Our study focused on a single marsh (North Creek Marsh, 37°36′40.20″N, 122°6′43.94′W) at Eden Landing Ecological Reserve in Hayward, CA, part of the South Bay Salt Pond Restoration Project (Figure 1). The SBSPRP is an adaptively-managed effort to restore over 15,000 acres of former salt-evaporation ponds to a mosaic of tidal wetlands and managed ponds (Trulio et al., 2007). North Creek Marsh is a 37.32 Ha restoration site initiated in 2006. The site was historically tidal wetland and was converted to industrial salt-evaporation in the late nineteenth century (Stanford et al., 2013). Tidal influence was returned to the area by breaching a levee at the southern end of the site. The restoration process is driven by tidal transport of sediment building the marsh plain to the appropriate level (Brew and Williams, 2010), then seed dispersal via tidal hydrochory driving the development of vegetation (Diggory and Parker, 2011). In addition to the passive restoration process via seed dispersal, the Invasive Spartina Project actively planted selected portions of the site with the native cordgrass Spartina foliosa, Distichlis spicata (saltgrass), and Grindelia stricta (marsh gumplant) (Hammond, 2016).

We determined mean annual salinity for each rain year (October–September) between 2009 and 2015 using Station 30 from the USGS SF Bay water quality archive (Cloern and Schraga, 2016). To explore potential differences between tidal heights, we determined mean higher high water (MHHW) and monthly maximum tide from the NOAA Redwood City Tide Gauge, the closest tide station with continuous data over the study period (https://tidesandcurrents.noaa.gov/). For salinity and MHHW, we subset the data for rain years 2009–2011 and 2012–2015 to correspond to the dates of our imagery and California's historic drought. To determine differences between the two periods, we performed a non-parametric Kruskal–Wallis test for salinity, MHHW and monthly maximum tide. To determine directional trends in salinity during the two periods, we used non-parametric generalized additive models to analyze salinity levels over time using the gam package in R (Hastie, 2013). Non-parametric tests were used due to non-normality of salinity data.

For 2009 and 2011, we obtained 0.8 m pan sharpened IKONOS imagery of the South Bay Salt Pond Restoration Project from the San Francisco Estuary Institute (^©Digital Globe Inc., 2011). For 2015, we obtained 0.5 m WorldView-2 imagery (^©Digital Globe Inc., 2015). Each set of imagery contained four spectral bands: red, green, blue, and near-infrared (NIR). To ensure phenological continuity between collection dates, all images were collected near peak biomass (June 23 2009, July 7 2011, and June 21 2015) at low tide to ensure maximum visibility of vegetation. The timing of collection is essential because tidal water frequently covers landscape features, such as vegetation patches, essential to change detection. To double check that intermediate years at our site did not exhibit anomalous vegetation growth that is not accounted for in our analysis, we reviewed Google Earth imagery (^©Google, 2017) for all available dates between June 2009 and June 2015. We did not find evidence of anomalous change or loss in the periods between our high resolution images.

To prepare the images for analysis, we re-projected the 2009 image from the GCS 1984 datum to the NAD 1983 datum to match the 2011 and 2015 images. We down-sampled all images to 0.8 m pixel resolution to match the lowest resolution images. We then geocorrected all images, resulting in an offset of 0.5 pixel maximum. Images were imported into eCognition (^©Trimble Inc.) software to perform OBIA. To allow for the most effective interpretation of vegetation patches, bands 4, 3, and 2 were visualized as RGB, respectively, and the Histogram Equalization stretch was applied across the image.

