Dynamic Space Use of Andalusian Rice Fields by Overwintering Lesser Black-Backed Gulls is Driven by Harvest-Related Flooding

Background: Research on the space use and behavior of waterbirds—as highly mobile of wetland habitats—yields important insights on human-wildlife interactions of ecological and societal importance under global change. The extent to which dynamic (within-season) changes in anthropogenic landscapes affects these interactions is poorly understood. Lesser black-backed gulls (Larus fuscus) are prominent biovectors of biological and articial materials, and have exhibited large population increases in parts of southern Europe in recent decades. Methods: We combined GPS tracking, earth observation, accelerometry and eld observations to study the space use of overwintering in a mixed rice eld landscape in Andalusia, Southern Spain. We used Manly selectivity metrics and classied remote sensing imagery to directly evaluate space use and habitat selection for these gulls and how it changed throughout the rice harvest cycle. Results: Analysis of over 45,000 GPS xes and 14 classied remotely-sensed images from winter 2016-17 showed dynamic space use driven by the harvest-related ooding across the rice harvest cycle. Prior to harvest, gulls foraged in rice paddies during the day and roosted in adjacent waterbodies (the River Guadalquivir, and sh ponds) at night. During harvest, they spent nearly 100% of their daily cycle within the rice elds, foraging in harvested paddies and roosting in post-harvest, ooded paddies. After harvest, they roosted in ooded elds at night and foraged at landlls in the surrounding landscape. Conclusion: Gull space use at and scales was closely linked to dynamic land and water management over the rice agricultural illustrating the scales activities The frequent early-spring elds


Background
Anthropogenic landscape change is the leading driver of biodiversity loss in the 21 st century and a major mechanism of continued global environmental change (Sala et al., 2000;Foley et al., 2005;Di Marco et al., 2018). Research on human-dominated landscapes is thus an important eld of study in ecology and conservation, yielding insights for predicting and understanding interactions between biodiversity and human systems in the Anthropocene. This is especially true in wetland environments, which are highly dynamic and strongly affected by local topography and hydrology. Natural wetlands have been reduced considerably in global extent while being replaced by anthropogenic wetlands such as rice elds and reservoirs (Davidson et al. 2018). As highly mobile animals that readily practice habitat supplementation (using multiple habitat types during their daily or life cycle), birds, and waterbirds especially, are important ecological connectors in heterogeneous landscapes (Green and Elmberg, 2014;Şekercioğlu et al., 2016).
The movement behavior and space use of waterbirds in human-dominated landscapes can thus have of Europe, and decreased availability of sheries discards from altered shing practices have led to a larger dependence of gulls on terrestrial and anthropogenic foods (Harris, 1965;Camphuysen, 1995;Oro, 1996). The spatial shift in Larus gulls and their growing use of food resources associated with humans raises concerns about their environmental and societal impacts, especially given their known role as biovectors, linking terrestrial and aquatic as well as anthropogenic and ecological systems (Martín-Veléz et al., 2019, 2020. For example, increased availability of terrestrial food in anthropogenic landscapes has led to roosting of large numbers of gulls on inland water bodies, with important impacts on water quality LBBG winter inland across Andalusia in southern Spain, but a combination of movement and census data suggest that in the rst half of the winter they concentrate in the extensive rice eld area (up to 38,000 ha.; Ramo et al. 2013) adjacent to the Doñana wetlands. These rice elds are the node with the greatest centrality within the network of important habitat sites for LBBG across the region (Martín-Vélez et al. 2020). During the rice harvest period, they feed on invasive cray sh exposed as rice plants are We combined high-resolution spatiotemporal GPS tracking and accelerometry data with remote sensing and eld observations to investigate the space use and movement behavior of LBBG overwintering in a rice eld landscape in Andalusia, Southwest Spain. Our objectives were to 1) quantify the spatiotemporal changes in the availability of different habitats within rice paddies throughout the harvest cycle, and determine how this in uences gull behavior and habitat selection within rice paddies, 2) determine changes in the selection of rice elds over nearby alternative habitats within the Guadalquivir river delta during the winter period, and whether these are driven by rice eld management, 3) establish the relationship between behavior and habitat use, using data from eld observations and accelerometers.

