Volume 5, Issue 1 e12844
Open Access

An effective online platform for crowd classification of coastal wetland loss

Sofia Eleni Spatharioti

Sofia Eleni Spatharioti

Microsoft Research, New York, New York, USA

Search for more papers by this author
Eliza Boetsch

Eliza Boetsch

Environmental Studies, Northeastern University, Boston, MA, USA

Search for more papers by this author
Scott Eustis

Scott Eustis

Healthy Gulf, New Orleans, Louisiana, USA

Search for more papers by this author
Kutub Gandhi

Kutub Gandhi

Khoury College of Computer Sciences, Northeastern University, Boston, MA, USA

Search for more papers by this author
Matt Rota

Matt Rota

Healthy Gulf, New Orleans, Louisiana, USA

Search for more papers by this author
Archana Apte

Archana Apte

Environmental Studies, Northeastern University, Boston, MA, USA

Search for more papers by this author
Seth Cooper

Seth Cooper

Khoury College of Computer Sciences, Northeastern University, Boston, MA, USA

Search for more papers by this author
Sara Wylie

Corresponding Author

Sara Wylie

Sociology & Anthropology, Health Sciences, Northeastern University, Boston, MA, USA


Sara Wylie, Northeastern University, 360 Huntington Ave, Boston, MA 02115, USA.

Email: [email protected]

Search for more papers by this author
First published: 07 December 2022

Sofia Eleni Spatharioti and Eliza Boetsch are joint first authors.

Seth Cooper and Sara Wylie are ioint last authors.

Funding information: National Science Foundation, Directorate for Computer and Information Science and Engineering, Grant/Award Number: 1816426


Wetland loss is increasing rapidly, and there are gaps in public awareness of the problem. By crowdsourcing image analysis of wetland morphology, academic and government studies could be supplemented and accelerated while engaging and educating the public. The Land Loss Lookout (LLL) project crowdsourced mapping of wetland morphology associated with wetland loss and restoration. We demonstrate that volunteers can be trained relatively easily online to identify characteristic wetland morphologies, or patterns present on the landscape that suggest a specific geomorphological process. Results from a case study in coastal Louisiana revealed strong agreement among nonexpert and expert assessments who agreed on classifications at least 83% and at most 94% of the time. Participants self-reported increased knowledge of wetland loss after participating in the project. Crowd-identified morphologies are consistent with expectations, although more work is needed to directly compare LLL results with previous studies. This work provides a foundation for using crowd-based wetland loss analysis to increase public awareness of the issue, and to contribute to land surveys or train machine learning algorithms.


Wetland loss is increasing rapidly, and there are gaps in public awareness of the problem. Crowdsourcing image analysis of wetland morphology could accelerate academic and government studies and engage and educate the public. The Land Loss Lookout (LLL) project crowdsources mapping of wetland morphologies associated with wetland loss and restoration. This paper describes results from LLL in which 672 nonexpert participants analyzed one or more of six sets of color near-infrared (CIR) images of Louisiana coastal wetlands for characteristic wetland morphologies associated with wetland loss and restoration.

The Mississippi River Delta, located on Louisiana's Gulf Coast, is one of the world's most fertile landscapes. Annually, the Mississippi River Delta (the delta) provides more than $12 billion and up to $47 billion in benefits through ecosystem services including hurricane and flood protection, drinking water, water quality, and sustaining fisheries (Batker et al., 2010). Coastal wetlands sequester carbon within soils and are net greenhouse gas sinks (Hopkinson et al., 2012). Despite the delta's economic and ecological value, wetlands loss is occurring rapidly in the region (Britsch & Dunbar, 1993; Olea & Coleman, 2014; Törnqvist et al., 2020; Turner & McClenachan, 2018). While this wetland loss is driven by multiple factors, the core issues are that sediment is not supplied to the Louisiana coast at the rate and quantity that it was before the Anthropocene, and that water is encroaching upon land areas (Britsch & Dunbar, 1993). Several drivers contribute to the sediment-water imbalance: subsidence, climate change (including global sea-level rise and erosion due to storms), channelization of rivers, flood-prevention infrastructure, dredging and canal-building (Couvillion et al., 2017). From 1996 to 2010, 87% of U.S. estuarine wetland loss occurred along the northern Gulf of Mexico, and of all estuarine land loss, 80% occurred in Louisiana alone (Gittman et al., 2019). From 1932 to 2016, coastal Louisiana (LA) lost about 25% of the 1932 land area, a net change of approximately −4833 km2. In localized areas like Plaquemines Parish, that rate is ~50% (Couvillion et al., 2017).

