Machine learning reveals hierarchical spatial patterns in salt marsh ditching along U.S. Atlantic Coast

Machine learning reveals hierarchical spatial patterns in salt marsh ditching along U.S. Atlantic Coast | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Machine learning reveals hierarchical spatial patterns in salt marsh ditching along U.S. Atlantic Coast Karen Aerni, Tom W. Bell, David L. Kimbro This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6387803/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 5 You are reading this latest preprint version Abstract The loss of salt marshes and their ecosystem services following anthropogenic disturbances necessitates restoration built on a scale-dependent understanding of how the prevalence and intensity of these disturbances are linked to ecosystem functioning. A conspicuous legacy modification of marshes, which lacks a standardized and scale-able assessment, is ditching. Consequently, U.S. Atlantic coast resource managers must devote limited resources to quantifying local-scale ditching or make restoration decisions based on a literature of subjective ditching assessments (low vs. high) from a subset of locations with contradictory impacts to ecosystem functions. Here, we combined freely available satellite imagery with machine learning to generate a multi-scale database of ditching prevalence and intensity in 634 marshes from Maine through Florida. The algorithm consistently detected ditches despite the heterogeneous appearance of this disturbance and marshes across regions, seasons, and tidal stages. In contrast to the oft-quoted historical ditching prevalence of 90%, the algorithm quantified a much lower current average of 38%, with the size of this discrepancy varying regionally from an average prevalence of 87% in the Gulf of Maine to 20% in the South Atlantic Bight. Ditching intensity showed further hierarchical spatial variation, but at the state and within-state levels, as opposed to regional level. Within regions, intensely ditched states (5% area removed) were opposed by mildly ditched states (1.9% area removed). With this standardized database of ditching prevalence and intensity, researchers and resource managers may now conduct scale-dependent assessments of ecosystem responses to ditching to inform restoration and management of this valuable habitat. anthropogenic impacts coastal wetland disturbance habitat mapping landscape ecology algorithm marsh restoration mosquito ditching remote sensing salt marsh Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction Salt marshes provide important and economically valuable ecosystem services worldwide by buffering shorelines from coastal erosion, filtering water, sequestering carbon, serving as a refuge for biodiversity, and providing space for recreation (Barbier et al. 2011 , Mcleod et al. 2011 , Temmerman et al. 2013 , Costanza et al. 2014 ). Despite their value, thousands of acres of salt marsh are lost annually due to anthropogenic activities like land conversion, eutrophication, and hydrological alterations (Dahl 2011 , Gedan et al. 2011 , Campbell et al. 2022 ) resulting in a 50% decrease in salt marsh aerial extent globally over the past century (Kennish 2001 , Davidson 2014 ). Additionally, because tidal inundation is an important driver of salt marsh ecosystem structure and functioning (Daiber 1974 , Gedan et al. 2009 ), marsh loss has been exacerbated in recent years by sea level rise, with future U.S. and global losses predicted to reach 20–60% during this century, respectively (Titus 1988 , Nicholls et al. 2007 , Watson et al. 2017 ). To prioritize effective marsh conservation and restoration, we need a better understanding of the distribution, extent, and degree of anthropogenic modifications within salt marshes. One expansive and potentially important anthropogenic manipulation of salt marshes is ditching. Historically, ditches were created as early as the 16th century to increase drainage and facilitate agriculture, some of which were maintained over the years and remain today; see Adamowicz et al. ( 2020 ) for a detailed history of agricultural ditching. Additionally, thousands of ditches termed “mosquito” or “parallel grid” ditches were constructed throughout marshes along the U.S. Atlantic Coast, primarily during the early 20th century, both to provide jobs as part of the New Deal and to reduce mosquito populations by facilitating predatory fish access to and draining mosquito breeding areas on the marshes (Glasgow 1938 , Crain et al. 2009 ). As a result, ditches occur in much higher concentrations than natural tidal channels such as creeks. Ditches also starkly differ from natural creeks because they are rigidly linear, narrow, and are not bordered by gently sloping mudflats which are important for ecosystem functioning (Nixon and Oviatt 1973 , Pethick 1992 , Bouma et al. 2016 ). Despite their presence in marshes for over a century, and regardless of the motivation behind their creation, the impacts of ditches on most aspects of marsh ecosystem functioning remain uncertain. For instance, while some research found that ditches drain marshes and thereby reduce ponding and tidal creek density (Lathrop et al. 2000 , Adamowicz and Roman 2005 , Smith and Pellew 2021 ), the geomorphic and hydrologic outcomes of other studies ranged from no or subtle impacts (Redfield 1972 , Corman et al. 2012 ) to increasing certain kinds of ponding (Smith et al. 2021 ). Similarly, the outcomes of biological responses studied to date comprise negative, positive, and negligible responses by birds, nekton, and invertebrates to ditches (Nixon 1982 , Clarke et al. 1984 , James-Pirri et al. 2012 , Tonjes 2013 ). Understanding this variability in ditching impacts may facilitate the success of on-going marsh management and restoration, including ditch remediation efforts that are increasingly being implemented (Adam 2019 , Burdick et al. 2020 ). Variability in ditching impacts may be the result of two key issues. First, the effects of ditching have not been systematically evaluated throughout the broad latitudinal range of salt marshes, which varies substantially in physical and biological conditions such as tidal range, winter ice coverage, and herbivory pressure (Pennings and Bertness 2001 , Marczak et al. 2011 , McCall and Pennings 2012 ). This regional environmental variability could be interacting with ditching to produce the observed inconsistent results among local-scale assessments scattered throughout the broad latitudinal range. Ditches are thought to occur in greater than 90% of marshes from Virginia to Maine (Bourn and Cottam 1950 ), though this estimate has not been rigorously confirmed or extended to the remaining one-third of the U.S. Atlantic coastline. The second, and non-mutually exclusive, issue concerns the previously unassessed variability in the intensity of ditching. In the majority of the published studies on which our current understanding is based, logistical constraints around measuring these sizeable features across landscapes necessitated that ditching be quantified as present versus absent or subjectively categorized as low or high intensity. Such categorizations may not only mask important non-linear or threshold ecosystem responses to ditching intensity, but also mask important regional variation in non-linear thresholds. Consequently, several individual studies have worked to objectively quantify ditching intensity and evaluate ecosystem responses (Correll et al. 2017 , McGarigal et al. 2017 ). However, these patch- and local-scale assessments could be bolstered by a comprehensive picture of spatial variation in the prevalence and intensity of ditching that can unravel any scale-dependent impacts of these historical modifications, as well as whether and how to mitigate them. Remote sensing is a promising tool for measuring landscape features across broad areas such as the 17 degrees of latitude spanned by U.S. Atlantic Coast salt marshes. This technique has been successfully applied to monitor wildfires across forest ranges, temporal changes in kelp cover along an entire coast, and sea surface temperatures across ocean basins (Minnich et al. 2000 , Bell et al. 2023 , Minnett et al. 2019 ). But successfully using remote sensing across a broad geographic range requires full spatial coverage by consistent data of a single type (e.g. multispectral imagery or lidar data). Unfortunately, spatial coverage of remote sensing data is lacking in some areas, can be cost prohibitive in others, and is generally at too coarse a resolution (10 to 30m) to detect small or narrow (< 1 m wide) features such as ditches (Woodcock et al. 2008 , Romijn et al. 2012 , Turner et al. 2015 ). One solution to these availability hurdles may be the ubiquitous, no-cost, high-resolution color satellite imagery from Google Earth. Indeed, Google Earth imagery is available for the over 600 salt marshes that vary in area from 1 to 100 km 2 between Maine and Florida. Thus, we hypothesized that this affordable remote sensing approach could be used as a first-step towards addressing the multi-scale knowledge gap regarding ditches. Given the vast amounts of information involved, translating remote sensing results into a database of ditching prevalence and intensity requires novel data processing techniques that can achieve segmentation (locating features in space) and overcome the very high color variability in the visual appearance of Google Earth sourced satellite imagery. This variability is particularly high for marshes, which vary latitudinally in marsh structure, seasonally in vegetation, and daily in tidal stage. Additionally, ditches themselves vary in length, width, orientation, and whether they intersect or parallel other ditches. While machine learning (ML) techniques such as clustering, classification and regression trees, and convolutional neural networks (CNNs) are promising (Maxwell et al. 2018 , Wüest et al. 2020 ), they struggle with segmentation tasks (i.e. ‘where is there a cat in this image’ not just ‘is there a cat in this image’) and their accuracy can be hindered by high superficial variability in imagery. However, a recent advancement in CNNs with substantial implications for remote sensing analysis of variable environments is the development of the U-Net architecture (Ronnenberger et al. 2015). U-Nets have long been successfully applied to biomedical imaging tasks and are now being considered for environmental image segmentation tasks (Du et al. 2020 , Li et al. 2021 ). In addition to allowing for segmentation, one of the most valuable aspects of U-Nets is that they reduce the amount of training data required to overcome cosmetic image variability by several magnitudes via data augmentation and skip connections. Consequently, the heterogeneous salt marshes and ditches of the U.S. Atlantic coast offer an ideal system for testing the utility of pairing U-Nets with no-cost satellite imagery from Google Earth to quantify multi-scale geographic variation in the prevalence and intensity of ditching. In this study, we harnessed machine learning and remote sensing to resolve uncertainty regarding the full spatial extent and variability of ditching. To accomplish this, we trained a U-Net CNN to detect ditches from satellite imagery. We then applied the trained algorithm to satellite imagery of all salt marshes along the U.S. Atlantic Coast greater than 1 km 2 to produce a multi-scale database of salt marsh ditching. Using this database, we then quantified variation in ditching prevalence and intensity at the regional (Gulf of Maine, Mid-Atlantic, South Atlantic), state, and within-state scales. This database will facilitate future studies into the contemporary impacts of this historical habitat modification and in turn help guide the prioritization and planning of marsh management and restoration efforts. Methods and Materials Study Area This study focused on salt marshes from the habitat’s northernmost U.S. range edge in Maine (45° N) to its southernmost range edge in Florida (28.5° N). Throughout this range, we identified 634 marshes greater than 1 km 2 that totaled to ~ 7600 km 2 across 13 states and three biogeographic regions (Fig. 1 a). At local scales, salt marsh communities are predominantly structured by tidal inundation along an elevation gradient where competition between plants dominates the more tolerable conditions at higher elevations and competitively inferior plants are displaced to the more stressful waterlogged conditions at lower elevations (Bertness 1991 , Levine et al. 1998 , Pennings et al. 2005 ). However, this organization across tidal elevation varies latitudinally because of regionally specific abiotic and biotic drivers (Pennings and Bertness 2001 , Marczak et al. 2011 , McCall and Pennings 2012 ). From 2000 to 2020, the Gulf of Maine (GOM) had the harshest winter conditions with mean daily temperatures of -6°C in Maine (compared to 16°C for Florida) and substantial ice cover throughout winter ( https://www.weather.gov/ ). In addition to smothering plants, ice shears off and translocates sediment and senesced grasses leading to patchier vegetation patterns (Niering and Warren 1980 , Ewanchuk and Bertness 2003 , 2004 , Argow et al. 2011 ). In contrast, summer is the harshest season for the South Atlantic Bight (SAB). There, plants grow year-round, but summer (July) mean daily high temperatures of 32°C (compared to 25°C in Maine), lead to elevated evapotranspiration rates, high soil salinity, and the increased formation of salt pans ( https://www.weather.gov/ ). Meanwhile, the Mid-Atlantic Bight (MAB) has a mix of these extremes at its opposite ends. Finally, while tidal range decreases (nonlinearly) from 5.6 m in GOM to 1 m in SAB ( https://tidesandcurrents.noaa.gov/ ), herbivory on marsh plants steadily increases (Smith and Odum 1981 , Silliman and Zieman 2001 , Pennings et al. 2009 ). In addition to these factors, marsh size and a diversity of other anthropogenic impacts such as eutrophication (Deegan et al. 2012 ) could individually and/or collectively interact with spatial patterns of ditching prevalence and intensity at multiple scales. Salt marsh identification and image acquisition To collate images of marsh habitat, we acquired wetland shapefiles from the National Wetlands Inventory for all 13 U.S. states bordering the Atlantic Ocean. We filtered these wetlands to coastal marshes using the ‘Select Attribute by Layer’ tool in ArcGIS Pro 2.8.0 (ESRI, Redlands, CA) to retain polygons classified as Estuarine Intertidal Emergent habitat (“WETLAND_TY = ‘Estuarine and Marine Wetland’ And ATTRIBUTE LIKE %EM%”), which includes the subclasses (numeric codes) of Persistent (1), Non-persistent (2), or Phragmites australis (5) regardless of any designated modifier codes. To assemble disconnected polygons into cohesive marshes, we aggregated individual polygons within 150 m of each other. We evaluated the effectiveness of this method for several New England marshes with which we are familiar and found that this distance successfully connected polygons separated by small creeks within individual marshes. Furthermore, this distance threshold did not connect polygons between adjacent marshes separated by creeks and upland habitat. To retain a manageable number of marshes (under 1000), we then filtered the aggregated marshes to those greater than 1 km 2 . Marsh polygon aggregation and filtering for size was performed in a single step using the ‘Aggregate Polygons’ Tool with ‘Aggregation Distance’ set to ‘150 Meters’ and ‘Minimum Area’ set to ‘1 SquareKilometers’. Finally, we drew a selection polygon along the state borders and copied the selected features to a new layer for subsequent use, thereby avoiding duplicated marshes since NWI state shapefiles had polygons extending into neighboring states. The preceding processing steps produced a database of 634 marshes which we converted to individual KML files of each polygon’s boundaries. KML files were favored over bounding boxes because adjacent marshes were often separated by a few hundred meters and, therefore, bounding boxes would have overlapped substantially and unnecessarily created repeated assessments of the same marsh images. High spatial resolution (< 0.5m) images covering the full extent of each marsh as viewed from an altitude of 400 m were downloaded from Google Earth, and only those images that contained a portion of the marsh that fell within the KML boundaries were retained for analysis. Since the algorithm was intended to handle heterogeneity, only one image was used to represent each marsh portion with no restrictions around the color properties of the marsh at that timepoint; however, to consistently measure current ditching among marshes, images were restricted to being from 2016 or later, a similar image quality, and were not from peak high tide which could make ditches appear wider. Algorithm development and ditch detection Algorithm structure and training To analyze marsh images for ditching metrics, we created a U-Net deep convolutional neural network (CNN) with seven convolutional layers: three compression layers, one representation layer, and three expansion layers. U-Nets are a special case of CNNs that allow for image segmentation because the network “funnels” information to a low-dimensional representation space and then “expands” back out to reproduce the image (thereby resembling a U shape) with features located. In this study, each layer of the CNN consisted of a stack of two 3-pixel x 3-pixel 2D convolutions followed by either a max pooling operation in the compression section or an up-sampling operation in the expansion section. Note that the pixel resolution in the initial compression layer matches that of the original image but decreases with subsequent compression layers. All layers used the ReLU activation function, except for the final output which used a sigmoid activation function. The output of the algorithm is a grayscale image with values between 0 and 1 representing the confidence that a given pixel is part of a ditch. Because the input size of the network was 256 x 256 pixels, we implemented a tiling algorithm that split individual satellite images into multiple tiles. All coding and machine learning were performed with Python (version 3.10.6) using the keras and tensorflow libraries (versions 2.9.0 and 2.10.0, respectively) on Google Colab. To train the algorithm, we generated a training dataset by handpicking 12 marshes from GA through ME that captured a range of variation in marsh appearance or color – such as might arise from latitudinal or seasonal variation – from which we haphazardly selected and labeled 262 images using Label Studio (version 0.9.1) with polygon segmentation labels. For each image, polygons were drawn around individual ditches and assigned a ditch label. As a result, a ditch was represented by an area that reflected their length and width. We created binary masks from these polygons where