Study Area
We conducted data collection and analysis at three scales, 1) the larger rice eld landscape, 2) the local rice eld complex, and 3) the interior of rice paddies ( Figure 1). The larger rice eld landscape is a rectangular 4200km 2 area centered around the rice elds but which includes the Doñana National and Natural parks, Veta la Palma aquaculture ponds, and adjacent agricultural areas within the Guadalquivir river delta ( Figure 1); this area encompassed >90% of all LBBG GPS xes during the study period, and was used as the study area for all static habitat selection analysis. The rice eld complex is a zone of intensive rice cultivation to the Northeast of Doñana National Park, and is the largest rice production area in Spain. It lies within a Biosphere Reserve (Green et al. 2018) and comprises several hundred small (4-35 ha) polygonal paddies divided by dikes, small roads, and irrigation canals and portions of the Guadalquivir river. The rice eld complex includes all landcover falling within a 250m buffer of rice paddy edges. It also includes larger urban settlements, fragments of shrubland and forest, relict natural marshland and extensive areas of dryland agriculture including cotton elds and fruit orchards. Finally, we examined habitat features in the interior of individual paddies, which we de ned as the area at least 45m interior of the paddy edge (Toral et al., 2011). We used this analytical scale for our dynamic habitat analysis of space use with respect to within-paddy condition.
Habitat availability in rice paddies follows a consistent pattern according to the annual rice harvesting cycle, which we divided into three phases. During the pre-harvest stage, all paddies contain unharvested rice. The active harvest phase begins with rice harvest starting in late September and early October, and paddies are tilled or plowed after harvest, and nally ooded. Flooded conditions in rice paddies persist from October to December or January, depending on ambient precipitation and evaporation rates (Toral et al., 2011). The active harvest phase is marked by high heterogeneity in habitat types, because the harvest, tilling, and ooding occur in a spatiotemporally staggered and piecemeal fashion because of limited harvesting equipment, leading to asynchronous transitions between within-paddy conditions (Toral et al., 2011). The post-harvest phase begins in mid-to late-January, when all elds have typically been harvested; for the purposes of this study, the post-harvest phase of the overwintering season ends with the departure of overwintering LBBG in mid-March.

GPS Tracking Data
We used data on the movements of LBBGs tracked using Global Positioning System (GPS) trackers We created two data subsets based on GPS xes that coincided with the dates of available imagery for our dynamic habitat analysis. The rst of these consisted only of data collected on the same day as each available remotely sensed image (24-hour window), and consisted of 3,082 xes, and the second (expanded dataset) included xes from one day before and one day after each image (72-hour window), for a total of 9,185 xes; this is slightly less than three times as many as the 24-hour window dataset, because two images were two days apart from the proceeding image, leading to an overlap in their 72hour windows. The 24-hour window dataset has a lower probability that harvest status changed between image collection and a given GPS x within that window but has reduced statistical power due to a smaller sample size, while the 72-hour window dataset increases sample size and statistical power but at slightly increases the risk of including GPS xes collected under different harvest conditions.
For each of the 6 remaining birds, we also used the UvA-BiTS virtual lab (Bouten, 2018) online client to project and visually check trajectories of un ltered data for each gull prior to quantitative analysis. This was done by subsampling data points to 3-hour intervals and then visually examining GPS trajectories for coarse-scale movement behavior within the rice eld landscape (e.g., movements between the rice paddies and other landscape features, or repeated daily movements to features outside of the rice eld landscape). Any forays outside of the rice eld landscape ( Fig. 1) were visually examined at a 20-minute interval resolution to identify additional destinations during the overwintering season.

