, conservation, and other topics (CRS 2023). Attempts to balance commodity supply and demand may occur through payments to farmers to make up the difference between production costs and low market prices, by promoting shifts to other crops, or by incentivizing removal of cropland from production (e.g., by putting it into a long-term conservation program, such as the Conservation Reserve Program [CRP]). Created in the 1985 Farm Bill, the CRP was designed to reduce soil erosion, control commodity overproduction, and provide wildlife habitat by establishing long-term leases that pay landowners to take marginal lands out of cultivation and plant perennial forms of land cover instead (Hohman and Halloum 2000). These policies illustrate how the Farm Bill may drive landscape patterns. However, farmers may respond to favorable short-term market conditions or to weather conditions (especially precipitation) on a time scale finer than the 4–6-year Farm Bill spans. Thus, differences in Farm Bill policies, commodity prices, and precipitation may be manifested in different cropping systems and thus in landscape pattern, but this is not well-understood.

The Yazoo-Mississippi Delta (regionally known as “the Delta”) is a portion of the Mississippi Alluvial Plain that is intensively altered for soybean, corn, cotton, and rice production (Cobb 1992; Saikku 2005; Ambinakudige and Intsiful 2020; Yasarer et al. 2020) (Fig. 1a). Crop rotation is common, resulting in land-cover compositional changes (Ambinakudige and Intsiful 2020; Yasarer et al. 2020). Yearly rotations generally involve two crops (Bennett et al. 2012), generating routine fluctuations in landscape patterns (Corry 2018, 2019; Jeanneret et al. 2021). However, previous studies had limited detail about distributions and timings of land-cover changes. Thus, we lack knowledge about spatio-temporal heterogeneity, whether other structural properties (configuration, connectivity, and context) mirror changes in land-cover composition over time, and where the most stable (and the most dynamic) portions of the Delta occur. Moreover, we lack knowledge of how federal policies, commodity prices, or precipitation may drive these patterns, even though such knowledge informs regional management for sustainable agriculture (Faulkner et al. 2011; Parajuli et al. 2013; Risal and Parajuli 2019; Risal et al. 2020; Lo and Pringle III 2021).

Fig. 1
figure 1

a Map depicting the Mississippi Alluvial Plain (stippling) within the US states of Missouri (MO), Arkansas (AR), Louisiana (LA), and Mississippi (MS) (state boundaries as thick teal lines); the Delta portion (dark gray) occurs within 19 Mississippi counties (thin teal lines). b–o) Time series of land-cover changes from 2008–2021 for nine land-cover classes in the 19 Mississippi counties that comprise the Delta: (1) Bolivar, (2) Carroll, (3) Coahoma, (4) Desoto, (5) Grenada, (6) Homes, (7) Humphreys, (8) Issaquena, (9) Leflore, (10) Panola, (11) Quitman, (12) Sharkey, (13) Sunflower, (14) Tallahatchie, (15) Tate, (16) Tunica, (17) Warren, (18) Washington, and (19) Yazoo

Full size image

Our objectives were to (1) quantify the spatio-temporal properties of land cover within the Delta. We sought to (2) identify areas of the Delta that were subject to the most (and least) landscape change through time. We (3) measured these changes via landscape metrics, which allowed us to (4) evaluate which metrics effectively characterized this agroecosystem and its dynamics. We (5) examined potential drivers of patterns (Farm Bill policies, commodity prices, precipitation). Lastly, we (6) discussed implications of these dynamic patterns of land-cover diversity for resource management.

Methods

Data

A summarized methodological workflow is in Supplementary Information (SI) Fig. S1. Land-cover data from 2008 to 2021 (hereafter, study period) for the 19 Mississippi counties of the Delta were downloaded from the United States Department of Agriculture National Agricultural Statistics Service’s Cropland Data Layer (USDA NASS CDL; https://nassgeodata.gmu.edu/CropScape/, Accessed June 2022; Boryan et al. 2011; Han et al. 2012). This land-cover data had a resolution of 30 m × 30 m pixels (cells). Of the > 140 land-cover categories within the CDL, only 36–47 occurred within the Delta in any given year. These were reclassified in ArcMap 10.8.1 (Esri, Redlands, CA) to nine mutually exclusive land-cover categories (Table 1). Given the regional importance of soybeans, corn, cotton, and rice (Ambinakudige and Intsiful 2020; Yasarer et al. 2020), those categories remained as-is from the CDL for analysis. All other crop types were less common (< 5% in total, and typically < 1%) and so were combined with fallow/idle croplands, hay fields, and pastures for livestock grazing into “other agriculture.” CRP lands occur in multiple habitat forms (e.g., grasslands, riparian, and/or timbered areas) and are not distinguishable from non-CRP regions in the CDL (to preserve landowner privacy). Thus, CRP of the Delta was likely distributed amongst the “other agriculture”, “wetlands”, and “upland forest” categories. Double cropping (the CDL’s designation for crop rotation within the same field in the same year) was rare (≤ 5.5% per year, SI Table S1). When double cropping included only one of the four primary crops, it was assigned as the primary crop type. Double cropping including two-primary crops were especially rare (< 0.5% per year, SI Table S1) and assigned to “other agriculture” because of the inability to assign it to an exclusive category.