Traditionally, government agencies and academic institutions analyze and identify coastal wetland loss (Boesch et al., 1994; Couvillion et al., 2017; Jones, 2016; Yuill et al., 2009). This research is vital, yet policy and economic action to prevent and restore wetland loss have been constrained in Louisiana by the political influence of powerful industrial actors (Houck 2015; Templet, 2022). The former Secretary of the Louisiana Department of Environmental Quality (LDEQ) recently described how “from 1988–1992…[he] witnessed firsthand the role oil and gas played in attempting to downgrade environmental laws, regulations and programs” (Templet, 2022). Louisiana's economy historically depends on industries that exacerbate wetland loss and climate change, particularly oil and gas extraction (Bagstad et al., 2007; Day et al., 2007; Houck, 1983; Mervis, 2020). In many parts of the Gulf of Mexico, including south Louisiana, subsidence is exacerbated by oil and gas industry activity including canal construction and fossil fuel extraction (Beland et al., 2017).

Traditional research can be infrequent, slow, and often fails to engage the public. This is problematic as climate change is occurring rapidly, and as sea levels continue to rise (Couvillion et al., 2017). At present-day rates of relative sea level rise, the submergence of the remaining ~15,000 km2 of coastal Louisiana marshland may be inevitable (Törnqvist et al., 2020). Scientists for Louisiana Coastal Protection and Restoration Authority (LA CPRA) models predict that in coastal Louisiana ​​“75% of the marsh loss was attributed to rising water levels” (Belhadjali, 2016; Schleifstein, 2020). The US public remains largely unaware of the issue and causes of wetland loss; for example, one study showed that Louisiana residents perceived no significant relationship between oil and gas activities and relative sea level rise risk (Altinay et al., 2020).

Citizen science offers potential solutions to these issues by engaging the general public in studying wetland loss, educating participants and enabling more frequent and systematic wetland loss studies. Recent citizen science and crowdsourcing projects show nonexperts can analyze land cover on a large scale. The Geo-wiki project (Fritz et al., 2009), relies on volunteers to improve global land cover maps, and Cropland Capture (Sturn et al., 2015) enlists citizen scientists to classify cropland in images. The DIY landcover project used paid crowdsourcing to identify regions of land cover types (Estes et al., 2016), and in MapMill public volunteers analyzed images for damage after Hurricane Sandy (Munro & Erle, 2013). These projects show nonexpert input can benefit remote sensing image processing for land use, land cover (LULC) analysis. Similarly, LLL engages the general public in the study of wetland loss.

Currently, the majority of crowdsourced citizen science projects analyze the natural world rather than anthropogenic industry. An examination of 888 citizen science publications found biology topics made up 72% of all projects, with the most common objective of studying species' distribution and diversity (Follett & Strezov, 2015). Driven primarily by professional scientists and conservation issues, researchers critique citizen science for depoliticizing environmental issues by focusing on natural science topics rather than more politically and socially contentious issues such as industrial accountability (Cooper et al., 2021; Kimura & Kinchy, 2016; Ottinger, 2017). Oil and gas corporations have even employed citizen science in public relations strategies (Blacker et al., 2021). Additionally, many citizen science research projects are determined by academic interests rather than community interests (Hendricks et al., 2022; Kasperowski & Kullenberg, 2019).

To address these issues, Healthy Gulf, a regional environmental organization dedicated to the protection and restoration of the Gulf Coast's natural resources, partnered with Northeastern University's Cartoscope to develop LLL. Cartoscope is a citizen science platform that supports nonprofits organizations, rather than academics, to develop citizen science projects and studies how to improve citizen scientists' engagement and enjoyment (Spatharioti, 2020). This paper describes LLL as an example of nonprofit lead citizen science, examines the efficacy of crowd labels in terms of participant engagement and participant accuracy, the project's contribution to analyzing wetland morphologies and whether LLL increased nonexperts' knowledge of wetland loss.