Remote-sensing Imagery, Classi cation, and Ancillary Spatial Data
We downloaded and classi ed a total of 14 images (ten from Landsat 5, 7 and 8 and four from Sentinel-1), with dates ranging from 2 September 2016 to 17 February 2017 (days 7 to 175) to detect and quantify the dynamic availability of habitat resources for gulls at the rice paddy scale throughout the overwintering period (hereafter "dynamic habitat analysis"). All images were geometrically and radiometrically corrected according to Aragonés et al. (2005). We classi ed harvest status landcover types within each paddy in the rice eld complex for each available image date. Following Toral et al.
(2011), we restricted the classi cation of rice paddies to the interior of each paddy by excluding a 45m buffer inwards from their outer edge to avoid spectral confusion with land-cover categories that are not found within paddies (e.g. shrubland, roads). Notably, this means that use of features outside of paddy interiors (e.g., dikes and roads between paddies) were quanti ed only in our static habitat analysis.
We classi ed these clipped images using K-means unsupervised classi cations with the RStoolbox package (Leutner et al., 2019) in R 3.5.2 (R Core team, 2019). We chose a value of K based on the largest consensus among a suite of 30 indices using the Nbclust package (Charrad et al., 2015) in R. Where no clear consensus among metrics was available, the parsimonious (smallest K) competing value was chosen. We used the resulting image with the highest number of habitat classes present as the basis for thematic classes used in all other classi ed images. We attempted at minimum to classify harvest status conditions that were relevant to habitat use by gulls (see also Martín Vélez et al. 2020), speci cally green vegetation or unharvested rice (low-density foraging activity), recently harvested rice (large amounts of active foraging), tilled rice (active foraging), and ooded areas (loa ng, sleeping, and occasional foraging). These classes were assigned using knowledge from eld visits and reference images.
We detected 7 thematic classes for harvest status within rice paddies: bare ground, dry tilled, wet tilled, deep ooded, shallow ooded, harvested rice, and green vegetation (either pre-harvest rice or other plants colonizing elds post-harvest). Deep-ooded and shallow-ooded paddies were detected based on spectral signatures, wherein deep-ooded elds showed a signature strongly indicative of water, while shallow-ooded paddies showed some mixture of water and bare ground, indicating that land was re ecting some light through a shallow layer of water; water depth in neither case likely exceeds 30cm, so the terms 'deep' and 'shallow' are based only on relative differences in water depth.
In addition to image-based classi cations of harvest status, we generated an additional map of habitat resources that were unlikely to change throughout the wintering season for an analysis of static habitat features at the rice eld landscape scale. For this we used a reclassi ed version of landcover data from the SIPNA database (Sistema de Información sobre el Patrimonio Natural de Andalucía, Junta de Andalucia, REDIAM, 2019), which included rivers and dikes, roads, non-rice agriculture, urban areas, natural and arti cial ponds, and other more permanent features in the larger rice eld landscape. We simpli ed thematic classi cations for this dataset to match behaviorally-relevant habitat types for LBBG for a nal habitat map of the larger rice eld landscape (for extent, see Fig. 1). This involved combining classes that presumably had no impact on gull biology into single classes, for example, combining industrial and residential development into "urban", and different types of fruit orchards into "non-rice agriculture". Areas known to be rice paddies were lumped into a single class, "Unclassi ed Rice" to re ect birds using rice paddy interiors in our static habitat analysis (for days without imagery and subsequent harvest status information; Additional File 2). Finally, we generated a shape le consisting of a 250m buffer around all rice paddies, to be used for determining whether or not given GPS xes were occurring within the rice eld complex or in another part of the rice eld landscape (e.g., land lls, more distant sections of the Guadalquivir river, aquaculture ponds or natural marshes, Fig. 1).