Table 1 Reclassification scheme we used, with the four primary crops of the Delta indicated
Full size table

The three most recent Farm Bills (“Food, Conservation, and Energy Act of 2008” [hereafter, FB Era I], “Agriculture Act of 2014” [FB Era II], and “Agriculture Improvement Act of 2018” [FB Era III]) were active during our focal span. We examined their potential influence on land-cover dynamics in the Delta by calculating average proportional landscape coverage of each land-cover class during each of their respective spans of time (2008–2013, 2014–2017, and 2018–2021). Those Farm Bill era averages of each class were then compared to the overall average of each class for the entire focal span. This allowed us to identify those Farm Bill eras that supported above-average and below-average land cover dynamics for each class, and to test specific predictions about the influence of each bill. For example, because the 2008 bill increased support for ethanol/biofuels through production of biomass crops, we anticipated seeing an increase in corn during FB Era I. The 2014 bill removed upland cotton as a covered commodity, but this was reinstated under protection of seed cotton by the Bipartisan Budget Act of 2018, so we expected to see a decline in FB Era II and recovery in cotton acreage in Era III. The 2018 bill included the Grassland Conservation Initiative to retain and expand pasturage, so we expected to see an increase in pasture (part of “other agriculture,” as it accounted for < 5% of the total landscape) in FB Era III. Finally, the 2008 Farm Bill capped CRP acreage at 32 million acres. The 2014 Farm Bill further capped CRP acreage down to 24 million acres; the 2018 Farm Bill increased the cap to 25 million acres. Thus, we expected to see declines in CRP prevalence on the landscape. Because CRP is not distinguishable in the CDL, we could not perform metric analyses on it. Instead, we report its prevalence at national and state levels (https://www.ers.usda.gov/data-products/chart-gallery/gallery/chart-detail/?chartId=106658; https://www.fsa.usda.gov).

To explore two other potential drivers of land cover, we examined growing-season precipitation and commodity prices for the four primary crops. Cumulative precipitation during the planting/growing season (March through June; hereafter, growing season) were obtained from the National Oceanic and Atmospheric Administration National Weather Service Greenville Area station (https://www.weather.gov/wrh/climate?wfo=jan). Commodity prices were obtained from the USDA Economic Research Service’s Commodity Costs and Return database (https://www.ers.usda.gov/data-products/commodity-costs-and-returns, Accessed 1 May 2023).

Metrics

Over 100 metrics quantify different aspects of landscape structure (Gustafson 1998; Frazier and Kedron 2017), with extensive attempts to identify the most parsimonious set needed to capture landscape heterogeneity (Li and Reynolds 1994; Riitters et al. 1995; Cushman et al. 2008). Emergent consensus indicated that landscapes are typically too complex and diverse for only a few metrics to be universally recommended, but “fishing expeditions” for informative metrics are discouraged (Gustafson 2019). Therefore, we deliberately chose 20 metrics to describe the structure of the Delta in terms of composition, configuration, and connectivity changes (Table 2). These metrics examined patterns at two scales: within individual land-cover classes and the overall landscape. Metrics were calculated in Fragstats 4.2.1 (https://www.fragstats.org) using the 8-cell rule to designate patches of a given land-cover type. Use of this rule minimized the influence of any misclassification errors that were present in the CDL (Boryan et al. 2011) and statistical artifacts associated with the modifiable areal unit problem (Wickham and Ritters 1995; Jelinski and Wu 1996; Wu et al. 2000). These metrics were examined with respect to their ability to characterize the Delta in terms of the diversity and amounts of land-cover types present in each year, land-cover patch characteristics such as shape, and how land-cover types were spatially arranged (e.g., in larger, contiguous patches vs. in smaller, dispersed patches). We chose a small subset of metrics that are readily interpretable and further narrowed the subset down to ones that exhibited trends in change over time (as opposed to stasis or to fluctuations without a clear pattern).