pixels within the ditch polygon are assigned the value 1 and all other pixels are set to 0. These labeled images and corresponding masks were then randomly partitioned into training data (90% of the labeled images) and validation data (10% of the labeled images) and subsequently split into 20 tiles of dimensions 256 x 256 pixels to match the U-Net's input size. Before proceeding with training, we checked the training and validation data sets to ensure that both of the randomly generated sets contained marsh images with ditches present and absent. During learning – accomplished through stochastic gradient descent – the U-Net made prediction masks for each image, compared that prediction to the true labeled mask and updated its learned weights each iteration. Loss was calculated using a binary cross-entropy score for each pixel in the predicted mask and summed across the full tile. Binary cross-entropy scores factor in how confidently the U-Net classified a pixel as ditch/not-ditch. Pixels incorrectly predicted with high confidence were penalized strongly and the penalty was fed back to the model using back-propagation to update model parameters. We trained the algorithm for 100 epochs using the ADAM optimizer (Kingma and Ba 2014 ). After reviewing the validation performance by plotting loss as a function of training time, we observed that although the model performed well on finding ditches, it sometimes incorrectly labeled other small water features or steep vertical creek banks as ditches. To improve algorithm performance, we hand-labeled 50 of the 262 ditch-labeled images with the two other water features typically seen in marshes – creeks and ponds. We then fine-tuned the algorithm by updating the weights to include a penalty for non-ditch water features (creeks and ponds) and training for 5 epochs at a smaller learning rate of e − 5 . The penalty was a mask that assigned a -10 weight to creeks and ponds, which was multiplied by the original algorithm’s output predictions such that a misclassification of a pixel as a ditch (value of 1) would then carry a -10x penalty. Note that more and fewer epochs were tried and all predictions visually evaluated to ultimately determine that 5 epochs was optimal. Ditch prevalence and intensity The trained and validated algorithm was used to analyze all 634 marshes for ditching prevalence and intensity. But these whole marsh assessments first required processing all the individual constituent images of a single marsh, which, in turn, required tiling each individual marsh image into 20 individual tiles measuring 256 x 256 pixels for input into the algorithm. The algorithm generated predictions for each individual tile using the aforementioned confidence-informed pixel classification from 0 through 1. To convert this prediction into a binary mask, we applied a threshold to each prediction of 0.35 such that every pixel less than 0.35 became a 0 (no ditch), and values equal to or greater than 0.35 became a 1 (a ditch). This value of 0.35 was chosen initially by a visual inspection and was later confirmed using a linear regression trained on the validation data. This thresholding step minimized overestimation by filtering out pixels that did not strongly indicate a ditch. An entire marsh was scored as having ditches present if any of the 20 tiles of any constituent marsh image contained pixels classified as ditch (pixel value of 1). The prevalence of ditched marshes was then calculated for each state, region, and the entire Atlantic coastline by calculating the ratio of ditched marshes:total marshes at each spatial scale. To determine ditching intensity of each marsh, we summed the total number of ditch pixels within each tile across all the constituent images of that marsh and converted that count into m 2 based on the spatial resolution of the images. For each marsh, the resulting total amount of ditch area was then divided by total marsh area (generated during earlier GIS processing) and multiplied by 100 to convert the ratio into a percent of marsh area ditched. If a marsh had no ditches present, its intensity was not recorded or included in the analysis of ditching intensity. The algorithm overestimated ditching intensity by around 1%, so we individually verified each marsh for ditching presence to generate an accurate inventory of prevalence. Comparative performance For a further evaluation of the algorithm’s accuracy, we compared our results to two existing salt marsh ditching datasets following the methods described in more detail in Online Resource 1. The first existing dataset served as the expert classification and comprises ditches in New Hampshire marshes that local experts hand digitized from aerial photos (Eberhardt and Burdick, 2008). To initiate this first comparison, we georectified twenty Google Earth images of the same New Hampshire marsh into QGIS (Online Resource 1). Because the layer of hand-digitized ditches consisted of line features (as opposed to polygons), we converted our algorithm results into a line feature format by hand drawing lines over the predicted ditches, thereby converting the ditch area predictions of our algorithm into comparable ditch lines and length estimates (Online Resource 1, Fig. S1). These lines were partitioned into two separate layers: one for true positives where our predictions overlapped with hand-digitized ditches and one for false positives where our predictions did not overlap with hand-digitized ditches. For each image, we summed the lengths of hand-digitized ditches (local expert layer), our true positive predictions, and our false positive predictions. Note that calculating true negatives was not possible for this analysis. Next, we evaluated the relative ditch-detecting performance of our algorithm against an existing regional (Virginia to Maine) computer vision-generated dataset of ditches (McGarigal et al. 2017 ), by comparing each to the hand-digitized layer of New Hampshire marshes created by local experts (Eberhardt and Burdick 2008). Similar to the first comparison, ditch predictions from the algorithm and the existing regional layer (McGarigal et al. 2017 ) were partitioned into two separate layers: one for true positives where each dataset’s predictions overlapped with hand-digitized ditches and one for false positives where predictions did not overlap with hand-digitized ditches. For each image, we summed the lengths of hand-digitized ditches (local expert layer), as well as the length of true and false positive predictions for our algorithm and the existing regional layer. Statistical Analysis We checked for algorithm overfitting by plotting its loss on the training dataset compared to the validation dataset. We then visually assessed model performance from predictions overlaid on the original aerial images. After confirming that overfitting was absent and that performance was accurate, we calculated mean (± sd) prevalence and intensity of ditching averaged across all three regions and 13 states. Next, we used an analysis of variance (ANOVA) to test for differences in ditch prevalence among the three regions, using state prevalence as replicates. The analysis of prevalence at a scale smaller than region level was not possible because there is only one prevalence value per state. In contrast, we used a Nested ANOVA to test for differences in ditch intensity among both regions and states, with state being nested within region and individual marshes serving as subsamples of states. Following a significant overall model (p < 0.05), we used Tukey’s Honest Significance Difference (HSD) to conduct post hoc mean comparisons of the prevalence and intensity results. For the Nested ANOVA, we also quantified the variance components using the ‘VCA’ package (version 1.4.5) to estimate the amount of variation explained at each spatial scale. Before conducting the ANOVAs, parametric assumptions of normally distributed residuals and homogeneous variances were tested with Shapiro’s test and Levene’s test, respectively. Ditching intensity values were square-root transformed to meet assumptions of normality and homoscedasticity. To compare the performance of the algorithm with an existing local-scale layer of ditches, we present the mean (± sd) percent of correctly identified ditches (averaged across 20 images) according to the total length of ditches quantified through hand-digitization by local experts. Of the 20 images, we also present the minimum (worst) and maximum (best) percent of correctly identified ditches by the algorithm. To assess the comparative performance of an existing regional layer, we generated similar summary statistics for predicted ditches of the same 20 images based on McGarigal et al. ( 2017 ). Results Algorithm training We successfully trained a U-Net to identify and locate ditches from satellite imagery (Fig. 2 a). The algorithm accurately distinguished ditches from other potentially similar features such as creeks or roads. Based on the validation data, the algorithm successfully learned (Fig. 2 b), as the binary cross entropy calculated loss steadily decreased and ultimately stabilized, indicating that learning potential was at maximum capacity. Importantly, the plot of loss did not show signs of overfitting, since training loss did not decrease sharply after stabilization (indicating pixel memorization). Equally important, validation loss did not synchronously increase after stabilization (indicating failure to generalize to novel images). To numerically evaluate the algorithm’s accuracy, we averaged ditching intensity estimates for the 395 marshes that were manually verified as zero intensity, and found the median overestimation for marshes with no ditching present was 1.2% (Fig. 2 c). Ditch prevalence and intensity We documented an overall ditching prevalence of 37.7% (Fig. 3 b), which is substantially lower than the previously reported 90%. However, this overall ditching prevalence masks significant regional variation among the three regions (Fig. 3, F 2,11 = 6.15, p < .05), with the mean ( ± sd) prevalence of both GOM (87.4% ± 19.1%) and MAB (66.7% ± 34.3%) being significantly greater than the previously unassessed SAB (20.3% ± 4.9%, TukeyHSD, p < .05). Based on coefficients of variation (CV), states within GOM (CV = 21.9) and SAB (CV = 24) were more homogeneous in ditching prevalence (with GOM being homogeneously ditched and SAB being homogeneously unditched) than states in MAB (CV = 51.4), though all regions exhibited substantial variation in ditch prevalence across states (Fig. 4 ). Only five states had ditching prevalence of approximately 90% or higher, including New Hampshire (100%), Massachusetts (96.9%), Rhode Island (100%), Connecticut (92.3%), and New York (86.7%). Virginia and Florida showed the lowest prevalence with only 15.8% and 16.2%, respectively. The remaining states had between 22–83% of their marshes ditched (Fig. 4 ). Because the algorithm successfully quantified ditching intensity, we were able to generate more resolution on the multi-scale patterns of ditches. Across the U.S. Atlantic coast, for marshes with ditches present, the mean (± sd) intensity (percentage) of marsh area composed of ditches is 3.8% (± 2.6%), with a maximum observed intensity of 14.1%. Unlike prevalence, we did not detect significant regional variation for intensity (nested ANOVA, F 2,11 = 1.94, p = 0.15, Fig. 5 a). The absence of significant regional variation is likely due to significantly high variation among states within each region (nested ANOVA, F 11,225 = 6.04, p < .001, Fig. 5 b). For example, GOM (regional mean ± sd intensity: 4.1% ± 2.7%) includes NH (5.1% ± 1.9%) and MA (4.8% ± 2.6%) with significantly higher intensity than ME (2.1% ± 1.5%; TukeyHSD, p < 0.05). Meanwhile, similar regional means in MAB (3.95% ± 2.7%) and SAB (3.4% ± 1.7%) masked a significant difference between NY (5.2% ± 3.8%) and VA (1.4% ± 0.5%; TukeyHSD, p < 0.05) in the former and NC (3.3% ± 2.1%) and FL (0.9% ± 0.9%; TukeyHSD, p < 0.001) in the latter (Fig. 5 ). Consequently, all three regions had coefficients of variation from 58–73%. Interestingly, the four states with the highest levels of ditching intensity (New York = 5.6%, New Hampshire = 5.1%, Massachusetts = 4.9%, Connecticut = 4.6%) occurred in the northern two regions. However, in contrast with those findings, the five states with the lowest mean intensities (Florida = 1.25%, Rhode Island = 1.37%, Virginia = 1.87%, Delaware = 2.18%, Maine = 2.35%) occurred throughout the U.S. Atlantic Coast. After calculating the variance components of the Nested ANOVA, we found that region only accounted for 1.5% of the total variation in ditching intensity, whereas state accounted for 22.5% of the total variation, while the remaining variation (76.2%) is attributable to within-state variation and error (Fig. 5 b). Comparative performance Of the ditches in a single New Hampshire marsh that were hand digitized by Eberhardt and Burdick (2008), our algorithm captured 69% ( ± 6.5%) of ditch length (true positives). Relative to the existing regional layer of ditches by McGarigal et al. ( 2017 ), the algorithm correctly identified 230% more of the true ditches because the existing regional ditch layer correctly identified 21% (9.1%) of ditches. Meanwhile, the false positive rate of the algorithm was relatively low 5.2% (3.7%). Discussion In this study, we successfully assessed ditching prevalence and intensity at multiple scales in salt marshes along the U.S. Atlantic Coast by applying machine learning to publicly available satellite imagery. In our assessment, the average prevalence of ditched marshes is much lower than previous reports in the literature (Bourn and Cottam 1950 ). But the size of this discrepancy depends on region, with the highest prevalence in the Gulf of Maine decreasing to lowest prevalence in the South Atlantic Bight. For marshes with ditches present, the amount of marsh altered by ditches did not vary regionally. In contrast, spatial variability in ditching intensity varied substantially within regions and states, where intensely ditched states in some regions are opposed by milder ditched states within the same region. Importantly, at least one state in each region has marshes that vary substantially in intensity, allowing future studies to examine functional responses to ditching intensity across scale-dependent biotic and abiotic drivers of salt marsh communities. In addition to demonstrating that a long-held narrative (“nearly all U.S. Atlantic Coast temperate marshes are ditched” (Tonjes 2013 )) depends on scale, this blending of ML with freely available data provides a comprehensive and valuable resource to further the conservation and restoration of salt marshes as well as the study of scale in macro- and community ecology. Ditching prevalence and intensity Across the latitudinal range of this temperate habitat along the U.S. Atlantic coast (spanning 17 degrees and 2000 km), there is little support for the previously held assertion that nearly all marshes north of Virginia have been modified by ditches, because nearly 50% of the large salt marshes (area > 1km 2 ) north of Virginia evaluated in this study actually lack ditches. Without this updated and larger-scale assessment, pursuing research on how the presence/absence of ditches affects diversity and ecosystem functioning in the Gulf of Maine and Mid-Atlantic Bight may not be sufficiently prioritized, because of the incorrectly presumed absence of control (unditched) marshes to which ditched marshes can be compared. Furthermore, the historical baseline provided by Bourn and Cottam ( 1950 ) is geographically restricted. By updating this baseline with information on the South Atlantic Bight, we now know that 22% of the marshes in this region have been ditched and therefore could also be considered for research and mitigation. Our results suggest that any investment decisions around research and mitigation of ditches would benefit from understanding two aspects of the spatial variability in ditches. First, the regional variation in ditching prevalence suggests that ditching effort (and/or recovery) was heterogeneous, with only GOM approaching high prevalence of 90%. Interestingly, the actual number of ditched marshes is relatively homogeneous throughout the study area (Fig. 4 b), suggesting that regional variation in prevalence is influenced at least in part by the increase in number of marshes from the GOM through the SAB due to geomorphological differences (Chapman 1960 ). Second, if we control for regional differences in overall marsh abundance by comparing prevalence among states within each region, we find significant spatial structure (Fig. 5 ). For instance, GOM and MAB both have states with prevalence rates above 90% (NH/MA and CT, respectively) that are balanced by states with lower prevalence such as ME (65%) and VA (6%), respectively. Furthermore, and in contrast to a regional evaluation that suggested a paucity of unditched or control marshes (Crain et al. 2009 ), our results demonstrated the presence of unditched marshes in every state except RI (MAB) and NH (GOM), which have a paucity of marshes greater than one km 2 in general (2 and 6, respectively). Consequently, there are sufficient ditched and unditched marshes within each region to robustly evaluate the impacts of ditching presence on diversity and ecosystem functioning, while controlling for background regional environmental differences. Similar to the study of disturbance in ecology and disease in epidemiology, the biogeography of prevalence in an anthropogenic modification of marshes will likely be insufficient for optimizing future research as well as restoration and management decisions (Brown and Vivas 2005 , Liu et al. 2007 , McClintock et al. 2010 ). This is because, whether it be key relationships between pathogen virulence and transmission or biodiversity and nutrient cycling, natural systems are replete with non-linear functional responses characterized by thresholds and critical tipping points (Holling 1973 , Scheffer et al. 2001 , Liu et al. 2007 ). By quantifying the full range of ditching intensity, our research laid the groundwork for detecting and understanding any such non-linear relationships. At the largest scale, we found the mean intensity of ditching (excluding unditched marshes) across the U.S. Atlantic coast to be 3.8%, greatly exceeding the only existing intensity estimate of approximately 1% for New England marshes (Crain et al. 2009 ). Nested within this large-scale estimate is the existence of hierarchically organized spatial patterning in ditch intensity. At the regional scale, ditching intensity did not differ significantly, accounting for little of the variance (1.5%). But within each region, we found substantial variation (22.5%) in ditch intensity (Fig. 5 ), with all three regions having states with high, moderate, and low intensities, as well as unditched marshes. There is also considerable variation at the next hierarchical