Accelerometer Data
Tri-axial accelerometer data (1-second segments of acceleration data in three dimensions collected concomitantly with GPS xes at 20hz frequency) were available for four of the tagged gulls. We removed accelerometer segments with more or less than 1-second of data to avoid bias, yielding a dataset of 10,641 accelerometer segments. We converted all accelerometer measurements to G's (equivalent to gravity at the Earth's surface, 9.8m/s 2 ), and calculated mean and standard error along each directional axis for each segment in order to calculate dynamic acceleration. Finally, we calculated overall dynamic body acceleration ( . These classes were terloco (terrestrial locomotion, walking), ap, ex ap (rapid, active apping), soar (gliding ight), manouvre ( apping and soaring combined), peck, , sit/stand (resting or inactive with the body level), walk, oat, and boat (resting on a moving object). We associated ex ap, peck, and terloco with feeding and food-searching related behaviors, ap, manouvre, and soar with travel, and boat, oat, and sit/stand with resting. These accelerometry-based behavioral classi cations differ from those outlined for eld observations (below), which were based only on visual identi cation and observation. We validated accelerometer classi cations by checking the correspondence of classi ed behaviors with other spatial characteristics of the associated GPS x for a subset of 100 xes (e.g., whether oating was associated with points in water, or whether instantaneous velocity was high enough for points classi ed as soaring or ying).

Field Observation Data
We conducted behavioral observations of overwintering LBBG in the rice eld complex to complement and validate the GPS and accelerometer data. We collected observational data between 0900 and 1800 hours from early November to the end of January during the 2018-2019 overwintering period (days 75-157). We constructed an ethogram of seven behaviors that were readily distinguishable in the eld: walking, searching-foraging, ying, swimming, bathing, sitting, and standing. We classi ed all observations according to these mutually exclusive behavior classes. Behavioral data consisted of instantaneous scans as well as focal bird observations. We performed instantaneous scans on paddies that contained LBBG and scanned the entire paddy (including surrounding dikes and airspace above it) for a 1-minute interval and counted the number of gulls performing a certain behavior. To reduce the potential for temporal autocorrelation in behaviors, repeated instantaneous scans in the same site were conducted no less than 10 minutes apart. For observations of focal birds, a single individual was monitored for up to 30 minutes, recording the amount of time spent on each behavior. Observations were ended if line of sight of the bird was lost. We also recorded time of day, harvest status (unharvested, actively being tilled, or harvested, and not ooded, partially ooded, or ooded), weather, and GPS location of the approximate paddy centroid. We conducted 142 instantaneous scans and 22 focal follows (totaling approximately 10h of observation) across 70 locations and 7 days of observation.

Space Use Analyses
We classi ed all GPS points according to whether they fell inside or outside of the rice eld complex, and whether they occurred during the day or night (according to daily sunrise and sunset times, using the R package RchivalTag (Bauer, 2018). We additionally classi ed all points by land use based on our static habitat map (rice eld landscape scale) using a simple overlap point extraction (R package 'raster'; Hijmans & van Etten, 2012). For our dynamic habitat analysis, we classi ed all xes that fell within the interior of rice paddies by harvest status using the same method. We conducted this analysis for both the 24-hour and 72-hour window data subsets. Accordingly, all GPS xes received a static habitat classi cation, and our imagery-based subsets (24-and 72-hour windows) received speci c information on harvest status at the paddy scale for our dynamic habitat analysis. We converted all date-time stamps of GPS xes to a "day after arrival", where day 1 was the day of the rst bird's arrival into the rice eld landscape (for the 2016/2017 overwintering season, this was 27 August, such that 1 January 2017 was day 128). We present all results in both calendar date and day after arrival.
We conducted space use analyses at population and individual scales, using simple overlays and pointselection functions. The relatively small position error for xes ltered for quality (~3m) lends credence to this type of analysis, although such errors could result in biases against detecting space use in thin, linear features like smaller roads and dikes. We conducted separate analyses for GPS xes during the day and night to examine differences in gull behavior across the daily cycle. We also conducted analyses on data that were subset according to the three phases of the overwintering harvest cycle: pre-harvest (days 1-50, or August to mid-October), active harvest (days 50-140, or mid-October to mid-January) and post-harvest (days 140-202, or mid-January to mid-March, when the last bird left the rice eld landscape), based on Toral et al. (2011), our own remote sensing datasets, and eld observations across years.
We used the static habitat (rice eld landscape scale) map to compare the selection of rice eld habitats vs. non-rice eld habitats and investigate the role of non-paddy habitat features embedded in the rice eld complex, while we used the image-based data subsets for a dynamic analysis of how rice paddy harvest status in uenced the use of the rice paddy interiors by LBBG. For both analyses, we calculated Manly selection ratios for habitat types following Fletcher and Fortin (2018) and using the widesII function in the package AdehabitatHS (Calenge and Basille, 2019). We calculated availability by randomly sampling 10,000 points from our habitat layers using the sampleRandom function in package Raster for classi ed remotely-sensed images (Hijmans et al., 2019) and the over function in package sp for our vector-based static habitat dataset (Pebesma et al., 2018). Availability and selection were calculated on a by-image (thus 24-hour or 72-hour) basis for the dynamic (image-based) habitat dataset. We calculated separate selection ratios by phase of the overwintering season and day versus night, as well as their factorial combination. Manly selectivity ratios were considered signi cant when their 95% con dence intervals did not overlap zero. Because results for dynamic habitat selection analyses were not qualitatively different between our 24-and 72-hour window data subsets, all results for habitat selection among paddies of different harvest status are presented using the 72-hour window subset (~10,000 GPS xes).