Table 2 Overview of metrics used to quantify landscape structure in the Yazoo-Mississippi Delta, 2008–2021
Full size table

Additionally, we developed three new landscape-scaled metrics of per-pixel land-cover richness, land-cover class change frequency, and land-cover class majority (Table 2) because there were no existing metrics to explicitly assess landscape context (that is, the overall setting in terms of homogeneity/heterogeneity in space and persistence over time). Land-cover richness was defined as the number of land-cover categories that occurred at a pixel during the study period. Land-cover class change frequency was defined as the count of land-cover category changes that occurred at a pixel on a year-to-year basis during the study period. Lastly, land-cover class majority was determined by tallying the most numerically prevalent land-cover type at each pixel; this allowed us to identify areas with similar land-use legacies, which may be informative of shared agroecological relationships therein (e.g., agrochemical applications). These contextual metrics allowed us to quantitatively describe transitions in land-cover richness over time (as well as the converse, persistence); we considered these metrics to inform a foundational understanding of this underrepresented aspect of landscape heterogeneity (Pontius et al. 2004).

Analyses

We plotted each of our 20 compositional, configurational, and connectivity metrics for each year to examine trends during our study period. We quantified, on a per-pixel basis, each of our contextual metrics for the same period. Contextual metrics were then mapped to identify portions of the Delta characterized by land-cover diversity or monoculture, and areas with stability or instability in land-cover type across years. In identifying areas of heterogeneity/homogeneity and stability/instability (i.e., context), we quantified the number of land-cover transitions experienced by each pixel over the study period. We examined (1) all transitions among land-cover classes, (2) transitions between consecutive cultivated land-cover classes, (3) transitions between consecutive non-cultivated land-cover types, and (4) transitions between consecutive cultivated and non-cultivated land-cover classes. Calculation of these transition dynamics was coded in Python via a Jupyter Notebook using the open-source scientific data analysis and computing packages pandas (https://pandas.pydata.org/) and NumPy (https://numpy.org/) (SI Document 1). The data were then converted into polygon features, which we refer to as “trajectories” because they contained information on the sequential nature of land-cover change, and we could quantify where those sequential changes had occurred. This enabled us to determine the proportional occurrence and area of those trajectories that contained evidence of sequential crop rotation. We examined sequential field rotation occurrence at three levels of intensity: at least once, at least half of our 13-year study period (i.e., at least seven times), and continuous. We then summarized those relationships by land cover. As an example: for the 1-land-cover class richness category, 9 trajectories were identified. These 9 trajectories represent portions of the Delta that only experienced a single type of land cover during the study period; we refer to this group of 9 trajectories collectively as “land-cover class core areas.” These analyses allowed us to ascribe patterns within the trajectories to processes such as crop rotation, urbanization, or flooding.

To examine the influence of potential drivers on landscape patterns, we first derived comparisons from the compositional metric PLAND over three Farm Bill eras (periods of time spanning each bill). Additionally, we tallied total CRP acreage in Mississippi for the years 2008, 2012, 2014, 2018, and 2021. Finally, to examine the influence of precipitation and commodity prices (Table 3) on fluctuations in the amount of each of the four primary crops, we used the dynlm package v0.3–6 (Zeileis 2019) in R v4.2.2 (R Core Team 2022) to conduct a dynamic linear model time series analysis. Data from 2007 to 2021 were used (2007 data were included to model possible lag effects); the variables were standardized with a z-transformation and plotted on an annual time scale. (Proportion of landscape of each land-cover is a simple and intuitive metric and is statistically correlated to many other metrics; we thus used the proportional land-cover metric as a representative proxy.) This analysis assessed the relationships between the proportion of landscape for each primary crop (log-transformed) and commodity price and precipitation (and their interactions). Six models were compared for each crop: models with 0 year lagged predictors (current year effect), 1 year lagged predictors (previous year effect), and both 0 and 1 year lagged predictors. In each of these cases, models with and without a price by precipitation interaction terms were fit. The model with the lowest AIC score for each crop was retained. Analytical code is available from Ag Data Commons (data.nal.usda.gov; see also SI Document 1, which includes detailed AIC values).

Table 3 Growing-season precipitation and commodity prices for our four primary crops for the years 2007–2021 (2007 data were included to model possible lag effects)
Full size table

Results

Landscape composition

The most land-cover diversity (and changes over time) occurred in the southern, eastern, and northern portions of the Delta (compare Figs. 1b–o). These areas exhibited high spatial and temporal heterogeneity with differing amounts of cotton and corn production by year. Rice production typically was concentrated in the western part of the Delta. Southern portions of the Delta were characterized by extensive wetlands.