level, where almost all states exhibit large within-state variation except for ME in GOM, RI and VA in MAB, and FL in SAB (Fig. 5 b). However, to determine exactly how much of the remaining variation in ditching intensity (76.2%) is due to within-state variation compared to error would require subsampling within marshes, which is worthy of future study. Consequently, our approach not only laid groundwork for detecting non-linear relationships, but it did so in a manner that promotes scale-dependent assessments and thus addresses a central challenge in ecology (Levin 1992 , Allen and Hoekstra 2015 , Jackson and Fahrig 2015 ). This research also paves the way for disentangling how ditches interact with established and projected environmental forcing of salt marsh community structure. By targeting sites encompassing the full gradient of ditching intensity, researchers in GOM could evaluate interactions with ice disturbances in the winter and wrack abundances in the spring. Meanwhile, in SAB, researchers can look at interactions with salt pans and herbivory pressure. Across all regions, researchers can evaluate the influence of tidal range, which has previously been proposed as determining the hydrological impacts of ditching (Taylor 1938 , Redfield 1972 , Tonjes 2013 ). It should also be noted that not all ditches drain equally, as some have become plugged over time and now hold water, which is another factor that researchers can address from this starting point. Equally important, researchers can investigate how ditching interacts with novel and increasingly influential drivers of salt marsh change. As with other disturbances, the legacy effects of ditching can impact the ecosystem’s response to successive disturbances, leading to ecological surprises and catastrophic shifts (Foster et al. 1998 , Paine et al. 1998 , Davies et al. 2009 ). Threats of particular concern include accelerated sea level rise in MAB (Sallenger et al. 2012 ), abundant tidal restrictions in GOM (Crain et al. 2009 ), and increasing range expansion of southern latitudinal predators and consumers such as blue crabs (Johnson, 2015 ) and fiddler crabs (Martínez-Soto and Johnson, 2020 ). Not only does this help us allocate research to understand ditching disturbance across abiotic and biotic forces at multiple scales, we can also strategically distribute ditch restoration experiments, thereby optimally informing and promoting marsh resilience. Detection of ditches by a U-Net CNN The algorithm developed in this study exceeded performance expectations (Fig. 2 ) and showed the applicability of using machine learning for regional and fine scale mapping of features within highly variable habitats. In general, the algorithm detected all ditches in an image and only infrequently missed a ditch, though it sometimes underestimated the length of a ditch. While it accurately detected and estimated ditches, it also minimally detected “ditches” where there were none. This was largely filtered out in the thresholding step, but not entirely, leading to all measurements of intensities being slightly overestimated by approximately 1% (Fig. 2 c). It is important to note that this overestimation can be partly attributed to plowed agricultural fields bordering marsh habitat and understandably triggering ditch detection, which can be filtered out by cropping this dataset to the original marsh polygons. Regardless, the overestimation was consistent for all marshes. Previous attempts to map ditches at any scale have been limited to a small number of marshes due to time-consuming manual processing of digital orthophotos and incomplete coverage of LiDAR data (Crain et al. 2009 , McGarigal et al. 2017 ). Relative to these high quality local-scale layers, our method can quickly, accurately (69% true positive rate, 5% false positive rate), and repeatedly assess ditching in marshes. Importantly, our method also improves upon the existing ditching layers by quantifying the area (length x width) of ditches, rather than just the length of ditches. Other attempts to map ditches (or other habitat features) through a broad geographical range that relied on rule-based computer vision approaches, perform poorly with cosmetically heterogeneous images and have failed to detect ditches measuring less than 75 m in length (Lathrop et al. 2006 , McGarigal et al. 2017 ). In contrast, our algorithm can be applied to any marsh throughout the range, relies on freely available aerial imagery, and provides an accurate estimation of ditching. Furthermore, relative to an existing regional layer of ditches (McGarigal et al. 2017 ), the predictive performance of our algorithm represents a 230% improvement. While machine learning and remote sensing have been successfully paired to map many different habitats (Wicaksono et al. 2019 , Lee et al. 2021 ), our study offers insights into challenges that arise when working in highly variable environments and distinguishing similar features (e.g. plant or wildlife species). By harnessing the data augmentation capability of U-Nets, we were able to simulate the tremendous cosmetic variability that would occur in salt marshes throughout space and time by using rescaling, zoom, and rotation to change the coloration, sizing and orientation of the images. During an early evaluation of this methodological approach, after training on only 80 images that showed similar visual appearances (bright green vegetation and clear blue creeks), algorithm performance was highly variable on test images with tanner or darker vegetation or lower contrast images. Because color variation notably affected performance, we prioritized marsh appearance variety in the training data set, which reduced failure to detect any ditches to a rarity. Future studies could avoid this issue by either strategically curating training data sources that show the full range of physical appearance variation and randomly selecting images within those sources, or by augmenting the data with a stronger color variation and adding brightness variation. During that early performance check, the algorithm also occasionally failed to differentiate ditches from similar water features such as creeks and open estuary (particularly with visible wave lines). We were able to improve algorithm accuracy by labeling additional water features in the training data and adding a negative weight (penalty) to these features. It is important to note that we found a tradeoff between maximizing ditch detection and minimizing creek noise. Heavy penalties imposed on creeks drastically reduced the amount of ditching detected, so only a minimal penalty could be imposed on creeks to not lose an unacceptable amount of ditch detection ability. Therefore, labeling a greater number of images and capturing even more marsh variation could potentially further improve algorithm performance. In summary, this study successfully uncovered the regional and local spatial variability of ditching prevalence and intensity by combining a U-Net CNN with heterogenous yet free aerial imagery and established remote sensing techniques. The resulting algorithm can quantify prevalence and intensity at multiple spatial scales, such that responses to intensity can be assessed within a single marsh, or among marshes separated by a few or thousands of kilometers. Furthermore, this approach allows for continuous monitoring of salt marshes which can capture temporal changes within these systems (as ditches fill in or widen) as imagery continues to be updated and the algorithm is reapplied. Thus, researchers now have a tool to test how ditches affect salt marsh ecosystems, investigate whether it has consequences for marsh resilience, and identify whether thresholds exist at which a certain amount of ditching has ecosystem-level consequences. This study also confirms the suitability of pairing U-Nets with remote sensing for producing accurate measurements of landscape features in highly variable environments, which provides a new avenue of biogeographical research for salt marshes and other habitats throughout their global range. Declarations Funding: This research was financially supported by National Science Foundation Award 1736943. Additional funding for was provided by Northeastern University Department of Marine and Environmental Sciences. Author Contributions: Author contributions: K.E.A., T.W.B. and D.L.K conceived the ideas; K.E.A. and T.W.B. collected the data; K.E.A. analyzed the data and wrote the initial draft, and all co-authors contributed to writing and editing thereafter. Conflict of Interest Statement: The authors have no conflicts of interest to declare. Availability of Data and Material: Data will be made available at time of publication and is available to reviewers upon request Acknowledgements: Special thanks to JPA Kaufman for his technical expertise and valuable advice in developing the machine learning methods. We also greatly appreciate insights and guidance provided during the early development of this project by JL Bowen, AR Hughes, and S Scyphers. References Abadi, M., P. Barham, J. Chen, Z. Chen, A. Davis, J. Dean, M. Devin, et al. 2016. “Tensorflow: a system for large-scale machine learning.” Osdi 16 : 265-283. Adam, P. 2019. Salt marsh restoration. In G.M.E. Perillo, E. Wolanski, D.R. Cahoon & C.S. Hopkinson, editors. Coastal Wetlands , pp. 817–61. Elsevier. Adamowicz, S. C., and C. T. Roman. 2005. New England salt marsh pools: A quantitative analysis of geomorphic and geographic features.” Wetlands 25 : 279-88. Adamowicz, S. C., G. Wilson, D. M. Burdick, W. Ferguson, and R. Hopping. 2020. Farmers in the marsh: lessons from history and case studies for the future. Allen, T. F., and T. W. Hoekstra. 2015. Toward a unified ecology . Columbia University Press, New York, NY. Argow, B. A., Z. J. Hughes, and D. M. FitzGerald. 2011. Ice raft formation, sediment load, and theoretical potential for ice-rafted sediment influx on northern coastal wetlands. Continental Shelf Research , 31 (12), 1294-1305. Barbier, E. B., S. D. Hacker, C. Kennedy, E. W. Koch, A. C. Stier, and B. R. Silliman. 2011. The value of estuarine and coastal ecosystem services.” Ecological Monographs , 81 (2): 169-93. Bell, T. W., K. C. Cavanaugh, V. R. Saccomanno, K. C. Cavanaugh, H. F. Houskeeper, N. Eddy, F. Schuetzenmeister, N. Rindlaub, and M. Gleason. 2023. “Kelpwatch: A new visualization and analysis tool to explore kelp canopy dynamics reveals variable response to and recovery from marine heatwaves.” Plos one , 18 (3): e0271477. Bertness, M. D. 1991. Zonation of Spartina patens and Spartina alterniflora in New England salt marsh. Ecology , 72 (1): 138-148. Bouma, T. J., J. van Belzen, T. Balke, J. van Dalen, P. Klaassen, A. M. Hartog, D. P. Callaghan, et al. 2016. Short-term mudflat dynamics drive long-term cyclic salt marsh dynamics. Limnology and Oceanography , 61 (6): 2261–75. Bourn, W. S., and C. Cottam. 1950. Some biological effects of ditching tidewater marshes. No. 19. U.S. Fish and Wildlife Service. Brown, M. T., and M. B. Vivas. 2005. Landscape development intensity index. Environmental monitoring and assessment , 101 : 289-309. Burdick, D. M., G. E. Moore, S. C. Adamowicz, G. M. Wilson, and C. R. Peter. 2020. Mitigating the legacy effects of ditching in a New England salt marsh. Estuaries and Coasts , 43 (7): 1672-1679. Campbell, A. D., L. Fatoyinbo, L. Goldberg, and D. Lagomasino. 2022. Global hotspots of salt marsh change and carbon emissions. Nature , 612 : 701-706. Chapman, V. J. 1960. Salt marshes and salt deserts of the world. In N. Polunin, editor. Plant Science Monographs (Salt Marshes), pp. 1-392. Interscience Publishers, Inc., London and New York, NY. Chollet, F., and others. 2015. Keras. GitHub. Retrieved from https://github.com/keras-team/keras Clarke, J. A., B. A. Harrington, T. Hruby, and F. E. Wasserman. 1984. The effect of ditching for mosquito control on salt marsh use by birds in Rowley, Massachusetts. Journal of Field Ornithology , 55 : 160-180. Corman, S. S., C. T. Roman, J. W. King, and P. G. Appleby. 2012. Salt marsh mosquito-control ditches: Sedimentation, landscape change, and restoration implications. Journal of Coastal Research , 28 (4): 874-880. Correll, M.D., Wiest, W.A., Hodgman, T.P., Shriver, W.G., Elphick, C.S., McGill, B.J., O'Brien, K.M. and Olsen, B.J., 2017. Predictors of specialist avifaunal decline in coastal marshes. Conservation Biology , 31 (1), pp.172-182. Costanza, R., R. de Groot, P. Sutton, S. van der Ploeg, S. J. Anderson, I. Kubiszewski, S. Farber, and R. K. Turner. 2014. Changes in the global value of ecosystem services. Global Environmental Change: Human and Policy Dimensions , 26 : 152-158. Crain, C. M., K. B. Gedan, and M. Dionne. 2009. Tidal restrictions and mosquito ditching in New England marshes. In B.R. Silliman, E.D. Grosholz and M.D. Bertness, editors. Human Impacts on Salt Marshes: A Global Perspective , pp. 149-169. University of California Press, Berkeley, CA. Dahl, T. E. 2011. Status and trends of wetlands in the conterminous United States 2004 to 2009. US Department of the Interior, US Fish and Wildlife Service, Fisheries and Habitat Conservation. Daiber, F. C. 1974. Salt marsh plants and future coastal salt marshes in relation to animals. In R.J. Reimold and W.H. Queen, editors. Ecology of Halophytes , pp. 475-508. Academic Press, Inc., New York, NY. Davidson, N. C. 2014. How much wetland has the world lost? Long-term and recent trends in global wetland area. Marine and Freshwater Research , 65 (10): 934-941. Davies, K. W., T. J. Svejcar, and J. D. Bates. 2009. Interaction of historical and nonhistorical disturbances maintains native plant communities. Ecological Applications , 19 (6): 1536-1545. Deegan, L. A., Johnson, D. S., Warren, R. S., Peterson, B. J., Fleeger, J. W., Fagherazzi, S., & Wollheim, W. M. (2012). Coastal eutrophication as a driver of salt marsh loss. Nature, 490(7420), 388-392. Du, G., X. Cao, J. Liang, X. Chen, and Y. Zhan. 2020. Medical image segmentation based on U-Net: A review. Journal of Imaging Science and Technology , 64 (2): 020508-1-020508-12 Ewanchuk, P. J., and M. D. Bertness. 2003. Recovery of a northern New England salt marsh plant community from winter icing. Oecologia , 136 (4): 616-626. Ewanchuk, P. J., and M. D. Bertness. 2004. Structure and organization of a northern New England salt marsh plant community. Journal of Ecology , 92 (1): 72-85. Foster, D. R., D. H. Knight, and J. F. Franklin. 1998. Landscape patterns and legacies resulting from large, infrequent forest disturbances. Ecosystems , 1 : 497-510. Gedan, K. B., A. H. Altieri, and M. D. Bertness. 2011. Uncertain future of New England salt marshes. Marine Ecology Progress Series , 434 : 229-237. Gedan, K. B., B. R. Silliman, and M. D. Bertness. 2009. Centuries of human-driven change in salt marsh ecosystems. Annual Review of Marine Science , 1 (1): 117-141. Glasgow, R. D. 1938. Mosquitoes and wild life as interrelated problems in human ecology. New York State Museum Bulletin , 316 : 7-20. Holling, C. S. 1973. Resilience and stability of ecological systems. Annual Review of Ecology and Systematics , 4 (1): 1-23. Jackson, H. B., and L. Fahrig. 2015. Are ecologists conducting research at the optimal scale? Global Ecology and Biogeography , 24 (1): 52-63. James‐Pirri, M. J., R. M. Erwin, D. J. Prosser, and J.D. Taylor. 2012. Responses of salt marsh ecosystems to mosquito control management practices along the Atlantic Coast (USA). Restoration Ecology, 20(3), 395-404. Johnson, D. S. 2015. The savory swimmer swims north: A northern range extension of the blue crab Callinectes sapidus ? Journal of Crustacean Biology , 35 (1): 105-110. Kennish, M. J. 2001. Coastal salt marsh systems in the U.S.: A review of anthropogenic impacts. Journal of Coastal Research , 17 (3): 731-748. Kingma, D. P., and J. Ba. 2014. Adam: A method for stochastic optimization. arXiv :1412.6980. Lathrop, R. G., M. B. Cole, and R. D. Showalter. 2000. Quantifying the habitat structure and spatial pattern of New Jersey (U.S.A.) Salt marshes under different management regimes. Wetlands Ecology and Management , 8 : 163-172. Lathrop, R. G., P. Montesano, and S. Haag. 2006. A multi-scale segmentation approach to mapping seagrass habitats using airborne digital camera imagery. Photogrammetric Engineering & Remote Sensing , 72 (6): 665-675. Lee, C. K. F., E. Nicholson, C. Duncan, and N. J. Murray. 2021. Estimating changes and trends in ecosystem extent with dense time‐series satellite remote sensing. Conservation Biology , 35 (1): 325-335. Levin, S. A. 1992. The problem of pattern and scale in ecology: the Robert H. MacArthur award lecture. Ecology , 73 (6): 1943-1967. Levine, J. M., J. S. Brewer, M. D. Bertness. 1998. Nutrients, competition and plant zonation in a New England salt marsh. Journal of Ecology , 285-292. Li, H., C. Wang, Y. Cui, and M. Hodgson. 2021. Mapping salt marsh along coastal South Carolina using U-Net. ISPRS Journal of Photogrammetry and Remote Sensing , 179 : 121-132. Liu, J., T. Dietz, S. R. Carpenter, M. Alberti, C. Folke, E. Moran, A. N. Pell, et al. 2007. Complexity of coupled human and natural systems. Science , 317 (5844): 1513-1516. Marczak, L. B., C. K. Ho, K. Więski, H. Vu, R. F. Denno, and S. C. Pennings. 2011. Latitudinal variation in top‐down and bottom‐up control of a salt marsh food web. Ecology , 92 (2): 276-281. Martínez-Soto, K. S., and D. S. Johnson. 2020. The density of the Atlantic marsh fiddler crab in its expanded range in the Gulf of Maine, USA. The Journal of Crustacean Biology , 40 (5): 544-548. Maxwell, A. E., T. A. Warner, and F. Fang. 2018. Implementation of machine-learning classification in remote sensing: An applied review. International Journal of Remote Sensing , 39 (9): 2784-2817. McCall, B. D., and S. C. Pennings. 2012. Geographic variation in salt marsh structure and function. Oecologia , 170 (3): 777-787. McClintock, B. T., J. D. Nichols, L. L. Bailey, D. I. MacKenzie, W. L. Kendall, and A. B. Franklin. 2010. Seeking a second opinion: uncertainty in disease ecology. Ecology Letters , 13 (6): 659-674. McGarigal, K., B. W. Compton, E. B. Plunkett, and J. Grand. 2017. Designing sustainable landscapes: salt marsh ditching metric. Report to the North Atlantic Conservation Cooperative, US Fish and Wildlife Service, Northeast Region . University of Massachusetts. Mcleod, E., G. L. Chmura, S. Bouillon, R. Salm, M. Björk, C. M. Duarte, C. E. Lovelock, W. H. Schlesinger, and B. R. Silliman. 2011. A blueprint for blue carbon: toward an improved understanding of the role of vegetated coastal habitats in sequestering CO2. Frontiers in Ecology and the Environment , 9 (10): 552-560. Minnett, P. J., A. Alvera-Azcárate, T. M. Chin, G. K. Corlett, C. L. Gentemann, I. Karagali, X. Li, et al. 2019. Half a century of satellite remote sensing of sea-surface temperature. Remote Sensing of Environment , 233 : 111366. Minnich, R. A., M. G. Barbour, J. H. Burk, and J. Sosa‐Ramirez. 