Abundance Data
We assessed the abundance of LBBG in the rice eld complex using two independent time series of LBBG abundance. First, we used aerial survey data collected monthly We also analyzed data from weekly daytime (0800 to 1700 hours) counts of gulls that were made via line transects on a boat along the Guadalquivir river from Spring 2008 to Fall 2009 (the only period for which these survey data were available). We selected only those survey observations collected within the rice eld complex and calculated total daily abundances for each survey date during the 2008-2009 overwintering period.

Analyses of Behavior
We analyzed the prevalence of different behaviors within the rice elds and throughout the overwintering season using classi ed accelerometer data (in 2016-17), protocols of focal birds (in 2018-2019) and instantaneous scans in the eld. We used chi-squared tests to detect signi cant relationships between paddy-scale harvest status and behavior prevalence and then used correspondence analysis (using the packages FactoMineR; Husson et al., 2019) to visualize these relationships in low-dimensional space. For behavioral data collected in the eld, behavior prevalence was measured by all gulls engaging in a certain behavior (focal birds), or the number of gulls performing that behavior at given point in time (instantaneous scan). For accelerometer data, behavioral prevalence was the proportion of xes (among all xes that had classi ed accelerometer data, controlled for sampling interval) classi ed as that behavior. We assessed the potential for autocorrelative bias in behavior by repeating calculations with instantaneous scan data including and excluding replicates from the same day and site. For accelerometer data, we conducted analyses separately using our static habitat dataset (landscape-scale, all data points) and our dynamic habitat dataset based on available imagery (paddy-scale, 24-and 72hour subsets).

Remote sensing of changing conditions in rice paddies
Remotely-sensed thematic classes within rice paddies changed progressively in their representation throughout the winter season (Fig. 2). Green vegetation and bare ground were the predominant thematic classes early in the winter season prior to harvest (days ~1-60), with the appearance of harvested rice and ooded paddies as harvest and tilling proceeded starting in early October. The greatest thematic class diversity within ponds was observed between late November and mid-December (day 90-110), where green vegetation, bare ground, dry and wet tilled, and shallow and deep ooded paddies were simultaneously available within the rice eld complex. This time period coincided with peak use of the rice eld complex by GPS-tracked gulls (see Fig. 3) and peak LBBG abundance from aerial surveys (Additional File 3). By late December, most paddies were deeply ooded after harvest, while beginning in January (post-harvest phase) water levels began to fall, resulting in a greater prevalence of shallow ooded and dry tilled paddies.