Of 23 metrics examined, 12 exhibited informative trends over time: proportional coverage of a given land-cover type, largest patch index, number of patches, average patch size, the coefficient of variation of average patch size, area-weighted average patch size, average Euclidean nearest-neighbor distance, average land-cover contiguity, and overall landscape contagion, as well as our contextual metrics of land-cover richness, land-cover change frequency, and land-cover class majority. Soybeans were the most common land-cover type throughout the Delta across all years (Fig. 2a) and experienced the greatest fluctuations in coverage (~ 30–48%) (SI Table S2; Fig. 2b). Wetlands were consistently second in terms of dominance (~ 23–28%). Corn (~ 6–16%), cotton (~ 3–10%), and the “other agriculture” category (~ 2–13%) were the next most abundant forms of land cover, varying in relative dominance by year. Rice was the rarest of the four primary crops (~ 2–7%); the remaining three land-cover classes were also relatively rare (< 5%), with minimal changes over time (Figs. 2a, b). The overall proportions of cultivated lands (soybeans, corn, cotton, rice, “other agriculture”) to non-cultivated lands (wetlands, water, upland forest, developed) were remarkably stable; departures of ± 5% only occurred twice over our study period (Fig. 2c). The Delta thus experienced no net loss of land-cover types into or out of agricultural production from 2008 to 2021. The largest patch index (Figs. 2d, e) indicated that the largest patch for each land-cover type generally stayed the same size over time, except for soybeans, which fluctuated greatly from year to year, with no overall increasing or decreasing trends (SI Table S3).

Fig. 2
figure 2

Composition metrics through time for each of the 9 classes: a Percentage of landscape (PLAND); b Box-and-whisker plot for PLAND; c) Proportion of cultivated (solid lines) and non-cultivated (hashed lines) lands in the Delta from 2008 to 2021; d Largest patch index (LPI); e) Box-and-whisker plot for LPI

Full size image

Landscape configuration

Fluctuations in the number of patches were especially evident for soybeans, corn, and other agriculture. Upland forest and rice accounted for the smallest area and fewest patches (SI Table S4; Figs. 3a, b).

Fig. 3
figure 3

Configuration metrics through time for each of the 9 classes [panel I]: a Number of patches (NP); b Box-and-whisker plot for NP; c) Patch area mean (AREA_MN); d) Box-and-whisker plot for (AREA_MN)

Full size image

Average patch size was largest for the four primary crops and for wetlands (SI Table S5; Figs. 3c, d). Average patch size for rice was larger than for most other land-cover types, indicating that what little rice was present was spatially contiguous. The developed class displayed fluctuations in number of patches (Figs. 3a, b). The developed class also displayed limited variation in area-weighted mean patch size (Figs. 4a, b) yet had greater variation in the coefficient of variation in mean patch size (Figs. 4c, d). The increasing trend in area-weighted mean patch size (SI Table S6; Figs. 4a, b) and stabilization of variation in patch size (SI Table S7; Figs. 4c, d) for soybeans signals coalescence of soybeans fields into larger patches. Similarly, when metrics of patch size (which indicate how large the contiguous clumps of a given land-cover type are) are examined in tandem with the number of patches of a given land-cover type, an increase in the number of patches and decrease in patch size over time for a given land-cover type can be an indicator of fragmentation. The trends of increasing patch size and a weaker one of a decrease in the number of patches of soybeans indicates that fewer but larger patches are coalescing, forming monocultures.

Fig. 4
figure 4

Configuration metrics through time for each of the 9 classes [panel II]: a Area-weighted patch area mean (AREA_AM); b Box-and-whisker plot AREA_AM; c Coefficient of variation of mean patch area (AREA_CV); d Box-and-whisker plot for AREA_CV

Full size image

Landscape connectivity

Mean Euclidean nearest-neighbor distances were consistently high for upland forest patches, signaling isolated patches separated by > 500 m on average. Rice patches were also spatially well-separated, potentially reflecting presence of poorly draining soils that are preferentially utilized for rice production in the Delta. Rice patches generally became more isolated over time (SI Table S8; Fig. 5a and b). The lowest nearest-neighbor distances of the nine land-cover types occurred for developed areas, soybeans, and wetlands, presumably from different mechanisms.

Fig. 5
figure 5

Connectivity metrics through time for each of the 9 classes: a Average Euclidean nearest-neighbor distance (ENN_MN); b Box-and-whisker plot for ENN_MN; c Average contiguity (CONTIG_MN); d Box-and-whisker plot for CONTIG_MN; e Landscape contagion of land-cover types (CONTAG)

Full size image

The mean contiguity index (Figs. 5c, d), where larger index values indicate the presence of more contiguous patches of a given land-cover type (LaGro 1991), reinforced the observation that wetland patches were relatively contiguous and stable (SI Table S9). Peaks seen for cotton and rice in some years correspond to peaks in average patch size for those land-cover types (Figs. 5c, d). When examined in conjunction with the proportion of landscape covered and patch size, these peaks indicate changes in landscape patterning. Specifically, despite a doubling of cotton area between 2008 and 2017, cotton typically existed in relatively small, dispersed patches; only in 2017 was there a coalescence of cotton patches to form larger, monocultural areas. Similarly, rice patches became increasingly contiguous despite an overall decrease in proportional coverage of this crop type. Overall landscape contagion was intermediate between complete dispersion and complete aggregation of land-cover types, indicating relative stability in spatial admixture (SI Tsble S10; Fig. 5e).