Object-based analyses were performed in eCognition Developer software version 8.8 (^©Trimble Inc.). As a first step, we generated primitive image objects as spatial units for wetland classification using the Multiscale Resolution Segmentation (MRS) tool which requires the parameters of scale, shape and compactness to control object size and heterogeneity. For all images, we used the red, green, blue, and infrared bands to classify imagery. To determine their values for our objectives, we worked through a series of scale parameter values in increments of 5, and both shape and compactness parameters in increments of 0.1. We assessed each combination of settings by trial and error to determine which combination of parameters best matched the visual distribution of vegetation at the site. Notably, due to the differences in the original resolution of image datasets, we had to individually adjust their MRS parameters to obtain primitive objects of comparable size. For the 2011 image, using a scale of 10 resulted in unrealistically small objects. Using scales of 40 and above did not capture enough of the surface variation, and after comparison of incremental steps, we determined that a scale of 30 most effectively captured the vegetation patterning on the marsh surface. We selected a scale of 25 for the 2015 image and a scale of 6 for the 2009 image. For all images, shape was given low weight (0.1) in the final classification, as shapes in wetland vegetation are highly dependent on patch size and do not conform to regular patterns across the marsh surface (Moffett and Gorelick, 2013). Compactness was given a medium weight (0.5). For all images, the four bands were given equal weight.

Following the segmentation process, we manually identified at least 50 training samples for each of the three main categories: Water/Channels, Mudflat, and Vegetation. Vegetation is included as a simple category since the majority of vegetation at the site consists of Salicornia pacifica, an early-colonizing marsh dominant (Krause, 2016). Jaumea carnosa (Fleshy Jaumea), Frankenia salina (Alkali Heath), the annual Salicornia europaea (common glasswort), G. stricta (marsh gumplant), and S. foliosa (California cordgrass) are present in lower densities due to natural recruitment (Krause, 2016) and planting (Hammond, 2016), but our imagery did not allow for differentiation between species. Samples were selected by examining the imagery and cross-referencing these observations with checks of Google Earth (^©2015 Google) imagery to verify vegetation patterns. This information was combined with expert knowledge on vegetation patterns from field visits conducted between 2013 and 2015. Once samples were selected, images were classified by including a supervised nearest neighbor process algorithm with the mean brightness, mean NIR and standard deviation of the red band selected as class-discriminating features. We initially included the Normalized Difference Vegetation Index (NDVI), which uses the red and infrared bands to detect green vegetation, as a classification parameter. However, this led to spurious identification of algae as vegetation, and misclassified vegetated areas with apparent mudfilms as mudflat, so we elected not to include it in the final process decision tree. Following sample selection and implementation of the nearest neighbor algorithm, images from all years were separately classified into the three categories using the classification algorithm in eCognition. Once images from each year had been classified, the resulting classifications were imported into ENVI to perform change detection analysis via simple spatial overlay. Images were masked to include only the marsh-plain area.

To perform accuracy assessment, we used the Random Points tool (Standard C Rand function) in ArcGIS v. 10.3 (Esri Inc.) to select between 54 and 87 points per category per year, excluding training samples, and visually identified cover categories. Samples that fell along object edges were excluded from the random point selection. Google Earth images (^©2015 Google) from each year were used to manually verify sample collection points. These points were imported as Regions of Interest (ROI) into ENVI v.5.2 (Harris Geospatial Inc.) software to perform accuracy analysis. The ROIs were used to populate the Confusion Matrix tool, which calculates standard accuracy metrics (overall accuracy, kappa, user's, and producer's accuracies for different classes) of a classified image based on verified samples.

Following classification, we analyzed vegetation patch dynamics. To determine the relationship between vegetation presence and channel structure, we digitized a vector of the major channels at the site, then created a distance raster using the Euclidean Distance tool in ArcGIS v. 10.3 (Esri Inc.). This tool calculates the distance from a specified feature and outputs a continuous raster with corresponding values. We generated 1,000 random points using the Random Points tool in ArcGis and extracted the vegetation layer from our classification for each year. Based on this data we used vegetation presence (1) and absence (0) to run a generalized linear model with a binomial distribution using the lme4 package in R (Bates et al., 2017). To determine changes in patch configuration across the three images, we ran patch statistics using FragStats v. 4 (McGarigal et al., 2015).