GPS Tracking Data
Arrival dates to the rice eld landscape ranged from August 27 (day 1) to November 7 2016 (day 73). Space use patterns with respect to the rice eld complex were distinct between the three phases of harvest (pre-harvest, active harvest, and post-harvest; Figs. 2-4). Two gulls did not arrive to the rice eld landscape until the beginning of the active harvest phase, so tracking data are limited to four individuals for the pre-harvest phase. In this phase, individuals spent time in rice eld complex during the day, but tended to roost outside the complex at night, typically in aquatic habitats like arti cial sh ponds at Veta la Palma, natural marshes within Doñana National Park, or the Guadalquivir river.
In the active harvest phase, which started on October 15 (day 50), individual gulls began also spending nights in the rice eld complex. Birds then spent almost the entire 24-hour cycle inside the rice eld complex until December 19-29 (day 115-125), at which point they began commuting with increasing frequency to locations outside of the rice eld complex (Figs. 3 & 4). Notably, these patterns were very consistent across sampled individuals (Fig. 4). These locations were, without exception, land lls at 22-67km distance from the rice eld complex. In December and January, gulls normally returned to the complex at night, but occasionally spent evenings in nearby water bodies, other ooded areas, or sections of the Guadalquivir river that were upstream or downstream of the rice eld complex. Daily distance traveled (including all movements, regardless of direction), averaged across all gulls, was 55.53 ± 3.04km, reaching a maximum on day 201 (March 15; 364.4km) and a minimum on day 71 (5 November, 10.6km). Mean daily distance traveled was greatest during the pre-and post-harvest stages of the overwintering season and reached its minimum during the active harvest phase (Fig. 3). The last tracked gull still present in the study area departed on Spring migration on March 16 (day 202).

Abundance Estimates
Total gull abundance in the nearby Guadalquivir river in the winter of 2008-2009 showed two peaks around days 30 and 150, which corresponds to the timing of transition into and out of the active harvest phase observed in the winter of 2016-2017. Gull abundance in the river reached its lowest numbers during the active harvest phase, when, according to GPS data, all birds were within the rice eld complex and rice paddies (Additional File 3).
Aerial counts of gulls inside the rice eld complex during 2016-2017 followed the opposite trend to those observed in the Guadalquivir river. The number of gulls reached a maximum of 3750 on day 83, remaining high until day 115 and then declining steeply thereafter, corresponding with the pattern we found in the tracked individuals. This corresponds with peak selection and use of the rice eld complex by tracked gulls in our study during the same winter (Additional File 3).

Habitat selection
Static habitat analysis A Manly selectivity analysis showed strong evidence for habitat selection both at the landscape scale (static analysis) and at the population scale (n=6 individuals; 2 >5000, p < 0.00001). Rice paddies were selected signi cantly more than expected by chance at this scale (Table 1). Other habitats were used proportionally to their availability, even when data were subset by day and night (Table 1; Additional Files 4-15). When analyses were further split into the three wintering phases, the Manly selectivity analysis showed that rice paddies were selected during the active-and post-harvest phases but not in the preharvest phase. During the pre-harvest phase, dikes and roads in the rice eld complex were also selected by gulls at the population level (Additional Files 4-15). At the individual level, most LBBG (three out of four individuals during the pre-harvest phase, four of six during the active phase, and six of six during the post-harvest phase) showed very high selectivity ratios for the Guadalquivir river during day and night, but due to large individual differences this was not signi cant at the population level despite a very high magnitude of selection (Table 1).
In the pre-harvest phase, several individuals showed preferences for the Veta la Palma shponds and rice paddies, but similarly, individual differences were large leading to con dence intervals overlapping zero at the population level (Additional Files 4-15). At the landscape scale, all individuals showed strong selection for rice paddies during the active harvest phase, both during the day and at night, but the use of dikes and canals, arti cial ponds, and roads strongly differed between individuals (Additional Files 4-15). The strongest selection for rice paddies was observed at night during the post-harvest phase. Signi cant diurnal selection for rice paddies was still evident in that phase, although the gulls spent most of their time at land lls outside of the rice eld landscape (Additional Files 4-15). Among all static habitat classes, rice paddies were the most used feature, accounting for >79% of xes on the rice eld landscape, with small dikes and canals (7.2%) and non-rice open space (fallow elds and other non-crop undeveloped areas; 5.6%) as the next most used features regardless of availability (Table 1).

Dynamic habitat selection
In our dynamic habitat analysis based on classi ed images of rice paddy interiors (images summarized in Fig. 2), during the pre-harvest phase, gulls showed a signi cant preference for green vegetation and unharvested rice as opposed to bare ground, even as the latter increased in availability (10 September -4 October, Table 2 Table 2; Additional Files [16][17][18]. At this point, gulls showed signi cant selection for dry and wettilled habitats and avoided green vegetation.