Landscape context

About a quarter of the Delta exhibited no spatial or temporal heterogeneity, representing stable monocultures (Franco et al. 2022). Such instances existed for each of the cultivated categories, especially soybeans. Soybean core areas were distributed throughout the Delta, with the highest concentration in the central and northern portions. Cotton had the second highest area of stable monoculture, with core areas primarily in the northern Delta. “Other agriculture” had many small core areas, often near the Delta periphery. Rice and corn had very small core areas (Table 4; Fig. 6). When non-cultivated lands were considered, the largest core areas were associated with wetlands, distributed throughout the Delta. Areas of water were also highly stable, with core areas typically associated with rivers or lakes; an aquaculture core was observed within the south-central Delta (Leflore and Sunflower Counties [counties 9 and 13 in Fig. 1]). Developed core areas included urban centers and road networks, which were diffusely scattered throughout the region. Lastly, upland forest core areas existed on the eastern periphery (Table 4; Fig. 6).

Table 4 Land-cover class core area distribution and proportional occurrences in the Yazoo-Mississippi Delta, 2008–2021
Full size table
Fig. 6
figure 6

Spatial distribution of land-cover class core areas in the Delta

Full size image

Landscape context as measured by per-pixel land-cover richness indicated that during our study period, most of the Delta was relatively stable in land cover (Table 5, Fig. 7). Transitions among three land-cover classes were the most common and occurred across ~ 33% of the landscape. Transitions among ≥ 5 classes were rare, at ~ 2% of the region in total (Fig. 7). Slightly over a quarter of the Delta never changed land-cover type (the 1-class richness category). Examining pixel trajectories for identification of instances of crop rotation, we found that sequential field rotation was pervasive yet highly variable in space and time. Portions of the Delta that experienced continuous field rotation were especially rare, at ~ 1% of the region in total (Table 5).

Table 5 Land-cover richness distribution and proportional occurrences of field rotations in the Yazoo-Mississippi Delta, 2008–2021
Full size table
Fig. 7
figure 7

Spatial distribution of context metrics: Land-cover Richness [panel I]

Full size image

Landscape context as measured by per-pixel land-cover change frequency indicated that most of the Delta experienced forms of land-cover change that included crop rotation as well as other forms of land-cover change, such as wetland inundation and urbanization (Table 6). Indeed, 73.25% of the Delta landscape experienced at least one land-cover class change, and on average, any given pixel experienced a change in land-cover type once every 5–6 years. Relatedly, 22.63% of the Delta experienced land-cover change during more than half (7 + years) of the study period. Furthermore, 1% of the landscape experienced a class change every year; these areas occurred within every county of the Delta, although they tended to be concentrated in central and southern counties (Table 6, Fig. 8). These temporal patterns of land-cover change in the Delta reflected spatial patterns, with relative stability in peripheral portions of the region and areas with high turnover in the central Delta (Fig. 8). The areas with the least land-cover stability were often occupied by corn and cotton, whereas the most stable areas were wetlands (compare Figs. 1, 7, and 8).

Table 6 Land-cover change frequency distribution in the Yazoo-Mississippi Delta, 2008–2021
Full size table
Fig. 8
figure 8

Spatial distribution of context metrics: Land-cover Change Frequency [panel II]

Full size image

Landscape context was also measured by per-pixel land-cover class majority (Fig. 9). Soybeans were the land-cover majority across 48.61% of the Delta, followed by wetlands as the land-cover majority across 25.57% of the Delta. Upland forest was the land-cover majority across only 0.30% of the Delta, ranking last (Table 7). Land-cover majority was indeterminate for small portions of the Delta (white gaps in Fig. 9).

Fig. 9
figure 9

Spatial distribution of context metrics: Land-cover Class Majority [panel III]

Full size image
Table 7 Land-cover class majority distribution in the Yazoo-Mississippi Delta, 2008–2021
Full size table

Drivers of landscape pattern

During FB Era I, corn, rice, water, and “other agriculture” featured above-average occurrences. Soybeans, cotton, wetlands, upland forest, and developed each had below-average occurrences. During FB Era II, only soybeans featured above-average occurrences, with below-average occurrences for all other classes except for “developed,” which was exactly average. During FB Era III, soybeans again experienced above-average occurrence, as did cotton, wetlands, upland forest, and developed. Below-average occurrences for corn, rice, water, and “other agriculture” were observed. With respect to CRP, there was a peak in acreage in 2008 nation-wide and in Mississippi (Table 8), with decreases since then. Thus, some predicted influences of the Farm Bill were upheld in our data. For example, the prevalence of corn was above average during FB Era I (as predicted, given the 2008 Farm Bill’s promotion of ethanol). However, cotton did not exhibit a decline and then recovery (instead, it increased throughout). Finally, “other agriculture” (which included pasture/grassland) was at its lowest coverage during FB Era III, despite the 2018 Grassland Conservation Initiative (SI Table S2). As expected, CRP acreage declined over time (Table 8).