2000. Californian mixed‐conifer forests under unmanaged fire regimes in the Sierra San Pedro Mártir, Baja California, Mexico. Journal of Biogeography , 27 (1): 105-129. Nicholls, R. J., P. P. Wong, V. R. Burkett, J. Codignotto, J. Hay, R. McLean, S. Ragoonaden, and C. D. Woodroffe. 2007. Coastal systems and low-lying areas. In M.L. Parry, O.F. Canziani, J.P. Palutikof, P.G. van der Linden, and C.E. Hanson, editors. Climate change 2007: impacts, adaptation and vulnerability. Contribution of Working Group II to the fourth assessment report of the Intergovernmental Panel on Climate Change , pp. 315-356. Cambridge University Press, Cambridge, UK. Niering, W. A., and R. S. Warren. 1980. Vegetation patterns and processes in New England salt marshes. Bioscience , 30 (5): 301-307. Nixon, S. W. 1982. The ecology of New England high salt marshes: a community profile. U.S. Fish and Wildlife Service, US Department of the Interior. Nixon, S. W., and C. A. Oviatt. 1973. Ecology of a New England salt marsh. Ecological Monographs , 43 (4): 463-498. Paine, R. T., M. J. Tegner, and E. A. Johnson. 1998. Compounded perturbations yield ecological surprises. Ecosystems , 1 : 535-545. Pennings, S. C., and M. D. Bertness. 2001. Salt marsh communities. In M.D. Bertness, S.D. Gaines, and M.E. Hay, editors. Marine community ecology , pp. 289-316. Sinauer Associates, Sunderland, MA. Pennings, S. C., M. B. Grant, and M. D. Bertness. 2005. Plant zonation in low-latitude salt marshes: disentangling the roles of flooding, salinity and competition. Journal of Ecology , 93 : 159-167. Pennings, S. C., C. K. Ho, C. S. Salgado, K. Więski, N. Davé, A. E. Kunza, and E. L. Wason. 2009. Latitudinal variation in herbivore pressure in Atlantic Coast salt marshes. Ecology , 90 (1): 183-195. Pethick, J. S. 1992. Saltmarsh geomorphology. In J.R.L. Allen and K. Pye, editors. Saltmarshes: Morphodynamics, conservation and engineering significance , pp. 41-62. Cambridge University Press, Cambridge, UK. Redfield, A. C. 1972. Development of a New England salt marsh. Ecological Monographs , 42 (2): 201-237. Romijn, E., M. Herold, L. Kooistra, D. Murdiyarso, and L. Verchot. 2012. Assessing capacities of non-Annex I countries for national forest monitoring in the context of REDD+. Environmental Science & Policy , 19 : 33-48. Ronneberger, O., P. Fischer, and T. Brox. 2015. U-net: Convolutional networks for biomedical image segmentation. a rXiv :1505.04597 Sallenger Jr, A. H., K. S. Doran, and P. A. Howd. 2012. Hotspot of accelerated sea-level rise on the Atlantic coast of North America. Nature Climate Change , 2 (12): 884-888. Scheffer, M., S. Carpenter, J. A. Foley, C. Folke, and B. Walker. 2001. Catastrophic shifts in ecosystems. Nature , 413 (6856): 591-596. Schuetzenmeister, A., and F. Dufey. 2020. VCA: variance component analysis. R package version 1.4.5. Retrieved from https://cran.r-project.org/web/packages/VCA/ Scott, M. E. 1988. The impact of infection and disease on animal populations: implications for conservation biology. Conservation biology, 2(1), 40-56. Silliman, B. R., and J. C. Zieman. 2001. Top‐down control of Spartina alterniflora production by periwinkle grazing in a Virginia salt marsh. Ecology , 82 (10): 2830-2845. Smith, J.A., Adamowicz, S.C., Wilson, G.M. and Rochlin, I., 2021. “Waffle” pools in ditched salt marshes: assessment, potential causes, and management. Wetlands Ecology and Management, pp.1-17. Smith III, T. J., and W. E. Odum. 1981. The effects of grazing by snow geese on coastal salt marshes. Ecology , 62 (1): 98-106. Smith, J.A. and Pellew, M., 2021. Pond dynamics yield minimal net loss of vegetation cover across an unditched salt marsh landscape. Estuaries and Coasts , 44 (6), pp.1534-1546. Taylor, N. 1938. A preliminary report on the salt marsh vegetation of Long Island, New York. New York State Museum Bulletin , 316 : 21-84. Temmerman, S., P. Meire, T. J. Bouma, P. M. J. Herman, T. Ysebaert, and H. J. De Vriend. 2013. Ecosystem-based coastal defence in the face of global change. Nature , 504 (7478): 79-83. Titus, J. G. 1988. Sea level rise and wetland loss: an overview. In J.G. Titus, editor. Greenhouse Effect, Sea Level Rise, and Coastal Wetlands , pp. 1-36. US Environmental Protection Agency, Washington, D.C. Tkachenko, M., M. Malyuk, A. Holmanyuk, and N. Liubimov. 2020. Label Studio: Data labeling software. Retrieved from https://github.com/heartexlabs/label-studio Tonjes, D. J. 2013. Impacts from ditching salt marshes in the mid-Atlantic and northeastern United States. Environmental Reviews , 21 (2): 116-126. Turner, W., C. Rondinini, N. Pettorelli, B. Mora, A. K. Leidner, Z. Szantoi, G. Buchanan, et al. 2015. Free and open-access satellite data are key to biodiversity conservation. Biological Conservation , 182 : 173-176. U. S. Fish and Wildlife Service. 2018. National Wetlands Inventory website. Available at https://www.fws.gov/wetlands/data/Data-Download.html, accessed 08/28/2018 Watson, E. B., K. B. Raposa, J. C. Carey, C. Wigand, and R. S. Warren. 2017. Anthropocene survival of southern New England’s salt marshes. Estuaries and Coasts , 40 (3): 617-625. Wicaksono, P., P. A. Aryaguna, and W. Lazuardi. 2019. Benthic habitat mapping model and cross validation using machine-learning classification algorithms. Remote Sensing , 11 (11): 1279. Woodcock, C. E., R. Allen, M. Anderson, A. Belward, R. Bindschadler, W. Cohen, F. Gao, et al. 2008. Free access to Landsat imagery. Science , 320 (5879): 1011. Wüest, R. O., N. E. Zimmermann, D. Zurell, J. M. Alexander, S. A. Fritz, C. Hof, H. Kreft, et al. 2020. Macroecology in the age of Big Data–Where to go from here? Journal of Biogeography , 47 (1): 1-12. Cite Share Download PDF Status: Under Review Version 1 posted Reviewers agreed at journal 28 Apr, 2025 Reviewers invited by journal 25 Apr, 2025 Editor invited by journal 19 Apr, 2025 Editor assigned by journal 07 Apr, 2025 First submitted to journal 06 Apr, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6387803","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":448068678,"identity":"400eb0a4-d479-4e48-8b71-50bc66f7e63c","order_by":0,"name":"Karen Aerni","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA9ElEQVRIiWNgGAWjYLCCBIYEORDNjCTGjEMtQosxkGJsBnF4iNIC1JTYQLQW/vbeZw8eMKSlb2fvff64oMLO3p4dKMJQYQ0yBCuQOHPc3CCBISd3Z89xw+YZZ5ITe3iAIgxn0nFqYbiRxiaRwFCRu+FGGmMzbxtzAo8EUISx7TBOLfL3n4G1pBuAtfyrt+eRB4ow/sOtxeAGG0hLTgJES8Nhxh4JoAhjA24thmdADjNIM9xw5hjjbJ5jxxN7wCLH0o1xaZE7foxN8kdFsrzB8TaGzzw11fbs7cfYJD7UWMvi9D7EeegCCXiVj4JRMApGwSggBADhj1ELUNXcvQAAAABJRU5ErkJggg==","orcid":"https://orcid.org/0000-0003-2191-1666","institution":"Northeastern University Marine Science Center","correspondingAuthor":true,"prefix":"","firstName":"Karen","middleName":"","lastName":"Aerni","suffix":""},{"id":448068679,"identity":"88864a4c-f79b-4a14-9188-0c3e6093eec6","order_by":1,"name":"Tom W. Bell","email":"","orcid":"","institution":"Woods Hole Oceanographic Institution Department of Applied Ocean Physics and Engineering","correspondingAuthor":false,"prefix":"","firstName":"Tom","middleName":"W.","lastName":"Bell","suffix":""},{"id":448068680,"identity":"4aac8d3c-fcde-490e-a71a-6d64b284660c","order_by":2,"name":"David L. Kimbro","email":"","orcid":"","institution":"Northeastern University Marine Science Center","correspondingAuthor":false,"prefix":"","firstName":"David","middleName":"L.","lastName":"Kimbro","suffix":""}],"badges":[],"createdAt":"2025-04-06 16:21:46","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6387803/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6387803/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":81645631,"identity":"570127eb-e82e-4f92-acef-770e89a4c78c","added_by":"auto","created_at":"2025-04-29 14:33:17","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":170853,"visible":true,"origin":"","legend":"\u003cp\u003eMap using albers projection and showing the (a) 13 states assessed in this study color coded by coastal bight with the Bourn and Cottam (1950) prevalence estimate overlaid as diagonal lines. In (b) satellite images show the variation in marsh color, ditch orientation and pattern, and overall image contrast throughout the coastal bights. A rectangle representing the color of the bight used in (a) has been drawn around images from the corresponding bight.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6387803/v1/02ad4ebca619cd458a0ef959.png"},{"id":81645653,"identity":"2614bcb7-0e1e-4a34-9de4-3c40d2b7b60a","added_by":"auto","created_at":"2025-04-29 14:33:18","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":670729,"visible":true,"origin":"","legend":"\u003cp\u003eA visual sample of algorithm performance on marshes from each coastal bight (a) where the left images of each row are the marsh images that were fed into the algorithm and the right images are the binary masks generated using the algorithm’s predictions. A line graph showing algorithm validity and performance (b) where the blue line represents training loss (how the algorithm performed on training images) and the orange line represents validation loss (how the algorithm performed on validation images it had not seen before) which shows the model successfully learned (loss decreased) to capacity (loss stabilizes) and was not overfit (training loss does not decline again to 0 nor does validation loss sharply increase). A density plot (c) showing the algorithm’s ditching intensity estimates for unditched marshes where the dashed line represents the median estimate of 1.2%\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6387803/v1/c27a775c44ed6d2e360944c5.png"},{"id":81646424,"identity":"49c498ad-de0b-4c62-9e35-823506a96407","added_by":"auto","created_at":"2025-04-29 14:41:17","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":274084,"visible":true,"origin":"","legend":"\u003cp\u003eA map using albers projection (a) showing the mean prevalence of ditched marshes along the U.S. Atlantic Coast within each coastal bight, where yellow values indicate high prevalence of ditching and blue values indicate low prevalence of ditching. Boxplots are shown in (b) to illustrate the variability of prevalence for the three coastal bights. The red dotted line represents the previously cited statistic that greater than 90% of marshes are ditched from Virginia to Maine (Mid-Atlantic Bight through Gulf of Maine), while the blue dashed line represents the average percentage of marshes ditched for the entire U.S. Atlantic Coast (37.7).\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6387803/v1/7381fcdddff569d16ee4e751.png"},{"id":81645632,"identity":"3b4b337b-fc2c-48b9-a537-f27a3f806852","added_by":"auto","created_at":"2025-04-29 14:33:17","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":365567,"visible":true,"origin":"","legend":"\u003cp\u003ePrevalence of ditching in marshes at the state level mapped as a percentage of total number of marshes within each state with ditching present (a) where dark blue represents the lowest percentages transitioning through to yellow for the highest percentages. A bargraph showing these same prevalence values reflected as a proportion of the number of marshes assessed within each state is shown in (b) where red represents marshes with ditching present and blue shows marshes with no ditching present.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-6387803/v1/97b794d4551e029483864bfd.png"},{"id":81645655,"identity":"78c30b62-254e-43d4-aee2-2169f2670f5d","added_by":"auto","created_at":"2025-04-29 14:33:19","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":334786,"visible":true,"origin":"","legend":"\u003cp\u003eBoxplots showing the variation in ditching intensity (calculated as the total ditch area/total marsh area and converted to a percentage) both within and among bights (a) and states (b) color coded by bight. For states, letters indicate significant differences among states within their respective bights as determined by Tukey HSD. Different letter sets were used for each bight, but should not be construed as differences among states in separate bights.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-6387803/v1/3e4e16ca7d4666d19289a2d1.png"},{"id":81647142,"identity":"33346bb9-04b5-44b2-91ed-4d85fad8cbdd","added_by":"auto","created_at":"2025-04-29 14:57:18","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2497790,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6387803/v1/0e1b4ce1-b331-46e2-872b-9f4da935b6a4.pdf"}],"financialInterests":"","formattedTitle":"Machine learning reveals hierarchical spatial patterns in salt marsh ditching along U.S. Atlantic Coast","fulltext":[{"header":"Introduction","content":"\u003cp\u003eSalt marshes provide important and economically valuable ecosystem services worldwide by buffering shorelines from coastal erosion, filtering water, sequestering carbon, serving as a refuge for biodiversity, and providing space for recreation (Barbier et al. \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2011\u003c/span\u003e, Mcleod et al. \u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e2011\u003c/span\u003e, Temmerman et al. \u003cspan citationid=\"CR77\" class=\"CitationRef\"\u003e2013\u003c/span\u003e, Costanza et al. \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2014\u003c/span\u003e). Despite their value, thousands of acres of salt marsh are lost annually due to anthropogenic activities like land conversion, eutrophication, and hydrological alterations (Dahl \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2011\u003c/span\u003e, Gedan et al. \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2011\u003c/span\u003e, Campbell et al. \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2022\u003c/span\u003e) resulting in a 50% decrease in salt marsh aerial extent globally over the past century (Kennish \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2001\u003c/span\u003e, Davidson \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2014\u003c/span\u003e). Additionally, because tidal inundation is an important driver of salt marsh ecosystem structure and functioning (Daiber \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e1974\u003c/span\u003e, Gedan et al. \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2009\u003c/span\u003e), marsh loss has been exacerbated in recent years by sea level rise, with future U.S. and global losses predicted to reach 20–60% during this century, respectively (Titus \u003cspan citationid=\"CR78\" class=\"CitationRef\"\u003e1988\u003c/span\u003e, Nicholls et al. \u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e2007\u003c/span\u003e, Watson et al. \u003cspan citationid=\"CR83\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). To prioritize effective marsh conservation and restoration, we need a better understanding of the distribution, extent, and degree of anthropogenic modifications within salt marshes.\u003c/p\u003e \u003cp\u003eOne expansive and potentially important anthropogenic manipulation of salt marshes is ditching. Historically, ditches were created as early as the 16th century to increase drainage and facilitate agriculture, some of which were maintained over the years and remain today; see Adamowicz et al. (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2020\u003c/span\u003e) for a detailed history of agricultural ditching. Additionally, thousands of ditches termed “mosquito” or “parallel grid” ditches were constructed throughout marshes along the U.S. Atlantic Coast, primarily during the early 20th century, both to provide jobs as part of the New Deal and to reduce mosquito populations by facilitating predatory fish access to and draining mosquito breeding areas on the marshes (Glasgow \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e1938\u003c/span\u003e, Crain et al. \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2009\u003c/span\u003e). As a result, ditches occur in much higher concentrations than natural tidal channels such as creeks. Ditches also starkly differ from natural creeks because they are rigidly linear, narrow, and are not bordered by gently sloping mudflats which are important for ecosystem functioning (Nixon and Oviatt \u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e1973\u003c/span\u003e, Pethick \u003cspan citationid=\"CR64\" class=\"CitationRef\"\u003e1992\u003c/span\u003e, Bouma et al. \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). Despite their presence in marshes for over a century, and regardless of the motivation behind their creation, the impacts of ditches on most aspects of marsh ecosystem functioning remain uncertain. For instance, while some research found that ditches drain marshes and thereby reduce ponding and tidal creek density (Lathrop et al. \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e2000\u003c/span\u003e, Adamowicz and Roman \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2005\u003c/span\u003e, Smith and Pellew \u003cspan citationid=\"CR75\" class=\"CitationRef\"\u003e2021\u003c/span\u003e), the geomorphic and hydrologic outcomes of other studies ranged from no or subtle impacts (Redfield \u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e1972\u003c/span\u003e, Corman et al. \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2012\u003c/span\u003e) to increasing certain kinds of ponding (Smith et al. \u003cspan citationid=\"CR73\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Similarly, the outcomes of biological responses studied to date comprise negative, positive, and negligible responses by birds, nekton, and invertebrates to ditches (Nixon \u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e1982\u003c/span\u003e, Clarke et al. \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e1984\u003c/span\u003e, James-Pirri et al. \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2012\u003c/span\u003e, Tonjes \u003cspan citationid=\"CR80\" class=\"CitationRef\"\u003e2013\u003c/span\u003e). Understanding this variability in ditching impacts may facilitate the success of on-going marsh management and restoration, including ditch remediation efforts that are increasingly being implemented (Adam \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2019\u003c/span\u003e, Burdick et al. \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2020\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eVariability in ditching impacts may be the result of two key issues. First, the effects of ditching have not been systematically evaluated throughout the broad latitudinal range of salt marshes, which varies substantially in physical and biological conditions such as tidal range, winter ice coverage, and herbivory pressure (Pennings and Bertness \u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e2001\u003c/span\u003e, Marczak et al. \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e2011\u003c/span\u003e, McCall and Pennings \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e2012\u003c/span\u003e). This regional environmental variability could be interacting with ditching to produce the observed inconsistent results among local-scale assessments scattered throughout the broad latitudinal range. Ditches are thought to occur in greater than 90% of marshes from Virginia to Maine (Bourn and Cottam \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e1950\u003c/span\u003e), though this estimate has not been rigorously confirmed or extended to the remaining one-third of the U.S. Atlantic coastline. The second, and non-mutually exclusive, issue concerns the previously unassessed variability in the intensity of ditching. In the majority of the published studies on which our current understanding is based, logistical constraints around measuring these sizeable features across landscapes necessitated that ditching be quantified as present versus absent or subjectively categorized as low or high intensity. Such categorizations may not only mask important non-linear or threshold ecosystem responses to ditching intensity, but also mask important regional variation in non-linear thresholds. Consequently, several individual studies have worked to objectively quantify ditching intensity and evaluate ecosystem responses (Correll et al. \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2017\u003c/span\u003e, McGarigal et al. \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). However, these patch- and local-scale assessments could be bolstered by a comprehensive picture of spatial variation in the prevalence and intensity of ditching that can unravel any scale-dependent impacts of these historical modifications, as well as whether and how to mitigate them.