Our results show that salinity was significantly higher during California's historic drought, and the magnitude of mean annual vegetation change was 10.4 times slower during this period compared to the lower salinity period that preceded it (Figures 2, 3). Mean salinity was 25.64 ppt for 2009–2010, and 23.99 ppt for 2010–2011, with an overall mean of 24.82 ppt (CV = 0.198) between 2009 and 2011. Mean salinity was 26.08 ppt for 2011–2012, 28.18 ppt for 2012–2013, 30.12 ppt for 2013–2014, and 29.50 ppt for 2014–2015, with a mean salinity of 28.47 ppt (CV = 0.10) between 2011 and 2015 (Figure 5). Salinity was significantly different between these two periods (p < 0.001, χ² = 18.40). Salinity significantly decreased between 2009 and 2011 (p < 0.001, F = 18.69) and significantly increased between 2011 and 2015 (p < 0.001, F = 16.50). Neither MHHW (p = 0.354, χ² = 0.86) nor monthly maximum tide was significantly different between the two periods (p = 0.354, χ² = 43.87) (Figure 2).

We obtained high classification accuracy for each of our cover categories in each year. For 2009, we obtained an overall accuracy of 92.42% and a Kappa Coefficient of 0.88. For 2011, we obtained an overall accuracy of 95.02% and a Kappa Coefficient of 0.92. For 2015, we obtained an overall accuracy of 96.83% and a Kappa Coefficient of 0.95. The lower overall accuracy in the 2009 image was due to over-classification of water on the marsh surface (Table 1). Vegetation, the focal target of post-restoration monitoring, was consistently classified with high user's and producer's accuracy exceeding 92% at all times (Table 1). It was most commonly misclassified with water in 2009 and 2015 and mudflat in 2011. Some of the overall classification error also occurred due to misclassification of water and mudflats that did not correspond to vegetation per se and thus was of lower concern for our objectives.

Total vegetation cover increased from 58,154 m² of the study area to 99,315 m² from 2009 to 2011, an increase of 70.77% at a mean rate of 20,580 m²/year. In contrast, vegetation cover increased from 99,315 m² in 2011 to 107,232 m² in 2015, a 7.97% change from the 2011 cover at a mean rate of 1,979 m²/year (Figures 3, 4). For all years, vegetation presence was significantly related to distance from channel, with areas closer to channel more likely to support vegetation, but the magnitude of the effect was notably larger in the 2015 image (2009: p < 0.001, z = −3.49; 2011: p = 0.002, z = −2.98; 2015: p < 0.001, z = −6.33). In the 2011 image, we observed some vegetation colonization of interior mudflat areas that did not persist in the 2015 image (Figures 4, 5). The overall number of patches decreased from 2009 (394 patches) to 2011 (282 patches) and increased in 2015 (473 patches). Mean patch area was the largest in 2011 (352 m²), intermediate in 2015 (226 m²), and smallest in 2009 (147 m²; Table 2).

Among non-vegetated surfaces, mudflats were the most prevalent cover class across all years, declining slightly in 2015, with total cover of 245,413 m² in 2009, 247,685 m² in 2011, and 230,752 m² in 2015. Since the amount of water in aerial images is highly dependent on the timing of image, tidal phase, and other stochastic factors, changes in water coverage should be interpreted with caution. In our images, water accounted for 69,764 m² in 2009, 26,188 m² in 2011, and 34,787 m² in 2015 (Figures 3, 4).

Our results demonstrate that drought may impact vegetation change rates in Mediterranean-type tidal wetland restoration projects, leading to non-linear recovery patterns. At North Creek Marsh, vegetation cover increased from 2009 to 2011 and from 2011 to 2015, but the mean annual rate of change during the first period, when Bay salinity was lower, was more than 10 times as rapid as change during the second period, when historic drought conditions elevated salinity levels in the Bay. By employing remotely sensed imagery to study this progression, we were able to scale up from previous field efforts that demonstrated the effect of lowered salinity on plant productivity (Zedler, 1983; Schile et al., 2011; Woo and Takekawa, 2012), and restoration trajectories (Chapple et al., 2017). Previous work from Southern California documented increased rates of S. foliosa establishment in response to increased sedimentation rates brought on by El Nino events (Ward et al., 2003) and increased Spartina biomass and structure in response to lowered salinity brought on by El Nino events (Zedler, 1983; Zedler et al., 1986). Our results show that freshwater availability may also influence the rate of vegetation expansion in recently restored wetlands dominated by S. pacifica. These larger-scale observations are supported by experimental results that demonstrate that increased salinity levels reduced S. pacifica biomass production (Schile et al., 2011; Woo and Takekawa, 2012). In contrast to our site, a similar restoration project in a more freshwater marsh without a notable drought period reached 90% vegetated over a 10 year period, with no evidence of slowing pace after initial gains (Tuxen et al., 2008). This indicates that restoration projects in higher salinity regions may exhibit more variable, less linear trajectories due to interannual variability in salinity. While increased inundation during periods of higher rainfall could be another factor influencing vegetation change, we found no significant difference in MHHW or monthly maximum tide between the two periods.