2.1 Site selection

LLL examined an area of the lower Barataria watershed on the West Bank of Jefferson Parish, and the East and West Banks of Plaquemines Parish, Louisiana, that is representative of where wetland loss is occurring throughout the state. This pilot study included six wetland morphologies common to most Gulf Coast wetlands (Table 1). Wetland loss patterns between 1932 and 1990 in this area were analyzed in two other studies in 2000, making it possible, at a later date, to compare land changes across time periods (Penland et al., 2000a, 2000b). As our study only involves image classification, we use the word “morphology” as defined in Wayne et al. (1993) to describe “the physical form of land loss areas…[that] cannot imply action or process” (125). Wayne et al.'s land loss classification scheme informed two systematic assessments of wetland loss in Louisiana (Penland et al., 2000a, 2000b) on which we base the six wetland morphologies analyzed in this study. The area was selected to limit the number of ambiguous drivers of land loss, as documented by Penland et al. (2000b). For example, the study area contained only one small area (less than 5km2 area, or ~ two images in our study) designated by Penland et al. as “Altered Hydrology- Multiple” because of the ambiguity of that process classification (Penland et al., 2000b).

TABLE 1. Wetland morphology categories examined in LLL and related scientific terminology, characteristics, associated mechanism and documentation in scientific literature
Project category (colloquial subproject title) Scientific terms for wetland morphology Characteristics Mechanisms behind this morphology Scientific sources
Shoreline erosion Shoreline erosion Concave arc patterns meeting in sharp points along a shoreline (Penland et al., 2000a) Marsh erosion from wind and tide generated waves in larger local water bodies Penland et al., 2000b “Erosion: Natural Wave”
Shipping Direct removal navigation Long, straight, relatively wide waterbody dredged to create a shipping channel. At the scale of our photography these channels have no endpoint within the field of view. Occasionally ships are visible Dredging

Penland et al., 2000b: “Direct Removal: Navigation Channel”a

Oil and gas Direct removal: oil/gas channel Irregular, smaller canals, often with dead-end branches, dredged for drilling access or pipelines Dredging

Penland et al., 2000b: “Direct Removal: Oil/Gas Channel”

Submergence: altered hydrology—oil/gas Many irregular ponds present next to the canals


Marsh submergence from the increased water levels due to water impoundment by oil and gas canals

Oil and gas canals can alter local hydrology, leading to marsh waterlogging that produces many irregular ponds in over-flooded low-level marshes (Swenson & Turner, 1987; Turner & Rao, 1990). Penland et al., 2000b: “Altered Hydrology: Oil and Gas.” We excluded areas designated as “Altered Hydrology: Multiple” in Penland et al. (2000b)
Agriculture/farming Agricultural ponds/failed land reclamation Quadrilateral shapes created by a drainage network, used to create an agricultural pond or created as a drainage area that later failed and flooded Flooding of land for agricultural ponds. Failure of “fastlands” and levees erected for an agricultural drainage system Penland et al., 2000b: “Direct Removal: Agricultural Pond” and “Failed Land Reclamation.”
Fastland Parallel, straight lines created by a network of drainage ditches, and quadrilateral borders of rectilinear fields Draining of wetland to produce fields We included unflooded agricultural lands behind the levee system (“Fastlands” layer retrieved from the LDNR in 2020)
Restoration Wetlands restoration Rectilinear areas of land, and areas of sand that appear gray in color-infrared imagery Sand areas created by pumping sediment into rectilinear impoundments to restore a marsh platform elevation (McMann et al., 2017) We refer to a layer from the LDNR, “Consistency Polygons.” As these boundaries are updated by LDNR for precision, we refer to a version downloaded in Spring 2017, which is the closest in date to the date of our study imagery (2016)b
Sea Level Rise


Relative sea level rise

Many irregular ponds caused by interior marsh flooding

If tidal water level is too high for too long during the growing season (if hydroperiod is too long), the marsh plants die off, leading ponds to appear where the lowest-lying marshes once were.

Tidal waters can be too high because of subsidence and sea-level rise due to climate change

Penland et al., 2000b does not name “sea level rise”: climate change was not a major driver of land loss in the 1990's and the topic was politically sensitive.