\u003c/p\u003e \u003cp\u003eRemote sensing is a promising tool for measuring landscape features across broad areas such as the 17 degrees of latitude spanned by U.S. Atlantic Coast salt marshes. This technique has been successfully applied to monitor wildfires across forest ranges, temporal changes in kelp cover along an entire coast, and sea surface temperatures across ocean basins (Minnich et al. \u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e2000\u003c/span\u003e, Bell et al. \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2023\u003c/span\u003e, Minnett et al. \u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). But successfully using remote sensing across a broad geographic range requires full spatial coverage by consistent data of a single type (e.g. multispectral imagery or lidar data). Unfortunately, spatial coverage of remote sensing data is lacking in some areas, can be cost prohibitive in others, and is generally at too coarse a resolution (10 to 30m) to detect small or narrow (\u0026lt; 1 m wide) features such as ditches (Woodcock et al. \u003cspan citationid=\"CR85\" class=\"CitationRef\"\u003e2008\u003c/span\u003e, Romijn et al. \u003cspan citationid=\"CR66\" class=\"CitationRef\"\u003e2012\u003c/span\u003e, Turner et al. \u003cspan citationid=\"CR81\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). One solution to these availability hurdles may be the ubiquitous, no-cost, high-resolution color satellite imagery from Google Earth. Indeed, Google Earth imagery is available for the over 600 salt marshes that vary in area from 1 to 100 km\u003csup\u003e2\u003c/sup\u003e between Maine and Florida. Thus, we hypothesized that this affordable remote sensing approach could be used as a first-step towards addressing the multi-scale knowledge gap regarding ditches.\u003c/p\u003e \u003cp\u003eGiven the vast amounts of information involved, translating remote sensing results into a database of ditching prevalence and intensity requires novel data processing techniques that can achieve segmentation (locating features in space) and overcome the very high color variability in the visual appearance of Google Earth sourced satellite imagery. This variability is particularly high for marshes, which vary latitudinally in marsh structure, seasonally in vegetation, and daily in tidal stage. Additionally, ditches themselves vary in length, width, orientation, and whether they intersect or parallel other ditches. While machine learning (ML) techniques such as clustering, classification and regression trees, and convolutional neural networks (CNNs) are promising (Maxwell et al. \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2018\u003c/span\u003e, Wüest et al. \u003cspan citationid=\"CR86\" class=\"CitationRef\"\u003e2020\u003c/span\u003e), they struggle with segmentation tasks (i.e. ‘where is there a cat in this image’ not just ‘is there a cat in this image’) and their accuracy can be hindered by high superficial variability in imagery. However, a recent advancement in CNNs with substantial implications for remote sensing analysis of variable environments is the development of the U-Net architecture (Ronnenberger et al. 2015). U-Nets have long been successfully applied to biomedical imaging tasks and are now being considered for environmental image segmentation tasks (Du et al. \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2020\u003c/span\u003e, Li et al. \u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). In addition to allowing for segmentation, one of the most valuable aspects of U-Nets is that they reduce the amount of training data required to overcome cosmetic image variability by several magnitudes via data augmentation and skip connections. Consequently, the heterogeneous salt marshes and ditches of the U.S. Atlantic coast offer an ideal system for testing the utility of pairing U-Nets with no-cost satellite imagery from Google Earth to quantify multi-scale geographic variation in the prevalence and intensity of ditching.\u003c/p\u003e \u003cp\u003eIn this study, we harnessed machine learning and remote sensing to resolve uncertainty regarding the full spatial extent and variability of ditching. To accomplish this, we trained a U-Net CNN to detect ditches from satellite imagery. We then applied the trained algorithm to satellite imagery of all salt marshes along the U.S. Atlantic Coast greater than 1 km\u003csup\u003e2\u003c/sup\u003e to produce a multi-scale database of salt marsh ditching. Using this database, we then quantified variation in ditching prevalence and intensity at the regional (Gulf of Maine, Mid-Atlantic, South Atlantic), state, and within-state scales. This database will facilitate future studies into the contemporary impacts of this historical habitat modification and in turn help guide the prioritization and planning of marsh management and restoration efforts.\u003c/p\u003e"},{"header":"Methods and Materials","content":"\u003ch3\u003eStudy Area\u003c/h3\u003e\u003cp\u003eThis study focused on salt marshes from the habitat’s northernmost U.S. range edge in Maine (45° N) to its southernmost range edge in Florida (28.5° N). Throughout this range, we identified 634 marshes greater than 1 km\u003csup\u003e2\u003c/sup\u003e that totaled to ~ 7600 km\u003csup\u003e2\u003c/sup\u003e across 13 states and three biogeographic regions (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ea). At local scales, salt marsh communities are predominantly structured by tidal inundation along an elevation gradient where competition between plants dominates the more tolerable conditions at higher elevations and competitively inferior plants are displaced to the more stressful waterlogged conditions at lower elevations (Bertness \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e1991\u003c/span\u003e, Levine et al. \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e1998\u003c/span\u003e, Pennings et al. \u003cspan citationid=\"CR62\" class=\"CitationRef\"\u003e2005\u003c/span\u003e). However, this organization across tidal elevation varies latitudinally because of regionally specific abiotic and biotic drivers (Pennings and Bertness \u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e2001\u003c/span\u003e, Marczak et al. \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e2011\u003c/span\u003e, McCall and Pennings \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e2012\u003c/span\u003e). From 2000 to 2020, the Gulf of Maine (GOM) had the harshest winter conditions with mean daily temperatures of -6°C in Maine (compared to 16°C for Florida) and substantial ice cover throughout winter (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.weather.gov/\u003c/span\u003e\u003cspan address=\"https://www.weather.gov/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). In addition to smothering plants, ice shears off and translocates sediment and senesced grasses leading to patchier vegetation patterns (Niering and Warren \u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e1980\u003c/span\u003e, Ewanchuk and Bertness \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2003\u003c/span\u003e, \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2004\u003c/span\u003e, Argow et al. \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2011\u003c/span\u003e). In contrast, summer is the harshest season for the South Atlantic Bight (SAB). There, plants grow year-round, but summer (July) mean daily high temperatures of 32°C (compared to 25°C in Maine), lead to elevated evapotranspiration rates, high soil salinity, and the increased formation of salt pans (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.weather.gov/\u003c/span\u003e\u003cspan address=\"https://www.weather.gov/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). Meanwhile, the Mid-Atlantic Bight (MAB) has a mix of these extremes at its opposite ends. Finally, while tidal range decreases (nonlinearly) from 5.6 m in GOM to 1 m in SAB (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://tidesandcurrents.noaa.gov/\u003c/span\u003e\u003cspan address=\"https://tidesandcurrents.noaa.gov/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e), herbivory on marsh plants steadily increases (Smith and Odum \u003cspan citationid=\"CR74\" class=\"CitationRef\"\u003e1981\u003c/span\u003e, Silliman and Zieman \u003cspan citationid=\"CR72\" class=\"CitationRef\"\u003e2001\u003c/span\u003e, Pennings et al. \u003cspan citationid=\"CR63\" class=\"CitationRef\"\u003e2009\u003c/span\u003e). In addition to these factors, marsh size and a diversity of other anthropogenic impacts such as eutrophication (Deegan et al. \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2012\u003c/span\u003e) could individually and/or collectively interact with spatial patterns of ditching prevalence and intensity at multiple scales.\u003c/p\u003e\u003ch2\u003eSalt marsh identification and image acquisition\u003c/h2\u003e\u003cp\u003eTo collate images of marsh habitat, we acquired wetland shapefiles from the National Wetlands Inventory for all 13 U.S. states bordering the Atlantic Ocean. We filtered these wetlands to coastal marshes using the ‘Select Attribute by Layer’ tool in ArcGIS Pro 2.8.0 (ESRI, Redlands, CA) to retain polygons classified as Estuarine Intertidal Emergent habitat (“WETLAND_TY = ‘Estuarine and Marine Wetland’ And ATTRIBUTE LIKE %EM%”), which includes the subclasses (numeric codes) of Persistent (1), Non-persistent (2), or \u003cem\u003ePhragmites australis\u003c/em\u003e (5) regardless of any designated modifier codes. To assemble disconnected polygons into cohesive marshes, we aggregated individual polygons within 150 m of each other. We evaluated the effectiveness of this method for several New England marshes with which we are familiar and found that this distance successfully connected polygons separated by small creeks within individual marshes. Furthermore, this distance threshold did not connect polygons between adjacent marshes separated by creeks and upland habitat. To retain a manageable number of marshes (under 1000), we then filtered the aggregated marshes to those greater than 1 km\u003csup\u003e2\u003c/sup\u003e. Marsh polygon aggregation and filtering for size was performed in a single step using the ‘Aggregate Polygons’ Tool with ‘Aggregation Distance’ set to ‘150 Meters’ and ‘Minimum Area’ set to ‘1 SquareKilometers’. Finally, we drew a selection polygon along the state borders and copied the selected features to a new layer for subsequent use, thereby avoiding duplicated marshes since NWI state shapefiles had polygons extending into neighboring states.\u003c/p\u003e\u003cp\u003eThe preceding processing steps produced a database of 634 marshes which we converted to individual KML files of each polygon’s boundaries. KML files were favored over bounding boxes because adjacent marshes were often separated by a few hundred meters and, therefore, bounding boxes would have overlapped substantially and unnecessarily created repeated assessments of the same marsh images. High spatial resolution (\u0026lt; 0.5m) images covering the full extent of each marsh as viewed from an altitude of 400 m were downloaded from Google Earth, and only those images that contained a portion of the marsh that fell within the KML boundaries were retained for analysis. Since the algorithm was intended to handle heterogeneity, only one image was used to represent each marsh portion with no restrictions around the color properties of the marsh at that timepoint; however, to consistently measure current ditching among marshes, images were restricted to being from 2016 or later, a similar image quality, and were not from peak high tide which could make ditches appear wider.\u003c/p\u003e\u003ch3\u003eAlgorithm development and ditch detection\u003c/h3\u003e\u003ch2\u003eAlgorithm structure and training\u003c/h2\u003e\u003cp\u003eTo analyze marsh images for ditching metrics, we created a U-Net deep convolutional neural network (CNN) with seven convolutional layers: three compression layers, one representation layer, and three expansion layers. U-Nets are a special case of CNNs that allow for image segmentation because the network “funnels” information to a low-dimensional representation space and then “expands” back out to reproduce the image (thereby resembling a U shape) with features located. In this study, each layer of the CNN consisted of a stack of two 3-pixel x 3-pixel 2D convolutions followed by either a max pooling operation in the compression section or an up-sampling operation in the expansion section. Note that the pixel resolution in the initial compression layer matches that of the original image but decreases with subsequent compression layers. All layers used the ReLU activation function, except for the final output which used a sigmoid activation function. The output of the algorithm is a grayscale image with values between 0 and 1 representing the confidence that a given pixel is part of a ditch. Because the input size of the network was 256 x 256 pixels, we implemented a tiling algorithm that split individual satellite images into multiple tiles. All coding and machine learning were performed with Python (version 3.10.6) using the keras and tensorflow libraries (versions 2.9.0 and 2.10.0, respectively) on Google Colab.\u003c/p\u003e\u003cp\u003eTo train the algorithm, we generated a training dataset by handpicking 12 marshes from GA through ME that captured a range of variation in marsh appearance or color – such as might arise from latitudinal or seasonal variation – from which we haphazardly selected and labeled 262 images using Label Studio (version 0.9.1) with polygon segmentation labels. For each image, polygons were drawn around individual ditches and assigned a ditch label. As a result, a ditch was represented by an area that reflected their length and width. We created binary masks from these polygons where pixels within the ditch polygon are assigned the value 1 and all other pixels are set to 0. These labeled images and corresponding masks were then randomly partitioned into training data (90% of the labeled images) and validation data (10% of the labeled images) and subsequently split into 20 tiles of dimensions 256 x 256 pixels to match the U-Net's input size. Before proceeding with training, we checked the training and validation data sets to ensure that both of the randomly generated sets contained marsh images with ditches present and absent. During learning – accomplished through stochastic gradient descent – the U-Net made prediction masks for each image, compared that prediction to the true labeled mask and updated its learned weights each iteration. Loss was calculated using a binary cross-entropy score for each pixel in the predicted mask and summed across the full tile. Binary cross-entropy scores factor in how confidently the U-Net classified a pixel as ditch/not-ditch. Pixels incorrectly predicted with high confidence were penalized strongly and the penalty was fed back to the model using back-propagation to update model parameters. We trained the algorithm for 100 epochs using the ADAM optimizer (Kingma and Ba \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e2014\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eAfter reviewing the validation performance by plotting loss as a function of training time, we observed that although the model performed well on finding ditches, it sometimes incorrectly labeled other small water features or steep vertical creek banks as ditches. To improve algorithm performance, we hand-labeled 50 of the 262 ditch-labeled images with the two other water features typically seen in marshes – creeks and ponds. We then fine-tuned the algorithm by updating the weights to include a penalty for non-ditch water features (creeks and ponds) and training for 5 epochs at a smaller learning rate of e\u003csup\u003e− 5\u003c/sup\u003e. The penalty was a mask that assigned a -10 weight to creeks and ponds, which was multiplied by the original algorithm’s output predictions such that a misclassification of a pixel as a ditch (value of 1) would then carry a -10x penalty. Note that more and fewer epochs were tried and all predictions visually evaluated to ultimately determine that 5 epochs was optimal.\u003c/p\u003e\u003ch3\u003eDitch prevalence and intensity\u003c/h3\u003e\u003cp\u003eThe trained and validated algorithm was used to analyze all 634 marshes for ditching prevalence and intensity. But these whole marsh assessments first required processing all the individual constituent images of a single marsh, which, in turn, required tiling each individual marsh image into 20 individual tiles measuring 256 x 256 pixels for input into the algorithm. The algorithm generated predictions for each individual tile using the aforementioned confidence-informed pixel classification from 0 through 1. To convert this prediction into a binary mask, we applied a threshold to each prediction of 0.35 such that every pixel less than 0.35 became a 0 (no ditch), and values equal to or greater than 0.35 became a 1 (a ditch). This value of 0.35 was chosen initially by a visual inspection and was later confirmed using a linear regression trained on the validation data. This thresholding step minimized overestimation by filtering out pixels that did not strongly indicate a ditch. An entire marsh was scored as having ditches present if any of the 20 tiles of any constituent marsh image contained pixels classified as ditch (pixel value of 1). The prevalence of ditched marshes was then calculated for each state, region, and the entire Atlantic coastline by calculating the ratio of ditched marshes:total marshes at each spatial scale.