Table 8 CRP acreage in Mississippi
Full size table

Some landscape changes through time were coincident with changes in precipitation (Table 3). The most apparent trend observed was that for years of above-average growing-season rainfall, there were corresponding increases in the percent of the Delta composed of wetlands, most notably in 2014 and from 2019 to 2021. However, in the case above, which land-cover class experienced the corresponding decreases among those years differed (2014: corn and “other agriculture”; 2019: soybeans; 2020: corn, cotton, and “other agriculture”; 2021: cotton and “other agriculture”). For example, during relatively dry growing seasons, we observed general increases in the amount of corn (2010–2013) and soybeans (2015, 2017–2018). Landscape change was also coincident with changes in commodity prices (Table 3). Each primary crop experienced above-average prices in 2011, 2013, and 2021. Peak prices occurred in 2012 for both soybeans and corn, whereas peak price for cotton was in 2011 and the peak price for rice occurred in 2008. Each of the four primary crops experienced below-average prices between 2015 and 2017. The extent of cotton (R2 = 0.21, p = 0.28) and rice (R2 = 0.06, p = 0.71) were not well-predicted by price or precipitation or their interaction. For corn (R2 = 0.61, p = 0.01), price and precipitation in the previous year best predicted corn extent in the current year (i.e., lag effect evident), with no interaction term included in the best model. Specifically, higher price and lower precipitation in the previous year were associated with higher extent of corn in the current year. For soybeans (R2 = 0.61, p = 0.02), we found that extent was best predicted by price, precipitation, and their interaction in the same year (no lag effect evident). Lower precipitation in the same year was associated with higher extent of soybean cultivation, and there was a positive interaction between soybean price and extent (the correlation between price and extent was more negative when precipitation is low, and more positive when precipitation was high). In summary, the amount of corn was positively associated with the previous year’s prices and negatively with precipitation whereas soybean acreage was lower in high-precipitation years and more dependent on commodity prices. For detailed AIC values results, see SI Document 1.

Discussion

Landscape change within the Delta during our study period was predominantly a function of agricultural activity and was broadly anticipated (Ambinakudige and Intsiful 2020; Yasarer et al. 2020). Wetlands were also key landscape features in the Delta, forming largely stable, contiguous patches. Other forms of land cover (water, developed, forest) were relatively rare. Our assessments illuminated contemporary shifting mosaic patterns across the Delta. Moreover, our chosen metrics characterized stable monoculture, land-cover class core areas, spatial arrangements within and among land-cover classes, crop rotations, and other trajectory behaviors.

It should be noted that different landscape patterns may yield the same values for some metrics, and many metrics are sensitive to data resolution and spatial extent, complicating comparisons of different landscapes (Remmel and Csillag 2003). Despite these limitations, use of multiple metrics as diagnostic pattern indicators can shed light on how human activities affect disturbance regimes, biodiversity, and ecosystem services (see review by Remmel and Mitchell 2022). Some metrics may be valuable elsewhere but not in the Delta (and vice versa). For example, measures of connectivity may be especially useful in examining changes in flow regimes in wetlands (due e.g., to drought, damming, etc.), whereas an assessment of patch density may be warranted if metapopulation structure is suspected for a focal organism.

Of the 23 metrics we used, a smaller suite of 12 was able to adequately characterize landscape pattern and change in terms of composition (percentage of landscape covered by a given land-cover type), configuration (number of patches of a given land-cover type, average patch size, the coefficient of variation of average patch size, area-weighted average patch size, largest patch index), connectivity (average Euclidean nearest-neighbor distance, average contiguity, landscape contagion), and context (land-cover richness, land-cover change frequency, land-cover class majority). These metrics exhibited trends over time; the others did not, being either relatively static or varying greatly and erratically (SI Figs. S2–S12). Some metrics are redundant (e.g., average patch size and area-weighted average patch size, or contiguity and contagion), so an even smaller subset of 10 could be used for greater parsimony.

In land-cover composition (SI Table S2, Figs. 2a, b), soybeans were dominant in coverage and appeared to coalesce into larger patches (Figs. 2d, e). The increased prevalence of soybeans is locally regarded as a function of price and reduced costs of production. It has also been locally regarded as a function of international market effects, especially increased competition associated with cotton (personal communication, Dr. Martin Locke [NSL Director] and Lee Vi Haas [WQERU Biologist and producer]). However, soybeans also experienced the greatest variation in coverage (Figs. 1b–o and 2a). The coalescence of soybeans into a larger monoculture has region-scale implications on fertilizer and water use, soil carbon, soil microbial diversity, yield, and other effects (Liu et al. 2006; Li et al. 2010). Such a trend towards landscape homogenization (potentially linked with federal agricultural policies such as the Farm Bill) occurs due to loss of land-cover diversity and structural heterogeneity, with knock-on negative effects on biodiversity through a reduction in resource availability. Incentivized conservation efforts (such as the Conservation Reserve Program in the US, or agri-environment schemes in Europe) are large-scale approaches to returning habitat value to cultivated landscapes (Batáry et al. 2015).