Our results also demonstrate that channel structure is a key determinant of where vegetation establishes, and may be even more important during periods of elevated salinity. Vegetation was significantly associated with channel proximity for all years, but between 2011 and 2015, the strength of the interaction between vegetation presence and channel proximity more than doubled. We also visually observed establishment of new vegetation patches in interior marsh areas in 2011 (Figure 4), but these patches did not persist in 2015. Channels drive the restoration process by improving drainage across the marsh surface and lowering salinity (Sanderson et al., 2000; Williams and Orr, 2002; Wallace et al., 2005; O'Brien and Zedler, 2006). Biomass production of S. pacifica is significantly influenced by elevated salinity in in poorly drained areas, but has no effect in well-drained areas adjacent to channels (Schile et al., 2011). Our results indicate that salinity levels likely interact with the channel structure at the site, allowing vegetation to persist and expand in areas adjacent to channels but precluding development in poorly drained interior areas. Under projected climate change scenarios, increased prevalence of drought is likely to reduce snowpack and increase salinity (Callaway et al., 2007). This may slow the overall rate of vegetation change and increase the importance of channel structure in the restoration process.

Our results also show how OBIA can be used to overcome some of the challenges with high resolution data to map vegetation change over time in developing tidal wetlands. The dynamic nature of tidal processes mean that images are often different from each other based on how mud and water appear in the image, which can present problems for comparing images from different years (Dronova, 2015; Campbell et al., 2017). Furthermore, local noise and spectral variation, especially pronounced at higher spatial resolution, pose considerable challenges for delineating wetland cover type patches as semantic entities (Moffett and Gorelick, 2013), particularly at early post-restoration stages with higher spatial heterogeneity (Tuxen and Kelly, 2008; Tuxen et al., 2008; Kelly et al., 2011). By using object-based methods, we were able to create realistic objects for our cover types that produced high levels of accuracy, allowing for comparison between years at high spatial resolution. While NDVI has historically been employed as a means of detecting vegetation, we found that classification parameters that relied too heavily on NDVI led to classification of areas with green algae on the mudflat surface as vegetation. By also taking into account spatial parameters, our object-based approach minimized spurious mapping of vegetation that may occur when using pixel-based change methods. Our results highlight the distinct benefit of using OBIA in assessing early stages of restoration project development to capture fine scale change and to streamline semi-automated vegetation detection despite some degree of required specificity of methods and parameters at individual dates. Although OBIA benefits in wetland analyses have long been recognized (Tuxen and Kelly, 2008; Dronova, 2015), this methodology is still under-utilized in the context of restoration monitoring (Klemas 2013) and offers powerful opportunities for cost-effective, spatially comprehensive, and repeated characterizations of vegetation development and landscape structure.

Notably, different algorithm parameters were needed for each image to produce images with the highest accuracy. We were able to attain a high level of accuracy across all three images, but accuracy was slightly lower in the in the 2009 imagery, when algae and surface water led to more confusion between classes, highlighting the importance of date-specific conditions on wetland surface analysis in tidal systems. Distributions of water and mud across the landscape were mapped differently in different years, due to different tidal heights at the time of collection and evolving morphology of landscape topography that likely led to retention of water in different areas across the years. We suggest that changes between mudflat and water should be interpreted with caution, since they are highly temporally variable and sensitive to when imagery was collected. While vegetation increased overall, there were also notable areas of localized vegetation loss (particularly in areas farther from channels), which indicates that the site is still evolving. We expect that efforts to monitor multiple restoration sites will likely need to create separate classifications for each site to minimize the impact of unique surface conditions at a given tidal stage and surface variability on classification effectiveness.