\u003c/p\u003e\u003cp\u003eTo determine ditching intensity of each marsh, we summed the total number of ditch pixels within each tile across all the constituent images of that marsh and converted that count into m\u003csup\u003e2\u003c/sup\u003e based on the spatial resolution of the images. For each marsh, the resulting total amount of ditch area was then divided by total marsh area (generated during earlier GIS processing) and multiplied by 100 to convert the ratio into a percent of marsh area ditched. If a marsh had no ditches present, its intensity was not recorded or included in the analysis of ditching intensity. The algorithm overestimated ditching intensity by around 1%, so we individually verified each marsh for ditching presence to generate an accurate inventory of prevalence.\u003c/p\u003e\u003ch3\u003eComparative performance\u003c/h3\u003e\u003cp\u003eFor a further evaluation of the algorithm’s accuracy, we compared our results to two existing salt marsh ditching datasets following the methods described in more detail in Online Resource 1. The first existing dataset served as the expert classification and comprises ditches in New Hampshire marshes that local experts hand digitized from aerial photos (Eberhardt and Burdick, 2008). To initiate this first comparison, we georectified twenty Google Earth images of the same New Hampshire marsh into QGIS (Online Resource 1). Because the layer of hand-digitized ditches consisted of line features (as opposed to polygons), we converted our algorithm results into a line feature format by hand drawing lines over the predicted ditches, thereby converting the ditch area predictions of our algorithm into comparable ditch lines and length estimates (Online Resource 1, Fig. S1). These lines were partitioned into two separate layers: one for true positives where our predictions overlapped with hand-digitized ditches and one for false positives where our predictions did not overlap with hand-digitized ditches. For each image, we summed the lengths of hand-digitized ditches (local expert layer), our true positive predictions, and our false positive predictions. Note that calculating true negatives was not possible for this analysis.\u003c/p\u003e\u003cp\u003eNext, we evaluated the relative ditch-detecting performance of our algorithm against an existing regional (Virginia to Maine) computer vision-generated dataset of ditches (McGarigal et al. \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2017\u003c/span\u003e), by comparing each to the hand-digitized layer of New Hampshire marshes created by local experts (Eberhardt and Burdick 2008). Similar to the first comparison, ditch predictions from the algorithm and the existing regional layer (McGarigal et al. \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2017\u003c/span\u003e) were partitioned into two separate layers: one for true positives where each dataset’s predictions overlapped with hand-digitized ditches and one for false positives where predictions did not overlap with hand-digitized ditches. For each image, we summed the lengths of hand-digitized ditches (local expert layer), as well as the length of true and false positive predictions for our algorithm and the existing regional layer.\u003c/p\u003e\u003ch2\u003eStatistical Analysis\u003c/h2\u003e\u003cp\u003eWe checked for algorithm overfitting by plotting its loss on the training dataset compared to the validation dataset. We then visually assessed model performance from predictions overlaid on the original aerial images. After confirming that overfitting was absent and that performance was accurate, we calculated mean (± sd) prevalence and intensity of ditching averaged across all three regions and 13 states. Next, we used an analysis of variance (ANOVA) to test for differences in ditch prevalence among the three regions, using state prevalence as replicates. The analysis of prevalence at a scale smaller than region level was not possible because there is only one prevalence value per state. In contrast, we used a Nested ANOVA to test for differences in ditch intensity among both regions and states, with state being nested within region and individual marshes serving as subsamples of states. Following a significant overall model (p \u0026lt; 0.05), we used Tukey’s Honest Significance Difference (HSD) to conduct post hoc mean comparisons of the prevalence and intensity results. For the Nested ANOVA, we also quantified the variance components using the ‘VCA’ package (version 1.4.5) to estimate the amount of variation explained at each spatial scale. Before conducting the ANOVAs, parametric assumptions of normally distributed residuals and homogeneous variances were tested with Shapiro’s test and Levene’s test, respectively. Ditching intensity values were square-root transformed to meet assumptions of normality and homoscedasticity.\u003c/p\u003e\u003cp\u003eTo compare the performance of the algorithm with an existing local-scale layer of ditches, we present the mean (± sd) percent of correctly identified ditches (averaged across 20 images) according to the total length of ditches quantified through hand-digitization by local experts. Of the 20 images, we also present the minimum (worst) and maximum (best) percent of correctly identified ditches by the algorithm. To assess the comparative performance of an existing regional layer, we generated similar summary statistics for predicted ditches of the same 20 images based on McGarigal et al. (\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2017\u003c/span\u003e).\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003eAlgorithm training\u003c/h2\u003e \u003cp\u003eWe successfully trained a U-Net to identify and locate ditches from satellite imagery (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea). The algorithm accurately distinguished ditches from other potentially similar features such as creeks or roads. Based on the validation data, the algorithm successfully learned (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eb), as the binary cross entropy calculated loss steadily decreased and ultimately stabilized, indicating that learning potential was at maximum capacity. Importantly, the plot of loss did not show signs of overfitting, since training loss did not decrease sharply after stabilization (indicating pixel memorization). Equally important, validation loss did not synchronously increase after stabilization (indicating failure to generalize to novel images). To numerically evaluate the algorithm\u0026rsquo;s accuracy, we averaged ditching intensity estimates for the 395 marshes that were manually verified as zero intensity, and found the median overestimation for marshes with no ditching present was 1.2% (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eDitch prevalence and intensity\u003c/h2\u003e \u003cp\u003eWe documented an overall ditching prevalence of 37.7% (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb), which is substantially lower than the previously reported 90%. However, this overall ditching prevalence masks significant regional variation among the three regions (Fig.\u0026nbsp;3, F\u003csub\u003e2,11\u003c/sub\u003e = 6.15, p\u0026thinsp;\u0026lt;\u0026thinsp;.05), with the mean (\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e\u0026plusmn;\u003c/span\u003e\u0026thinsp;sd) prevalence of both GOM (87.4% \u0026plusmn; 19.1%) and MAB (66.7% \u0026plusmn; 34.3%) being significantly greater than the previously unassessed SAB (20.3% \u0026plusmn; 4.9%, TukeyHSD, p\u0026thinsp;\u0026lt;\u0026thinsp;.05). Based on coefficients of variation (CV), states within GOM (CV\u0026thinsp;=\u0026thinsp;21.9) and SAB (CV\u0026thinsp;=\u0026thinsp;24) were more homogeneous in ditching prevalence (with GOM being homogeneously ditched and SAB being homogeneously unditched) than states in MAB (CV\u0026thinsp;=\u0026thinsp;51.4), though all regions exhibited substantial variation in ditch prevalence across states (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). Only five states had ditching prevalence of approximately 90% or higher, including New Hampshire (100%), Massachusetts (96.9%), Rhode Island (100%), Connecticut (92.3%), and New York (86.7%). Virginia and Florida showed the lowest prevalence with only 15.8% and 16.2%, respectively. The remaining states had between 22\u0026ndash;83% of their marshes ditched (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eBecause the algorithm successfully quantified ditching intensity, we were able to generate more resolution on the multi-scale patterns of ditches. Across the U.S. Atlantic coast, for marshes with ditches present, the mean (\u0026plusmn;\u0026thinsp;sd) intensity (percentage) of marsh area composed of ditches is 3.8% (\u0026plusmn;\u0026thinsp;2.6%), with a maximum observed intensity of 14.1%. Unlike prevalence, we did not detect significant regional variation for intensity (nested ANOVA, F\u003csub\u003e2,11\u003c/sub\u003e = 1.94, p\u0026thinsp;=\u0026thinsp;0.15, Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ea). The absence of significant regional variation is likely due to significantly high variation among states within each region (nested ANOVA, F\u003csub\u003e11,225\u003c/sub\u003e = 6.04, p\u0026thinsp;\u0026lt;\u0026thinsp;.001, Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eb). For example, GOM (regional mean\u0026thinsp;\u0026plusmn;\u0026thinsp;sd intensity: 4.1% \u0026plusmn; 2.7%) includes NH (5.1% \u0026plusmn; 1.9%) and MA (4.8% \u0026plusmn; 2.6%) with significantly higher intensity than ME (2.1% \u0026plusmn; 1.5%; TukeyHSD, p\u0026thinsp;\u0026lt;\u0026thinsp;0.05).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eMeanwhile, similar regional means in MAB (3.95% \u0026plusmn; 2.7%) and SAB (3.4% \u0026plusmn; 1.7%) masked a significant difference between NY (5.2% \u0026plusmn; 3.8%) and VA (1.4% \u0026plusmn; 0.5%; TukeyHSD, p\u0026thinsp;\u0026lt;\u0026thinsp;0.05) in the former and NC (3.3% \u0026plusmn; 2.1%) and FL (0.9% \u0026plusmn; 0.9%; TukeyHSD, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001) in the latter (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e). Consequently, all three regions had coefficients of variation from 58\u0026ndash;73%. Interestingly, the four states with the highest levels of ditching intensity (New York\u0026thinsp;=\u0026thinsp;5.6%, New Hampshire\u0026thinsp;=\u0026thinsp;5.1%, Massachusetts\u0026thinsp;=\u0026thinsp;4.9%, Connecticut\u0026thinsp;=\u0026thinsp;4.6%) occurred in the northern two regions. However, in contrast with those findings, the five states with the lowest mean intensities (Florida\u0026thinsp;=\u0026thinsp;1.25%, Rhode Island\u0026thinsp;=\u0026thinsp;1.37%, Virginia\u0026thinsp;=\u0026thinsp;1.87%, Delaware\u0026thinsp;=\u0026thinsp;2.18%, Maine\u0026thinsp;=\u0026thinsp;2.35%) occurred throughout the U.S. Atlantic Coast. After calculating the variance components of the Nested ANOVA, we found that region only accounted for 1.5% of the total variation in ditching intensity, whereas state accounted for 22.5% of the total variation, while the remaining variation (76.2%) is attributable to within-state variation and error (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eb).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eComparative performance\u003c/h2\u003e \u003cp\u003eOf the ditches in a single New Hampshire marsh that were hand digitized by Eberhardt and Burdick (2008), our algorithm captured 69% (\u003cspan type=\"Underline\" class=\"Underline\" name=\"Emphasis\"\u003e\u0026plusmn;\u003c/span\u003e\u0026thinsp;6.5%) of ditch length (true positives). Relative to the existing regional layer of ditches by McGarigal et al. (\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2017\u003c/span\u003e), the algorithm correctly identified 230% more of the true ditches because the existing regional ditch layer correctly identified 21% (9.1%) of ditches. Meanwhile, the false positive rate of the algorithm was relatively low 5.2% (3.7%).\u003c/p\u003e \u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, we successfully assessed ditching prevalence and intensity at multiple scales in salt marshes along the U.S. Atlantic Coast by applying machine learning to publicly available satellite imagery. In our assessment, the average prevalence of ditched marshes is much lower than previous reports in the literature (Bourn and Cottam \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e1950\u003c/span\u003e). But the size of this discrepancy depends on region, with the highest prevalence in the Gulf of Maine decreasing to lowest prevalence in the South Atlantic Bight. For marshes with ditches present, the amount of marsh altered by ditches did not vary regionally. In contrast, spatial variability in ditching intensity varied substantially within regions and states, where intensely ditched states in some regions are opposed by milder ditched states within the same region. Importantly, at least one state in each region has marshes that vary substantially in intensity, allowing future studies to examine functional responses to ditching intensity across scale-dependent biotic and abiotic drivers of salt marsh communities. In addition to demonstrating that a long-held narrative (\u0026ldquo;nearly all U.S. Atlantic Coast temperate marshes are ditched\u0026rdquo; (Tonjes \u003cspan citationid=\"CR80\" class=\"CitationRef\"\u003e2013\u003c/span\u003e)) depends on scale, this blending of ML with freely available data provides a comprehensive and valuable resource to further the conservation and restoration of salt marshes as well as the study of scale in macro- and community ecology.\u003c/p\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eDitching prevalence and intensity\u003c/h2\u003e \u003cp\u003eAcross the latitudinal range of this temperate habitat along the U.S. Atlantic coast (spanning 17 degrees and 2000 km), there is little support for the previously held assertion that nearly all marshes north of Virginia have been modified by ditches, because nearly 50% of the large salt marshes (area\u0026thinsp;\u0026gt;\u0026thinsp;1km\u003csup\u003e2\u003c/sup\u003e) north of Virginia evaluated in this study actually lack ditches. Without this updated and larger-scale assessment, pursuing research on how the presence/absence of ditches affects diversity and ecosystem functioning in the Gulf of Maine and Mid-Atlantic Bight may not be sufficiently prioritized, because of the incorrectly presumed absence of control (unditched) marshes to which ditched marshes can be compared. Furthermore, the historical baseline provided by Bourn and Cottam (\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e1950\u003c/span\u003e) is geographically restricted. By updating this baseline with information on the South Atlantic Bight, we now know that 22% of the marshes in this region have been ditched and therefore could also be considered for research and mitigation.\u003c/p\u003e \u003cp\u003eOur results suggest that any investment decisions around research and mitigation of ditches would benefit from understanding two aspects of the spatial variability in ditches. First, the regional variation in ditching prevalence suggests that ditching effort (and/or recovery) was heterogeneous, with only GOM approaching high prevalence of 90%. Interestingly, the actual number of ditched marshes is relatively homogeneous throughout the study area (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eb), suggesting that regional variation in prevalence is influenced at least in part by the increase in number of marshes from the GOM through the SAB due to geomorphological differences (Chapman \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e1960\u003c/span\u003e). Second, if we control for regional differences in overall marsh abundance by comparing prevalence among states within each region, we find significant spatial structure (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e). For instance, GOM and MAB both have states with prevalence rates above 90% (NH/MA and CT, respectively) that are balanced by states with lower prevalence such as ME (65%) and VA (6%), respectively. Furthermore, and in contrast to a regional evaluation that suggested a paucity of unditched or control marshes (Crain et al. \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2009\u003c/span\u003e), our results demonstrated the presence of unditched marshes in every state except RI (MAB) and NH (GOM), which have a paucity of marshes greater than one km\u003csup\u003e2\u003c/sup\u003e in general (2 and 6, respectively). Consequently, there are sufficient ditched and unditched marshes within each region to robustly evaluate the impacts of ditching presence on diversity and ecosystem functioning, while controlling for background regional environmental differences.