Among the other primary crops, corn was next-most variable in coverage, with areas of majority coverage within Humphreys, Leflore, Sharkey, and Tallahatchie counties (counties 7, 9, 12, and 14 in Fig. 1). Cotton had the third-highest variation in coverage, with areas of majority coverage in Coahoma, Quitman, and Tunica counties (counties 3, 11, and 16 in Fig. 1). Rice area had been described as stagnant (Ambinakudige and Intsiful 2020), yet in later years of our study, we observed a decreasing trend in rice cultivation. Among the four major regional crops, rice had the lowest variation in coverage, with limited areas of majority coverage, and even fewer persistent land-cover class core areas. As noted previously, rice production is concentrated on poorly draining soils in the Delta and requires production practices that may be cost-intensive (e.g., standing water). Coupled, these factors likely explain in part the observed landscape patterns for rice. In early years of our study, a diffuse distribution of rice occurred in central and northern Delta counties. These historically rice-dominated areas were typically rotated with and then generally replaced by soybeans; this may reflect a shift in irrigation practices (Yasarer et al. 2020). Moreover, as rice production was not well-predicted by price or precipitation, it would suggest that other, more localized drivers (e.g., machinery access, labor availability) are likely important determinants for regional rice cultivation. For “other agriculture,” variation in coverage was the second highest among all nine land-cover types. “Other agriculture” had diffusely distributed majority occurrence areas, generally more common in eastern and southern portions of the Delta, and often near wetlands. Class core areas for “other agriculture” were often near the periphery of the Delta.

Among non-agricultural land-cover classes, wetlands accounted for the most surface area during all years, with stability (persistence) in coverage, likely because they are too wet to farm and/or occur in areas that flood periodically. Large areas of wetlands were present along the Mississippi River and eastern limits of the Delta (often as components of or proximal to protected lands, e.g., Delta National Forest). Extensive flood control and aquifer management activities in the Delta likely account for some of the persistence we observed (Cobb 1992; Saikku 2005; Dakhlalla et al. 2016; Alhassan et al. 2019; Yasarer et al. 2020). Observed wetland stability was also likely influenced by wetland type, as large, bottomland-forested wetlands were more readily identifiable in the CDL. Persistent aquaculture was especially present in Leflore and Sunflower counties (counties 9 and 13 in Fig. 1).

As evidenced by variation in land-cover class changes, much of the Delta was dynamic. For any given pixel, the land cover richness ranged from one to eight classes, with most pixels fluctuating among three land-cover types (Table 5, Fig. 7). This three-class variation is consistent with regional crop rotation of soybeans, corn, and cotton (Ambinakudige and Intsiful 2020). The highest interannual variation in number of patches (Figs. 3a, b) was in the “other agriculture” class. This pattern may be due to (1) our methodological approach for double cropping among primary crops, albeit a relatively rare consideration (SI Table S1); (2) fluctuations with crops due to crop rotation, including to the fallow/idle cropland category; and/or (3) a consequence of flooding (fostering delayed planting/crop failure). Limited variation in the number of patches of upland forest, water, and wetland classes is consistent with the stability we observed in coverage for these land-cover types. The presence of numerous developed patches despite this land-cover type’s relative rarity on the landscape is indicative of the dispersed small settlements in the Delta. However, the high variability in the number and sizes of these developed patches may be inherent to the Cropland Data Layer, as “urban” (i.e., the developed land-cover type) is known to be difficult to characterize (McIntyre 2011), and impervious surfaces are difficult to distinguish from some other land-cover types in remotely sensed imagery (Heintzman and McIntyre 2019).

Rice exhibited the greatest interannual fluctuations in mean patch size (SI Table S5, Figs. 3c, d). In later years of our study, the number of patches of rice generally declined and mean patch size generally increased. Similar later-year patterns were present for cotton as well. Notably, cotton had the overall highest mean patch size in 2017, when it also had the third-fewest number of patches. We attribute these patterns in part to crop rotation displacement of corn, resulting in a relatively denser distribution of cotton patches. Soybeans, although consistently among the top-4 land-cover classes with respect to mean patch area, also exhibited reciprocal oscillations between larger number of patches and lower mean patch area (e.g., in 2009 and 2017), which we attribute to rotation with cotton and corn or “other agriculture.” (These variations were particularly noticeable when controlling for the greater overall amount of soybeans on the landscape [Fig. 1b–o].) For corn, mean patch area was generally stable during early years, with greater oscillations beginning in 2013, and increasing to a class-specific peak in 2021; we attribute this peak to displacement of cotton and “other agriculture.”