Analysis of Accelerometer data
Behavioral time budgets of the gulls remained relatively constant throughout the overwintering season.
Gulls spent the majority (72%) of time resting (behavioral class sit/stand), with terrestrial locomotion (7.8%), apping (6.1%), and feeding and behaviors related to food-searching and foraging (peck, ex ap; total 5.7%) as the next most prevalent behaviors. ODBA showed no clear seasonal pattern, but peaks in ODBA occurred around dawn and dusk. All individuals spent more time on behaviors related to foodsearching, foraging and locomotion (terloco, peck, manouvre, ap, soar) during the day and more time standing still during the night, with increased apping at dusk and dawn (Additional File 19) but with no obvious seasonal trends (Additional Files 19-21; but see daily distance traveled, Fig. 3).

Chi-squared analyses indicated a statistically signi cant association between behaviors and both static
and dynamic (within-paddy) habitat features during the day ( 2 =1417, p < 0.00001) and night ( 2 =820, p < 0.00001). Floating was predictably restricted to aquatic habitats (arti cial wetlands, rivers and streams) and resting occurred in marshlands, roads, and dikes within the rice eld complex (Additional . Food-searching and foraging-related behaviors (terloco/walk, peck) were mainly observed inside rice paddies at all stages of harvest, especially in shallow ooded and wet-tilled paddies. During the active-and post-harvest phases, non-rice open areas were associated with ap and soar behaviors.
During the day, non-rice agriculture and open space were associated with ying and maneuvering, and dikes, canals and paddy interiors were associated with sitting and resting during the entire overwintering season. Shallow ooded ponds were associated with food-searching and foraging-related behaviors (peck and terrestrial locomotion/walk) and arti cial ponds and rivers and streams were associated with oating. During the night, most habitats were associated with sitting and resting, especially roads and dikes and deep-ooded paddies, and rivers and streams and arti cial ponds were associated with oating (Additional Files 22-27).

Field Observations of Behavior
For instantaneous scan data, we found no qualitative difference in results when including replicate site visits in the same day. The number of gulls in an instantaneous scan ranged from one to several hundred, although the scan typically only resulted in the sampling of up to ~70 individuals per scan due to the oneminute time limit. The maximum number of replicate scans for the same site and day was three, and scans conducted per day ranged from 9-34. Ethograms of focal birds and instantaneous scans largely corresponded with one another, showing that the majority of time was spent in resting behaviors (sitting and standing; 58% and 51%, respectively in ethograms and scans). Active behaviors were observed less frequently ( ying 13% and 5%, respectively, walking 4% and 12%, and swimming 9% and 13%). The amount of time spent foraging amounted to 11% and 14%, respectively. Chi-squared analyses indicated signi cant deviations from behavioral patterns expected under random behavior in rice elds (focal follow: 2 =42300, p < 0.00001; instantaneous scan: 2 =913, p < 0.00001), and correspondence analysis of both datasets showed clear associations between ooded elds and bathing, swimming, and preening behaviors (Additional File 28). Partially ooded elds were more used for foraging and walking, while dry elds were associated with ying (travel) and foraging. Standing showed no strong association with any ooding pattern and sitting was more associated with ooded or partially ooded elds (Additional File 28).