\u003c/p\u003e \u003cp\u003eSimilar to the study of disturbance in ecology and disease in epidemiology, the biogeography of prevalence in an anthropogenic modification of marshes will likely be insufficient for optimizing future research as well as restoration and management decisions (Brown and Vivas \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2005\u003c/span\u003e, Liu et al. \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e2007\u003c/span\u003e, McClintock et al. \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e2010\u003c/span\u003e). This is because, whether it be key relationships between pathogen virulence and transmission or biodiversity and nutrient cycling, natural systems are replete with non-linear functional responses characterized by thresholds and critical tipping points (Holling \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e1973\u003c/span\u003e, Scheffer et al. \u003cspan citationid=\"CR69\" class=\"CitationRef\"\u003e2001\u003c/span\u003e, Liu et al. \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e2007\u003c/span\u003e). By quantifying the full range of ditching intensity, our research laid the groundwork for detecting and understanding any such non-linear relationships. At the largest scale, we found the mean intensity of ditching (excluding unditched marshes) across the U.S. Atlantic coast to be 3.8%, greatly exceeding the only existing intensity estimate of approximately 1% for New England marshes (Crain et al. \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2009\u003c/span\u003e). Nested within this large-scale estimate is the existence of hierarchically organized spatial patterning in ditch intensity. At the regional scale, ditching intensity did not differ significantly, accounting for little of the variance (1.5%). But within each region, we found substantial variation (22.5%) in ditch intensity (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e), with all three regions having states with high, moderate, and low intensities, as well as unditched marshes. There is also considerable variation at the next hierarchical level, where almost all states exhibit large within-state variation except for ME in GOM, RI and VA in MAB, and FL in SAB (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eb). However, to determine exactly how much of the remaining variation in ditching intensity (76.2%) is due to within-state variation compared to error would require subsampling within marshes, which is worthy of future study. Consequently, our approach not only laid groundwork for detecting non-linear relationships, but it did so in a manner that promotes scale-dependent assessments and thus addresses a central challenge in ecology (Levin \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e1992\u003c/span\u003e, Allen and Hoekstra \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2015\u003c/span\u003e, Jackson and Fahrig \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2015\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThis research also paves the way for disentangling how ditches interact with established and projected environmental forcing of salt marsh community structure. By targeting sites encompassing the full gradient of ditching intensity, researchers in GOM could evaluate interactions with ice disturbances in the winter and wrack abundances in the spring. Meanwhile, in SAB, researchers can look at interactions with salt pans and herbivory pressure. Across all regions, researchers can evaluate the influence of tidal range, which has previously been proposed as determining the hydrological impacts of ditching (Taylor \u003cspan citationid=\"CR76\" class=\"CitationRef\"\u003e1938\u003c/span\u003e, Redfield \u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e1972\u003c/span\u003e, Tonjes \u003cspan citationid=\"CR80\" class=\"CitationRef\"\u003e2013\u003c/span\u003e). It should also be noted that not all ditches drain equally, as some have become plugged over time and now hold water, which is another factor that researchers can address from this starting point. Equally important, researchers can investigate how ditching interacts with novel and increasingly influential drivers of salt marsh change. As with other disturbances, the legacy effects of ditching can impact the ecosystem\u0026rsquo;s response to successive disturbances, leading to ecological surprises and catastrophic shifts (Foster et al. \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e1998\u003c/span\u003e, Paine et al. \u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e1998\u003c/span\u003e, Davies et al. \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2009\u003c/span\u003e). Threats of particular concern include accelerated sea level rise in MAB (Sallenger et al. \u003cspan citationid=\"CR68\" class=\"CitationRef\"\u003e2012\u003c/span\u003e), abundant tidal restrictions in GOM (Crain et al. \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2009\u003c/span\u003e), and increasing range expansion of southern latitudinal predators and consumers such as blue crabs (Johnson, \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2015\u003c/span\u003e) and fiddler crabs (Mart\u0026iacute;nez-Soto and Johnson, \u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Not only does this help us allocate research to understand ditching disturbance across abiotic and biotic forces at multiple scales, we can also strategically distribute ditch restoration experiments, thereby optimally informing and promoting marsh resilience.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eDetection of ditches by a U-Net CNN\u003c/h2\u003e \u003cp\u003eThe algorithm developed in this study exceeded performance expectations (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e) and showed the applicability of using machine learning for regional and fine scale mapping of features within highly variable habitats. In general, the algorithm detected all ditches in an image and only infrequently missed a ditch, though it sometimes underestimated the length of a ditch. While it accurately detected and estimated ditches, it also minimally detected \u0026ldquo;ditches\u0026rdquo; where there were none. This was largely filtered out in the thresholding step, but not entirely, leading to all measurements of intensities being slightly overestimated by approximately 1% (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec). It is important to note that this overestimation can be partly attributed to plowed agricultural fields bordering marsh habitat and understandably triggering ditch detection, which can be filtered out by cropping this dataset to the original marsh polygons. Regardless, the overestimation was consistent for all marshes.\u003c/p\u003e \u003cp\u003ePrevious attempts to map ditches at any scale have been limited to a small number of marshes due to time-consuming manual processing of digital orthophotos and incomplete coverage of LiDAR data (Crain et al. \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2009\u003c/span\u003e, McGarigal et al. \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). Relative to these high quality local-scale layers, our method can quickly, accurately (69% true positive rate, 5% false positive rate), and repeatedly assess ditching in marshes. Importantly, our method also improves upon the existing ditching layers by quantifying the area (length x width) of ditches, rather than just the length of ditches. Other attempts to map ditches (or other habitat features) through a broad geographical range that relied on rule-based computer vision approaches, perform poorly with cosmetically heterogeneous images and have failed to detect ditches measuring less than 75 m in length (Lathrop et al. \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2006\u003c/span\u003e, McGarigal et al. \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). In contrast, our algorithm can be applied to any marsh throughout the range, relies on freely available aerial imagery, and provides an accurate estimation of ditching. Furthermore, relative to an existing regional layer of ditches (McGarigal et al. \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2017\u003c/span\u003e), the predictive performance of our algorithm represents a 230% improvement.\u003c/p\u003e \u003cp\u003eWhile machine learning and remote sensing have been successfully paired to map many different habitats (Wicaksono et al. \u003cspan citationid=\"CR84\" class=\"CitationRef\"\u003e2019\u003c/span\u003e, Lee et al. \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2021\u003c/span\u003e), our study offers insights into challenges that arise when working in highly variable environments and distinguishing similar features (e.g. plant or wildlife species). By harnessing the data augmentation capability of U-Nets, we were able to simulate the tremendous cosmetic variability that would occur in salt marshes throughout space and time by using rescaling, zoom, and rotation to change the coloration, sizing and orientation of the images. During an early evaluation of this methodological approach, after training on only 80 images that showed similar visual appearances (bright green vegetation and clear blue creeks), algorithm performance was highly variable on test images with tanner or darker vegetation or lower contrast images. Because color variation notably affected performance, we prioritized marsh appearance variety in the training data set, which reduced failure to detect any ditches to a rarity. Future studies could avoid this issue by either strategically curating training data sources that show the full range of physical appearance variation and randomly selecting images within those sources, or by augmenting the data with a stronger color variation and adding brightness variation. During that early performance check, the algorithm also occasionally failed to differentiate ditches from similar water features such as creeks and open estuary (particularly with visible wave lines). We were able to improve algorithm accuracy by labeling additional water features in the training data and adding a negative weight (penalty) to these features. It is important to note that we found a tradeoff between maximizing ditch detection and minimizing creek noise. Heavy penalties imposed on creeks drastically reduced the amount of ditching detected, so only a minimal penalty could be imposed on creeks to not lose an unacceptable amount of ditch detection ability. Therefore, labeling a greater number of images and capturing even more marsh variation could potentially further improve algorithm performance.\u003c/p\u003e \u003cp\u003eIn summary, this study successfully uncovered the regional and local spatial variability of ditching prevalence and intensity by combining a U-Net CNN with heterogenous yet free aerial imagery and established remote sensing techniques. The resulting algorithm can quantify prevalence and intensity at multiple spatial scales, such that responses to intensity can be assessed within a single marsh, or among marshes separated by a few or thousands of kilometers. Furthermore, this approach allows for continuous monitoring of salt marshes which can capture temporal changes within these systems (as ditches fill in or widen) as imagery continues to be updated and the algorithm is reapplied. Thus, researchers now have a tool to test how ditches affect salt marsh ecosystems, investigate whether it has consequences for marsh resilience, and identify whether thresholds exist at which a certain amount of ditching has ecosystem-level consequences. This study also confirms the suitability of pairing U-Nets with remote sensing for producing accurate measurements of landscape features in highly variable environments, which provides a new avenue of biogeographical research for salt marshes and other habitats throughout their global range.\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003ch2\u003eFunding:\u003c/h2\u003e \u003cp\u003eThis research was financially supported by National Science Foundation Award 1736943. Additional funding for was provided by Northeastern University Department of Marine and Environmental Sciences.\u003c/p\u003e \u003cp\u003eAuthor Contributions:\u003c/p\u003e \u003cp\u003eAuthor contributions: K.E.A., T.W.B. and D.L.K conceived the ideas; K.E.A. and T.W.B. collected the data; K.E.A. analyzed the data and wrote the initial draft, and all co-authors contributed to writing and editing thereafter.\u003c/p\u003e \u003cp\u003eConflict of Interest Statement:\u003c/p\u003e \u003cp\u003eThe authors have no conflicts of interest to declare.\u003c/p\u003e \u003cp\u003eAvailability of Data and Material:\u003c/p\u003e \u003cp\u003eData will be made available at time of publication and is available to reviewers upon request\u003c/p\u003e\u003ch2\u003eAcknowledgements:\u003c/h2\u003e \u003cp\u003eSpecial thanks to JPA Kaufman for his technical expertise and valuable advice in developing the machine learning methods. We also greatly appreciate insights and guidance provided during the early development of this project by JL Bowen, AR Hughes, and S Scyphers.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAbadi, M., P. Barham, J. Chen, Z. Chen, A. Davis, J. Dean, M. Devin, et al. 2016. \u0026ldquo;Tensorflow: a system for large-scale machine learning.\u0026rdquo; \u003cem\u003eOsdi\u003c/em\u003e \u003cstrong\u003e16\u003c/strong\u003e: 265-283.\u003c/li\u003e\n\u003cli\u003eAdam, P. 2019. Salt marsh restoration. In G.M.E. Perillo, E. Wolanski, D.R. Cahoon \u0026amp; C.S. Hopkinson, editors. \u003cem\u003eCoastal Wetlands\u003c/em\u003e, pp. 817\u0026ndash;61. Elsevier.\u003c/li\u003e\n\u003cli\u003eAdamowicz, S. C., and C. T. Roman. 2005. New England salt marsh pools: A quantitative analysis of geomorphic and geographic features.\u0026rdquo; \u003cem\u003eWetlands \u003c/em\u003e\u003cstrong\u003e25\u003c/strong\u003e: 279-88.\u003c/li\u003e\n\u003cli\u003eAdamowicz, S. C., G. Wilson, D. M. Burdick, W. Ferguson, and R. Hopping. 2020. Farmers in the marsh: lessons from history and case studies for the future.\u003c/li\u003e\n\u003cli\u003eAllen, T. F., and T. W. Hoekstra. 2015. \u003cem\u003eToward a unified ecology\u003c/em\u003e. Columbia University Press, New York, NY.\u003c/li\u003e\n\u003cli\u003eArgow, B. A., Z. J. Hughes, and D. M. FitzGerald. 2011. Ice raft formation, sediment load, and theoretical potential for ice-rafted sediment influx on northern coastal wetlands. \u003cem\u003eContinental Shelf Research\u003c/em\u003e, \u003cem\u003e31\u003c/em\u003e(12), 1294-1305.\u003c/li\u003e\n\u003cli\u003eBarbier, E. B., S. D. Hacker, C. Kennedy, E. W. Koch, A. C. Stier, and B. R. Silliman. 2011. The value of estuarine and coastal ecosystem services.\u0026rdquo; \u003cem\u003eEcological Monographs\u003c/em\u003e, \u003cstrong\u003e81\u003c/strong\u003e(2): 169-93.\u003c/li\u003e\n\u003cli\u003eBell, T. W., K. C. Cavanaugh, V. R. Saccomanno, K. C. Cavanaugh, H. F. Houskeeper, N. Eddy, F. Schuetzenmeister, N. Rindlaub, and M. Gleason. 2023. \u0026ldquo;Kelpwatch: A new visualization and analysis tool to explore kelp canopy dynamics reveals variable response to and recovery from marine heatwaves.\u0026rdquo; \u003cem\u003ePlos one\u003c/em\u003e, \u003cstrong\u003e18\u003c/strong\u003e(3): e0271477.\u003c/li\u003e\n\u003cli\u003eBertness, M. D. 1991. Zonation of \u003cem\u003eSpartina patens\u003c/em\u003e and \u003cem\u003eSpartina alterniflora\u003c/em\u003e in New England salt marsh. \u003cem\u003eEcology\u003c/em\u003e, \u003cstrong\u003e72\u003c/strong\u003e(1): 138-148.\u003c/li\u003e\n\u003cli\u003eBouma, T. J., J. van Belzen, T. Balke, J. van Dalen, P. Klaassen, A. M. Hartog, D. P. Callaghan, et al. 2016. Short-term mudflat dynamics drive long-term cyclic salt marsh dynamics. \u003cem\u003eLimnology and Oceanography\u003c/em\u003e, \u003cstrong\u003e61\u003c/strong\u003e(6): 2261\u0026ndash;75.\u003c/li\u003e\n\u003cli\u003eBourn, W. S., and C. Cottam. 1950. Some biological effects of ditching tidewater marshes. No. 19. U.S. Fish and Wildlife Service.\u003c/li\u003e\n\u003cli\u003eBrown, M. T., and M. B. Vivas. 2005. Landscape development intensity index. \u003cem\u003eEnvironmental monitoring and assessment\u003c/em\u003e, \u003cstrong\u003e101\u003c/strong\u003e: 289-309.\u003c/li\u003e\n\u003cli\u003eBurdick, D. M., G. E. Moore, S. C. Adamowicz, G. M. Wilson, and C. R. Peter. 2020. Mitigating the legacy effects of ditching in a New England salt marsh. \u003cem\u003eEstuaries and Coasts\u003c/em\u003e, \u003cstrong\u003e43\u003c/strong\u003e(7): 1672-1679.\u003c/li\u003e\n\u003cli\u003eCampbell, A. D., L. Fatoyinbo, L. Goldberg, and D. Lagomasino. 2022. Global hotspots of salt marsh change and carbon emissions. \u003cem\u003eNature\u003c/em\u003e, \u003cstrong\u003e612\u003c/strong\u003e: 701-706.\u003c/li\u003e\n\u003cli\u003eChapman, V. J. 1960. Salt marshes and salt deserts of the world. In N. Polunin, editor. \u003cem\u003ePlant Science Monographs\u003c/em\u003e (Salt Marshes), pp. 1-392. Interscience Publishers, Inc., London and New York, NY.\u003c/li\u003e\n\u003cli\u003eChollet, F., and others. 2015. Keras. GitHub. Retrieved from https://github.com/keras-team/keras\u003c/li\u003e\n\u003cli\u003eClarke, J. A., B. A. Harrington, T. Hruby, and F. E. Wasserman. 1984. The effect of ditching for mosquito control on salt marsh use by birds in Rowley, Massachusetts. \u003cem\u003eJournal of Field Ornithology\u003c/em\u003e, \u003cstrong\u003e55\u003c/strong\u003e: 160-180.\u003c/li\u003e\n\u003cli\u003eCorman, S. S., C. T. Roman, J. W. King, and P. G. Appleby. 2012. Salt marsh mosquito-control ditches: Sedimentation, landscape change, and restoration implications. \u003cem\u003eJournal of Coastal Research\u003c/em\u003e, \u003cstrong\u003e28\u003c/strong\u003e(4): 874-880.\u003c/li\u003e\n\u003cli\u003eCorrell, M.D., Wiest, W.A., Hodgman, T.P., Shriver, W.G., Elphick, C.S., McGill, B.J., O\u0026apos;Brien, K.M. and Olsen, B.J., 2017. Predictors of specialist avifaunal decline in coastal marshes. \u003cem\u003eConservation Biology\u003c/em\u003e, \u003cem\u003e31\u003c/em\u003e(1), pp.172-182.\u003c/li\u003e\n\u003cli\u003eCostanza, R., R. de Groot, P. Sutton, S. van der Ploeg, S. J. Anderson, I. Kubiszewski, S. Farber, and R. K. Turner. 2014. Changes in the global value of ecosystem services. \u003cem\u003eGlobal Environmental Change: Human and Policy Dimensions\u003c/em\u003e, \u003cstrong\u003e26\u003c/strong\u003e: 152-158.\u003c/li\u003e\n\u003cli\u003eCrain, C. M., K. B. Gedan, and M. Dionne. 2009. Tidal restrictions and mosquito ditching in New England marshes. In B.R. Silliman, E.D. Grosholz and M.D. Bertness, editors. \u003cem\u003eHuman Impacts on Salt Marshes: A Global Perspective\u003c/em\u003e, pp. 149-169. University of California Press, Berkeley, CA.\u003c/li\u003e\n\u003cli\u003eDahl, T. E. 2011. Status and trends of wetlands in the conterminous United States 2004 to 2009. US Department of the Interior, US Fish and Wildlife Service, Fisheries and Habitat Conservation.\u003c/li\u003e\n\u003cli\u003eDaiber, F. C. 1974. Salt marsh plants and future coastal salt marshes in relation to animals. In R.J. Reimold and W.H. Queen, editors. \u003cem\u003eEcology of Halophytes\u003c/em\u003e, pp. 475-508. Academic Press, Inc., New York, NY.\u003c/li\u003e\n\u003cli\u003eDavidson, N. C. 2014. How much wetland has the world lost? Long-term and recent trends in global wetland area. \u003cem\u003eMarine and Freshwater Research\u003c/em\u003e, \u003cstrong\u003e65\u003c/strong\u003e(10): 934-941.\u003c/li\u003e\n\u003cli\u003eDavies, K. W., T. J. Svejcar, and J. D. Bates. 2009. Interaction of historical and nonhistorical disturbances maintains native plant communities. \u003cem\u003eEcological Applications\u003c/em\u003e, \u003cstrong\u003e19\u003c/strong\u003e(6): 1536-1545.\u003c/li\u003e\n\u003cli\u003eDeegan, L. A., Johnson, D. S., Warren, R. S., Peterson, B. J., Fleeger, J. W., Fagherazzi, S., \u0026amp; Wollheim, W. M. (2012). Coastal eutrophication as a driver of salt marsh loss. Nature, 490(7420), 388-392.\u003c/li\u003e\n\u003cli\u003eDu, G., X. Cao, J. Liang, X. Chen, and Y. Zhan. 2020. Medical image segmentation based on U-Net: A review. \u003cem\u003eJournal of Imaging Science and Technology\u003c/em\u003e, \u003cstrong\u003e64\u003c/strong\u003e(2): 020508-1-020508-12\u003c/li\u003e\n\u003cli\u003eEwanchuk, P. J., and M. D. Bertness. 2003. Recovery of a northern New England salt marsh plant community from winter icing. \u003cem\u003eOecologia\u003c/em\u003e, \u003cstrong\u003e136\u003c/strong\u003e(4): 616-626.\u003c/li\u003e\n\u003cli\u003eEwanchuk, P. J., and M. D. Bertness. 2004. Structure and organization of a northern New England salt marsh plant community. \u003cem\u003eJournal of Ecology\u003c/em\u003e, \u003cstrong\u003e92\u003c/strong\u003e(1): 72-85.