Upland forest and rice had the greatest Euclidean nearest-neighbor distances and most variation in that metric (SI Table S8, Figs. 5a, b). The upland forest class was especially rare, with patches restricted to the extreme eastern fringe of the Delta. Although the distances among upland forest patches began high and remained so, the range of variation was far less than that for rice. The range of variation in distances among rice patches began to increase from 2010 onward, indicative of possible elimination within crop rotation regimes (being largely replaced by soybeans).

With respect to drivers of these patterns, changes within the Delta were influenced by precipitation and commodity price. During years with above-average growing season precipitation, we observed significant increases in wetlands and increases in “other agriculture,” likely representing flooded croplands being fallow/idle. During years with below-average growing season precipitation, we observed increases in soybeans and corn. Because the Delta is extensively irrigated via groundwater extraction (Dakhlalla et al. 2016; Yang et al. 2019; Yasarer et al. 2020; Lo and Pringle III 2021), some effects of precipitation on landscape composition may have been masked, especially if irrigation allowed producers to take advantage of higher commodity prices despite potentially dry conditions. Although we observed increases in soybeans during other dry times, our data suggest that the previous year’s conditions were not a strong predictor of soybean extent. Both cotton and rice were not well-predicted by price or precipitation. Comparatively, the influence of the Farm Bill as a driver of agricultural change was not as clear as the effect of commodity prices and precipitation on landscape cover. This is perhaps not surprising, given the long time spans associated with each Bill; changing prices and precipitation require more immediate responses. The landscape homogenization associated with agricultural intensification with respect to soybeans thus appears to be more a function of local to regional factors rather than federal-level policies.

These findings have potential relevance in resource management. For example, areas of spatial homogeneity and temporal stability may be relatively more vulnerable to pest outbreaks (Dalin et al. 2009). Additionally, the identification of the Delta as a shifting mosaic has implications on biodiversity conservation. For instance, Conservation Reserve Program lands would likely be the most fixed and stable portions of the landscape, at least on the decadal time scale of most CRP leases; if leases are not renewed, however, then it would be useful to know the locations of potential alternate habitats for displaced biodiversity, whether they are likely reachable, etc. Likewise, provision of ecosystem services such as pollination from native bees would vary in intensity over space and time, making it difficult to establish where supplemental pollination from domestic honey bees (Apis mellifera) would be needed. Finally, industries and individuals seeking to promote sustainable and regenerative agriculture may influence landscape change for climate adaptation and to address food insecurity (Schattman et al. 2023). Our results suggest that such efforts would be most effective, at least on a regional scale, when allowing for short-term responses associated with commodity prices and weather. Our approach thus provides a useful model to management agencies for identifying areas of spatio-temporal heterogeneity or stability.

Conclusions

Our study quantified changes in landscape patterns in the Delta. The metrics we calculated indicated relative stability in landscape composition over our study period, but with fluctuations in configuration, connectivity, and context that indicated latent landscape changes that metrics of composition did not detect.

Although the Farm Bill can provide billions of dollars in expenditures that influence the types and amounts of crops grown, this funding may be smaller and less responsive than changes in commodity prices. The overall effects of the different Farm Bills were mixed with respect to proportional coverage. Coupled with the general decline in CRP acreage, these results support the observation of a homogenizing agricultural environment in the Delta.

Abundances of soybeans and corn were significantly associated with precipitation and commodity prices, although their responses to temporal variation differed (lags for corn but not for soybeans). Patterns of cotton and rice were not explained by either precipitation or commodity prices. As our models were able to explain only ~ 60% of the variation in coverage of soybeans and corn (and did not perform well for cotton or rice), other drivers (e.g., irrigation, soil type or texture, crop insurance, development of transgenic crops, etc.) are likely important within the Delta (Yasarer et al. 2020).

Given the relative stability of most land-cover classes, similar ratios of cultivated vs. non-cultivated lands, and no net loss or gain of cultivated lands, the Delta appears to be a shifting mosaic whereby land-cover categories are not commonly eliminated but rather are repositioned. This also suggests that the Delta has likely plateaued in terms of cultivated land since the remaining lands (wetlands, forests, urbanized areas) are not optimal for agricultural production. Thus, it is likely that future agricultural productivity can only be increased in the Delta through management strategies other than conversion of more land to cultivation.

In summary, our study addresses knowledge gaps in spatio-temporal heterogeneity of agroecosystems (Hopfenmüller et al. 2014; Corry 2019; Coffin et al. 2021) via an assessment of shifting mosaic dynamics within the Yazoo-Mississippi Delta. Our insights can guide agroecosystem optimization (Jeanneret et al. 2021; Tscharntke et al. 2021) and be a model for examining similar agroecosystems (Yang et al. 2019; Jeanneret et al. 2021).