In addition to the effects of wet years and drought, the trends we observed are likely influenced by a combination of other factors. In the commonly accepted models of tidal wetland development, sedimentation rates are expected to slow as the marsh plain reaches equilibrium with tidal inundation (Morris et al., 2002; Williams and Orr, 2002; D'Alpaos et al., 2012; Schile et al., 2014), which could explain the observed decrease in the rate of vegetation expansion we observed. However, sedimentation data collected at the site shows that annual sedimentation rates between the breach date in 2006 and 2013 were marginally slower (1.21 cm/year) than between 2013 and 2016 (1.33 cm/year), when drought conditions persisted (Krause, 2016). This indicates that the decreased rate of vegetation expansion is not due to decreased rates of sedimentation. Further, between 2012 and 2015, S. foliosa was planted across the study site (Hammond, 2016). Since these plantings were largely adjacent to areas of existing vegetation, they may have contributed to the expansion we observed, which means that rates of natural expansion during the drought years may have been even lower than our results indicate. Lastly, our analysis of tidal height data shows that differences in tidal inundation did not differ between the wet and dry periods.

The inability to detect species-level trends is an important limitation of our study. In addition to the S. foliosa plantings, the tidal wetland sub-dominant species F. salina (Alkali Heath) and J. carnosa (Fleshy Jaumea) were also present at the site in very low densities (Krause, 2016). Work from older restoration and reference sites in the north SF Bay indicates that Bay salinity can also influence the dynamics of sub-dominant species (Chapple et al., 2017), which may be a promising direction for future studies in these areas. However, S. pacifica is the dominant species in the early stages of restoration in the area, and is responsible for the majority of vegetation cover. One of the major implications of rates of vegetation change is the ability of developing restoration projects to keep pace with sea level rise (Goals Project, 2015), so for the purposes of our study understanding overall rates of vegetation change is appropriate. Advancing this OBIA-based monitoring framework to develop a capacity to detect species-level transitions in the future is an important research need that could benefit from the advances in high-resolution hyperspectral platforms (Santos et al., 2011; Lucieer et al., 2014).

Our results demonstrate that considering non-linear post-restoration site development trajectories that are dependent on weather may be crucial for structuring adaptive management decisions in variable climates. A detailed understanding of how weather interacts with site geomorphology to influence outcomes is important for planning effective restoration efforts (Holmgren and Scheffer, 2001; Vaughn and Young, 2010; Sitters et al., 2012; Chapple et al., 2017). Importantly, slower progress of vegetation is not entirely negative, as the intermediate habitat mosaic of vegetation, mudflat, and water provides habitat for a number of avian species (Moss, 2015). However, given that the rapid re-vegetation of tidal wetland restoration projects is considered to be one of the best means of allowing developing sites to keep pace with sea-level rise (Goals Project, 2015), understanding the role of weather in determining these rates will be essential for managing projects that are resilient to climate change.

Developing reproducible remote sensing techniques is a promising, potentially cost effective means of monitoring change in these projects over time. Future efforts should explore change over multiple sites to discern how generalized these weather-dependent trends are and how transferable image classification settings are between sites. Sampling restoration sites across a range of salinity levels in the SF Bay would allow for an exploration of how the spatial context of sites might influence their temporal development. Since field sampling is limited by time, scale, funding, and spatial resolution, remotely sensed products hold high promise for addressing these issues.

From a restoration management perspective, our findings supported other work demonstrating that channel edges are hotspots of vegetation development (Sanderson et al., 2000; Wallace et al., 2005; O'Brien and Zedler, 2006). Attempts to add diversity into developing marshes should focus on these areas, a practice which is already in place in the SF Bay (Hammond, 2016). Since we show that interior mudflat areas away from channels may be slow to develop vegetation, proactive manipulation of elevation in these areas prior to restoring tidal access may be one way to speed vegetation development. Further, efforts to actively manipulate channel structure may also help speed the development of vegetation establishment. These actions are likely to be more necessary in areas where salinity levels are currently higher, but may become necessary across a range of sites as climate change shifts salinity distributions in the SF Bay (Callaway et al., 2007). Proactive geomorphic intervention is likely to make these projects more resilient to the impacts of sea level rise.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.