\u003c/li\u003e\n\u003cli\u003eFoster, D. R., D. H. Knight, and J. F. Franklin. 1998. Landscape patterns and legacies resulting from large, infrequent forest disturbances. \u003cem\u003eEcosystems\u003c/em\u003e, \u003cstrong\u003e1\u003c/strong\u003e: 497-510.\u003c/li\u003e\n\u003cli\u003eGedan, K. B., A. H. Altieri, and M. D. Bertness. 2011. Uncertain future of New England salt marshes. \u003cem\u003eMarine Ecology Progress Series\u003c/em\u003e, \u003cstrong\u003e434\u003c/strong\u003e: 229-237.\u003c/li\u003e\n\u003cli\u003eGedan, K. B., B. R. Silliman, and M. D. Bertness. 2009. Centuries of human-driven change in salt marsh ecosystems. \u003cem\u003eAnnual Review of Marine Science\u003c/em\u003e, \u003cstrong\u003e1\u003c/strong\u003e(1): 117-141.\u003c/li\u003e\n\u003cli\u003eGlasgow, R. D. 1938. Mosquitoes and wild life as interrelated problems in human ecology. \u003cem\u003eNew York State Museum Bulletin\u003c/em\u003e, \u003cstrong\u003e316\u003c/strong\u003e: 7-20.\u003c/li\u003e\n\u003cli\u003eHolling, C. S. 1973. Resilience and stability of ecological systems. \u003cem\u003eAnnual Review of Ecology and Systematics\u003c/em\u003e, \u003cstrong\u003e4\u003c/strong\u003e(1): 1-23.\u003c/li\u003e\n\u003cli\u003eJackson, H. B., and L. Fahrig. 2015. Are ecologists conducting research at the optimal scale? \u003cem\u003eGlobal Ecology and Biogeography\u003c/em\u003e, \u003cstrong\u003e24\u003c/strong\u003e(1): 52-63.\u003c/li\u003e\n\u003cli\u003eJames‐Pirri, M. J., R. M. Erwin, D. J. Prosser, and J.D. Taylor. 2012. Responses of salt marsh ecosystems to mosquito control management practices along the Atlantic Coast (USA). Restoration Ecology, 20(3), 395-404. \u003c/li\u003e\n\u003cli\u003eJohnson, D. S. 2015. The savory swimmer swims north: A northern range extension of the blue crab \u003cem\u003eCallinectes sapidus\u003c/em\u003e? \u003cem\u003eJournal of Crustacean Biology\u003c/em\u003e, \u003cstrong\u003e35\u003c/strong\u003e(1): 105-110.\u003c/li\u003e\n\u003cli\u003eKennish, M. J. 2001. Coastal salt marsh systems in the U.S.: A review of anthropogenic impacts. \u003cem\u003eJournal of Coastal Research\u003c/em\u003e, \u003cstrong\u003e17\u003c/strong\u003e(3): 731-748.\u003c/li\u003e\n\u003cli\u003eKingma, D. P., and J. Ba. 2014. Adam: A method for stochastic optimization. \u003cem\u003earXiv\u003c/em\u003e:1412.6980.\u003c/li\u003e\n\u003cli\u003eLathrop, R. G., M. B. Cole, and R. D. Showalter. 2000. Quantifying the habitat structure and spatial pattern of New Jersey (U.S.A.) Salt marshes under different management regimes. \u003cem\u003eWetlands Ecology and Management\u003c/em\u003e, \u003cstrong\u003e8\u003c/strong\u003e: 163-172.\u003c/li\u003e\n\u003cli\u003eLathrop, R. G., P. Montesano, and S. Haag. 2006. A multi-scale segmentation approach to mapping seagrass habitats using airborne digital camera imagery. \u003cem\u003ePhotogrammetric Engineering \u0026amp; Remote Sensing\u003c/em\u003e, \u003cstrong\u003e72\u003c/strong\u003e(6): 665-675.\u003c/li\u003e\n\u003cli\u003eLee, C. K. F., E. Nicholson, C. Duncan, and N. J. Murray. 2021. Estimating changes and trends in ecosystem extent with dense time‐series satellite remote sensing. \u003cem\u003eConservation Biology\u003c/em\u003e, \u003cstrong\u003e35\u003c/strong\u003e(1): 325-335.\u003c/li\u003e\n\u003cli\u003eLevin, S. A. 1992. The problem of pattern and scale in ecology: the Robert H. MacArthur award lecture. \u003cem\u003eEcology\u003c/em\u003e, \u003cstrong\u003e73\u003c/strong\u003e(6): 1943-1967.\u003c/li\u003e\n\u003cli\u003eLevine, J. M., J. S. Brewer, M. D. Bertness. 1998. Nutrients, competition and plant zonation in a New England salt marsh. \u003cem\u003eJournal of Ecology\u003c/em\u003e, 285-292.\u003c/li\u003e\n\u003cli\u003eLi, H., C. Wang, Y. Cui, and M. Hodgson. 2021. Mapping salt marsh along coastal South Carolina using U-Net. \u003cem\u003eISPRS Journal of Photogrammetry and Remote Sensing\u003c/em\u003e, \u003cstrong\u003e179\u003c/strong\u003e: 121-132.\u003c/li\u003e\n\u003cli\u003eLiu, J., T. Dietz, S. R. Carpenter, M. Alberti, C. Folke, E. Moran, A. N. Pell, et al. 2007. Complexity of coupled human and natural systems. \u003cem\u003eScience\u003c/em\u003e, \u003cstrong\u003e317\u003c/strong\u003e(5844): 1513-1516.\u003c/li\u003e\n\u003cli\u003eMarczak, L. B., C. K. Ho, K. Więski, H. Vu, R. F. Denno, and S. C. Pennings. 2011. Latitudinal variation in top‐down and bottom‐up control of a salt marsh food web. \u003cem\u003eEcology\u003c/em\u003e, \u003cstrong\u003e92\u003c/strong\u003e(2): 276-281.\u003c/li\u003e\n\u003cli\u003eMart\u0026iacute;nez-Soto, K. S., and D. S. Johnson. 2020. The density of the Atlantic marsh fiddler crab in its expanded range in the Gulf of Maine, USA. \u003cem\u003eThe Journal of Crustacean Biology\u003c/em\u003e, \u003cstrong\u003e40\u003c/strong\u003e(5): 544-548.\u003c/li\u003e\n\u003cli\u003eMaxwell, A. E., T. A. Warner, and F. Fang. 2018. Implementation of machine-learning classification in remote sensing: An applied review. \u003cem\u003eInternational Journal of Remote Sensing\u003c/em\u003e, \u003cstrong\u003e39\u003c/strong\u003e(9): 2784-2817.\u003c/li\u003e\n\u003cli\u003eMcCall, B. D., and S. C. Pennings. 2012. Geographic variation in salt marsh structure and function. \u003cem\u003eOecologia\u003c/em\u003e, \u003cstrong\u003e170\u003c/strong\u003e(3): 777-787.\u003c/li\u003e\n\u003cli\u003eMcClintock, B. T., J. D. Nichols, L. L. Bailey, D. I. MacKenzie, W. L. Kendall, and A. B. Franklin. 2010. Seeking a second opinion: uncertainty in disease ecology. \u003cem\u003eEcology Letters\u003c/em\u003e, \u003cstrong\u003e13\u003c/strong\u003e(6): 659-674.\u003c/li\u003e\n\u003cli\u003eMcGarigal, K., B. W. Compton, E. B. Plunkett, and J. Grand. 2017. Designing sustainable landscapes: salt marsh ditching metric. \u003cem\u003eReport to the North Atlantic Conservation Cooperative, US Fish and Wildlife Service, Northeast Region\u003c/em\u003e. University of Massachusetts.\u003c/li\u003e\n\u003cli\u003eMcleod, E., G. L. Chmura, S. Bouillon, R. Salm, M. Bj\u0026ouml;rk, C. M. Duarte, C. E. Lovelock, W. H. Schlesinger, and B. R. Silliman. 2011. A blueprint for blue carbon: toward an improved understanding of the role of vegetated coastal habitats in sequestering CO2. \u003cem\u003eFrontiers in Ecology and the Environment\u003c/em\u003e, \u003cstrong\u003e9\u003c/strong\u003e(10): 552-560.\u003c/li\u003e\n\u003cli\u003eMinnett, P. J., A. Alvera-Azc\u0026aacute;rate, T. M. Chin, G. K. Corlett, C. L. Gentemann, I. Karagali, X. Li, et al. 2019. Half a century of satellite remote sensing of sea-surface temperature. \u003cem\u003eRemote Sensing of Environment\u003c/em\u003e, \u003cstrong\u003e233\u003c/strong\u003e: 111366.\u003c/li\u003e\n\u003cli\u003eMinnich, R. A., M. G. Barbour, J. H. Burk, and J. Sosa‐Ramirez. 2000. Californian mixed‐conifer forests under unmanaged fire regimes in the Sierra San Pedro M\u0026aacute;rtir, Baja California, Mexico. \u003cem\u003eJournal of Biogeography\u003c/em\u003e, \u003cstrong\u003e27\u003c/strong\u003e(1): 105-129.\u003c/li\u003e\n\u003cli\u003eNicholls, R. J., P. P. Wong, V. R. Burkett, J. Codignotto, J. Hay, R. McLean, S. Ragoonaden, and C. D. Woodroffe. 2007. Coastal systems and low-lying areas. In M.L. Parry, O.F. Canziani, J.P. Palutikof, P.G. van der Linden, and C.E. Hanson, editors. \u003cem\u003eClimate change 2007: impacts, adaptation and vulnerability. Contribution of Working Group II to the fourth assessment report of the Intergovernmental Panel on Climate Change\u003c/em\u003e, pp. 315-356. Cambridge University Press, Cambridge, UK.\u003c/li\u003e\n\u003cli\u003eNiering, W. A., and R. S. Warren. 1980. Vegetation patterns and processes in New England salt marshes. \u003cem\u003eBioscience\u003c/em\u003e, \u003cstrong\u003e30\u003c/strong\u003e(5): 301-307.\u003c/li\u003e\n\u003cli\u003eNixon, S. W. 1982. The ecology of New England high salt marshes: a community profile. U.S. Fish and Wildlife Service, US Department of the Interior.\u003c/li\u003e\n\u003cli\u003eNixon, S. W., and C. A. Oviatt. 1973. Ecology of a New England salt marsh. \u003cem\u003eEcological Monographs\u003c/em\u003e, \u003cstrong\u003e43\u003c/strong\u003e(4): 463-498.\u003c/li\u003e\n\u003cli\u003ePaine, R. T., M. J. Tegner, and E. A. Johnson. 1998. Compounded perturbations yield ecological surprises. \u003cem\u003eEcosystems\u003c/em\u003e, \u003cstrong\u003e1\u003c/strong\u003e: 535-545.\u003c/li\u003e\n\u003cli\u003ePennings, S. C., and M. D. Bertness. 2001. Salt marsh communities. In M.D. Bertness, S.D. Gaines, and M.E. Hay, editors.\u003cem\u003e Marine community ecology\u003c/em\u003e, pp. 289-316. Sinauer Associates, Sunderland, MA.\u003c/li\u003e\n\u003cli\u003ePennings, S. C., M. B. Grant, and M. D. Bertness. 2005. Plant zonation in low-latitude salt marshes: disentangling the roles of flooding, salinity and competition. \u003cem\u003eJournal of Ecology\u003c/em\u003e, \u003cstrong\u003e93\u003c/strong\u003e: 159-167.\u003c/li\u003e\n\u003cli\u003ePennings, S. C., C. K. Ho, C. S. Salgado, K. Więski, N. Dav\u0026eacute;, A. E. Kunza, and E. L. Wason. 2009. Latitudinal variation in herbivore pressure in Atlantic Coast salt marshes. \u003cem\u003eEcology\u003c/em\u003e, \u003cstrong\u003e90\u003c/strong\u003e(1): 183-195.\u003c/li\u003e\n\u003cli\u003ePethick, J. S. 1992. Saltmarsh geomorphology. In J.R.L. Allen and K. Pye, editors. \u003cem\u003eSaltmarshes: Morphodynamics, conservation and engineering significance\u003c/em\u003e, pp. 41-62. Cambridge University Press, Cambridge, UK.\u003c/li\u003e\n\u003cli\u003eRedfield, A. C. 1972. Development of a New England salt marsh. \u003cem\u003eEcological Monographs\u003c/em\u003e, \u003cstrong\u003e42\u003c/strong\u003e(2): 201-237.\u003c/li\u003e\n\u003cli\u003eRomijn, E., M. Herold, L. Kooistra, D. Murdiyarso, and L. Verchot. 2012. Assessing capacities of non-Annex I countries for national forest monitoring in the context of REDD+. \u003cem\u003eEnvironmental Science \u0026amp; Policy\u003c/em\u003e, \u003cstrong\u003e19\u003c/strong\u003e: 33-48.\u003c/li\u003e\n\u003cli\u003eRonneberger, O., P. Fischer, and T. Brox. 2015. U-net: Convolutional networks for biomedical image segmentation. a\u003cem\u003erXiv\u003c/em\u003e:1505.04597\u003c/li\u003e\n\u003cli\u003eSallenger Jr, A. H., K. S. Doran, and P. A. Howd. 2012. Hotspot of accelerated sea-level rise on the Atlantic coast of North America. \u003cem\u003eNature Climate Change\u003c/em\u003e, \u003cstrong\u003e2\u003c/strong\u003e(12): 884-888.\u003c/li\u003e\n\u003cli\u003eScheffer, M., S. Carpenter, J. A. Foley, C. Folke, and B. Walker. 2001. Catastrophic shifts in ecosystems. \u003cem\u003eNature\u003c/em\u003e, \u003cstrong\u003e413\u003c/strong\u003e(6856): 591-596.\u003c/li\u003e\n\u003cli\u003eSchuetzenmeister, A., and F. Dufey. 2020. VCA: variance component analysis. R package version 1.4.5. Retrieved from https://cran.r-project.org/web/packages/VCA/\u003c/li\u003e\n\u003cli\u003eScott, M. E. 1988. The impact of infection and disease on animal populations: implications for conservation biology. Conservation biology, 2(1), 40-56. \u003c/li\u003e\n\u003cli\u003eSilliman, B. R., and J. C. Zieman. 2001. Top‐down control of \u003cem\u003eSpartina alterniflora\u003c/em\u003e production by periwinkle grazing in a Virginia salt marsh. \u003cem\u003eEcology\u003c/em\u003e, \u003cstrong\u003e82\u003c/strong\u003e(10): 2830-2845.\u003c/li\u003e\n\u003cli\u003eSmith, J.A., Adamowicz, S.C., Wilson, G.M. and Rochlin, I., 2021. \u0026ldquo;Waffle\u0026rdquo; pools in ditched salt marshes: assessment, potential causes, and management. Wetlands Ecology and Management, pp.1-17. \u003c/li\u003e\n\u003cli\u003eSmith III, T. J., and W. E. Odum. 1981. The effects of grazing by snow geese on coastal salt marshes. \u003cem\u003eEcology\u003c/em\u003e, \u003cstrong\u003e62\u003c/strong\u003e(1): 98-106.\u003c/li\u003e\n\u003cli\u003eSmith, J.A. and Pellew, M., 2021. Pond dynamics yield minimal net loss of vegetation cover across an unditched salt marsh landscape. \u003cem\u003eEstuaries and Coasts\u003c/em\u003e, \u003cem\u003e44\u003c/em\u003e(6), pp.1534-1546.\u003c/li\u003e\n\u003cli\u003eTaylor, N. 1938. A preliminary report on the salt marsh vegetation of Long Island, New York. \u003cem\u003eNew York State Museum Bulletin\u003c/em\u003e, \u003cstrong\u003e316\u003c/strong\u003e: 21-84.\u003c/li\u003e\n\u003cli\u003eTemmerman, S., P. Meire, T. J. Bouma, P. M. J. Herman, T. Ysebaert, and H. J. De Vriend. 2013. Ecosystem-based coastal defence in the face of global change. \u003cem\u003eNature\u003c/em\u003e, \u003cstrong\u003e504\u003c/strong\u003e(7478): 79-83.\u003c/li\u003e\n\u003cli\u003eTitus, J. G. 1988. Sea level rise and wetland loss: an overview. In J.G. Titus, editor. \u003cem\u003eGreenhouse Effect, Sea Level Rise, and Coastal Wetlands\u003c/em\u003e, pp. 1-36. US Environmental Protection Agency, Washington, D.C.\u003c/li\u003e\n\u003cli\u003eTkachenko, M., M. Malyuk, A. Holmanyuk, and N. Liubimov. 2020. Label Studio: Data labeling software. Retrieved from https://github.com/heartexlabs/label-studio\u003c/li\u003e\n\u003cli\u003eTonjes, D. J. 2013. Impacts from ditching salt marshes in the mid-Atlantic and northeastern United States. \u003cem\u003eEnvironmental Reviews\u003c/em\u003e, \u003cstrong\u003e21\u003c/strong\u003e(2): 116-126.\u003c/li\u003e\n\u003cli\u003eTurner, W., C. Rondinini, N. Pettorelli, B. Mora, A. K. Leidner, Z. Szantoi, G. Buchanan, et al. 2015. Free and open-access satellite data are key to biodiversity conservation. \u003cem\u003eBiological Conservation\u003c/em\u003e, \u003cstrong\u003e182\u003c/strong\u003e: 173-176.\u003c/li\u003e\n\u003cli\u003eU. S. Fish and Wildlife Service. 2018. National Wetlands Inventory website. Available at https://www.fws.gov/wetlands/data/Data-Download.html, accessed 08/28/2018\u003c/li\u003e\n\u003cli\u003eWatson, E. B., K. B. Raposa, J. C. Carey, C. Wigand, and R. S. Warren. 2017. Anthropocene survival of southern New England\u0026rsquo;s salt marshes. \u003cem\u003eEstuaries and Coasts\u003c/em\u003e, \u003cstrong\u003e40\u003c/strong\u003e(3): 617-625.\u003c/li\u003e\n\u003cli\u003eWicaksono, P., P. A. Aryaguna, and W. Lazuardi. 2019. Benthic habitat mapping model and cross validation using machine-learning classification algorithms. \u003cem\u003eRemote Sensing\u003c/em\u003e, \u003cstrong\u003e11\u003c/strong\u003e(11): 1279.\u003c/li\u003e\n\u003cli\u003eWoodcock, C. E., R. Allen, M. Anderson, A. Belward, R. Bindschadler, W. Cohen, F. Gao, et al. 2008. Free access to Landsat imagery. \u003cem\u003eScience\u003c/em\u003e, \u003cstrong\u003e320\u003c/strong\u003e(5879): 1011.\u003c/li\u003e\n\u003cli\u003eW\u0026uuml;est, R. O., N. E. Zimmermann, D. Zurell, J. M. Alexander, S. A. Fritz, C. Hof, H. Kreft, et al. 2020. Macroecology in the age of Big Data\u0026ndash;Where to go from here? \u003cem\u003eJournal of Biogeography\u003c/em\u003e, \u003cstrong\u003e47\u003c/strong\u003e(1): 1-12.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"estuaries-and-coasts","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"esco","sideBox":"Learn more about [Estuaries and Coasts](https://www.springer.com/journal/12237)","snPcode":"12237","submissionUrl":"https://www.editorialmanager.com/esco/","title":"Estuaries and Coasts","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"anthropogenic impacts, coastal wetland, disturbance, habitat mapping, landscape ecology, algorithm, marsh restoration, mosquito ditching, remote sensing, salt marsh","lastPublishedDoi":"10.21203/rs.3.rs-6387803/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6387803/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe loss of salt marshes and their ecosystem services following anthropogenic disturbances necessitates restoration built on a scale-dependent understanding of how the prevalence and intensity of these disturbances are linked to ecosystem functioning. A conspicuous legacy modification of marshes, which lacks a standardized and scale-able assessment, is ditching. Consequently, U.S. Atlantic coast resource managers must devote limited resources to quantifying local-scale ditching or make restoration decisions based on a literature of subjective ditching assessments (low vs. high) from a subset of locations with contradictory impacts to ecosystem functions. Here, we combined freely available satellite imagery with machine learning to generate a multi-scale database of ditching prevalence and intensity in 634 marshes from Maine through Florida. The algorithm consistently detected ditches despite the heterogeneous appearance of this disturbance and marshes across regions, seasons, and tidal stages. In contrast to the oft-quoted historical ditching prevalence of 90%, the algorithm quantified a much lower current average of 38%, with the size of this discrepancy varying regionally from an average prevalence of 87% in the Gulf of Maine to 20% in the South Atlantic Bight. Ditching intensity showed further hierarchical spatial variation, but at the state and within-state levels, as opposed to regional level. Within regions, intensely ditched states (5% area removed) were opposed by mildly ditched states (1.9% area removed). With this standardized database of ditching prevalence and intensity, researchers and resource managers may now conduct scale-dependent assessments of ecosystem responses to ditching to inform restoration and management of this valuable habitat.\u003c/p\u003e","manuscriptTitle":"Machine learning reveals hierarchical spatial patterns in salt marsh ditching along U.S. Atlantic Coast","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-04-29 14:33:12","doi":"10.21203/rs.3.rs-6387803/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewerAgreed","content":"","date":"2025-04-28T17:03:27+00:00","index":0,"fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-04-25T14:28:24+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"Estuaries and Coasts","date":"2025-04-19T10:53:46+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-04-07T04:20:26+00:00","index":"","fulltext":""},{"type":"submitted","content":"Estuaries and Coasts","date":"2025-04-06T12:21:20+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"estuaries-and-coasts","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"esco","sideBox":"Learn more about [Estuaries and Coasts](https://www.springer.com/journal/12237)","snPcode":"12237","submissionUrl":"https://www.editorialmanager.com/esco/","title":"Estuaries and Coasts","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"33029098-7520-4a8d-8ae7-57c73c9639ff","owner":[],"postedDate":"April 29th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2025-04-29T14:33:12+00:00","versionOfRecord":[],"versionCreatedAt":"2025-04-29 14:33:12","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6387803","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6387803","identity":"rs-6387803","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}