Discussion
We analyzed the space use of LBBGs in the Andalusian rice eld landscape and changes in habitat characteristics within rice paddies across time to examine the ne-scale drivers of movement behavior in this important and widespread taxon. Our spatial analyses show clear, consistent, and highly structured changes in LBBG use of the rice paddies throughout the overwintering season. These changes correspond with three phases of the harvest cycle, and the spatiotemporal patterns of abundance. They are likely driven by the availability of spatial resources (i.e., foraging and roost sites), speci cally with respect to the proximity and reliability of food resources, a major driver of LBBG space use in anthropogenic landscapes during the breeding season (Spelt et al., 2019). These patterns were highly consistent among sampled individuals and were strongly supported by eld observations of abundance and behavior during the same and different years. The use of and behavior in rice elds showed strong differences between day and night (diurnal feeding and nocturnal resting) as well as the between the different phases of harvest at the landscape scale. Rice paddies were used more often when recently harvested and when ooded. Gulls used recently harvested paddies for daytime foraging and loa ng, which parallels Spelt et al. (2019)'s observations that LBBG sought terrestrial prey after plowing and other disturbance in other types of agricultural elds. Rice paddy interiors were most used during the active harvest phase when a wide range of paddy conditions occurred, making both foraging and resting habitats simultaneously available. During that period, all birds spent 100% of their recorded time (day and night) within the rice eld complex. Daily movements were strongly reduced during that phase, as birds reduced commuting time by nding all of their resources within the rice eld complex (Figure 3 Aerial and boat-based surveys complemented the patterns observed through GPS telemetry and showed peak abundances of gulls inside the rice eld complex when rice paddy interiors showed the greatest diversity of available conditions (active harvest phase), while gulls stayed more in the rivers when only one habitat type was prevalent (pre-and post-harvest phases). These ndings are consistent with previous studies at a broader scale which found that gull numbers peak at harvest time as birds exploit cray sh (Procamburus clarkii) and other resources that are exposed during and immediately after harvest Our study shows that LBBG traveled daily to land lls most often during the post-harvesting phase and that they did so during daylight, suggesting that their role as biovectors in rice elds is largely restricted to this phase and that deposition of potentially harmful materials in areas of human concern occurs during roosting, which is predominantly at night. Indeed, at the Andalusian scale Martín-Veléz et al. (2020) showed that LBBG strongly increased their time spent at land lls, leading to increased connectivity between land lls and other habitats such as rice elds during the same time period across several overwintering seasons. The present study illustrates the ne-scale dynamics behind this pattern, including a shift to using rice elds almost exclusively for roosting and land lls for foraging. This pattern could result in major in uxes of other contaminants (e.g., heavy metals and plastics) into elds where food for human consumption will be grown throughout the following spring and summer, a phenomenon observed in other roosting areas for land ll-feeding gulls (

Conclusions
We found that overwintering LBBG occupying an anthropogenic rice agriculture landscape change their movement patterns and space use over the course of the winter according to the habitat resources available in rice-growing paddies across a winter harvesting cycle. Behavioral observations in the eld and using classi ed accelerometer data indicated that these changes in space use were to nd habitats for food acquisition and resting, and that metabolic needs and behavioral time budgets did not change over the course of the winter. LBBG's rapid increase in anthropogenic landscapes and potential for transporting diverse materials including agents potentially harmful to human health makes a mechanistic understanding of their movement behavior important and relevant to management and conservation.   Figure 1 A false-color infrared photograph of the study area, showing the rice eld landscape (the total area for which all land cover types were classi ed for our static habitat analysis), the rice eld complex, and the interior of the rice paddies (where within-paddy habitat availability was assessed using remotely-sensed imagery). The protected wetlands of Doñana National Park and the nearby Veta la Palma shponds are also shown. The availability (percent cover) and use (% gull daily time within) of rice paddies in different conditions across the course of the season, based on classi ed remotely-sensed images of the rice paddy interiors. Both availability and use are assessed at the paddy scale; see Table 1 for landscape-scale habitat use.

List Of Abbreviations
The phases of the gull overwintering season, demarcated by changes in commuting behavior, are shown at the top, and months of the year below the Y-axis. Day after arrival is based on the time since the arrival of the rst bird to the rice eld landscape (27 August).  The hourly proportion of time spent in the rice eld complex averaged across all birds for each day of the overwintering cycle. Birds switched from spending daylight hours in the complex and roosting elsewhere in the rice eld landscape (pre-harvest phase), to spending 100% of their daily cycle in the complex (active harvest phase), to roosting in the complex but spending daytime hours outside (post-harvest phase). Day after arrival is based on the time since the arrival of the rst bird to the rice eld landscape (27 August).