The closest category in Penland 2000b is “Submergence: Natural Waterlogging,” defined as submergence due to natural subsidence, where immediate human causes of submergence have been eliminated

We use the term “Sea level rise” as in 2016 LA CPRA recognized “Sea level rise” as the leading cause of land loss from 2017 to 2067 (Belhadjali, 2016; Schleifstein, 2020)

We draw on pattern characteristics guided by computer models projecting changes in plant community distribution and composition in response to environmental conditions (J. Visser & Duke-Sylvester, 2017)

  • a We did not create a separate project for “Erosion: Navigation Wave,” for erosion occurring in navigation channels (as described by Penland 2000b) because the “navigation wave erosion” produces the same pattern as natural wave erosion. Therefore, participants classifying shoreline erosion would capture erosion happening in shipping channels.
  • b We did not refer to Penland et al. (2000b) for our restoration data as that paper analyzed images from 1990, when restoration was not a prevalent land pattern, and therefore does not include a restoration layer.

Aerial images for this region are available through the Louisiana Department of Natural Resources (LDNR) for 2016, 2012, and 2008. LA CPRA computer models for this area forecast high-level of wetland loss (Peyronnin et al., 2013; J. M. Visser et al., 2013). A large number of restoration projects were completed in the area by 2016 (LA CPRA, 2017). For Healthy Gulf's nonprofit mission to protect the Gulf, this area is of interest as industrial permits are being considered in this region which would destroy wetlands (F.R. 87 FR 18787, 2022).

2.2 Classifying wetland morphologies in LLL

The research team identified six wetland morphologies with visually identifiable patterns, that were expected in the study region based on previous studies (Penland et al., 2000a, 2000b; J. Visser & Duke-Sylvester, 2017). The six morphologies, and their correspondence with categories from Penland et al. (2000b) and additional literature are described in Table 1. Wetland morphologies were characterized based on CIR aerial photographs (Appendix S1), where the contrast between red (photosynthesizing plant life), white/gray (bare ground or sand) and blue (water) make the land/water boundary readily discernible (Mondejar & Tongco, 2019).

To ensure that wetland morphologies were understandable to nonspecialists, our tutorials translated scientific terminology into more colloquial terms (Table 1). Rather than asking laypeople to identify the characteristics of a typical morphology, we asked if they could see a “pattern”; and we described that morphology based on lay description of a signature classification of the visual characteristics.

As multiple wetland morphologies can be present within a photograph, instead of having participants choose which morphology best represented the area, we divided the entire study into six subprojects, and asked participants a simpler “Yes or No” question, as to whether one of the six morphologies was present or absent. Study photographs were evaluated for all six morphologies.

Some related wetland morphologies were combined for efficiency. For instance, instead of teaching participants to identify oil and gas canals alone, we taught them to recognize pools that form in marshes, because oil and gas canals can reduce marsh drainage. Therefore, wetland morphologies associated with waterlogging were included in the “Oil and Gas” wetland morphology category and the “Sea Level Rise” wetland morphology category. This enables comparison across the “Oil and Gas” and “Sea Level Rise” subproject results to assess where waterlogging occurs with and without oil and gas canals.

A full geomorphological study would extend beyond a visual assessment of images and involve fieldwork particularly because a local driver of water level and a regional driver of water level are indistinguishable in aerial photographs–and in many areas both the local hydrological impacts of oil and gas canals, and the regional water level changes from relative sea level rise are likely raising the same water level. Therefore, this study did not assign a specific mechanism to each wetland morphology, but rather assessed whether laypeople could categorize wetland morphologies as effectively as experts.

2.3 Study design

Based on these categories we developed six parallel subprojects for each wetland morphology. Participants were randomly sorted into one of the six subprojects upon entering LLL and given a tutorial for one wetland morphology, to reduce their cognitive overload.

Tutorials were developed by Cartoscope and Healthy Gulf (see Supplementary materials) and two external experts reviewed the tutorials prior to sharing the projects. The tutorials taught participants how to examine CIR aerial photographs and to identify the unique patterns visible for each wetland morphology. Tutorials began with an annotated CIR aerial image, with written descriptions of the morphology's key characteristics. Participants were then presented with five to six unmarked practice images. Participants selected whether the morphology is present or absent in the practice images and then informed whether their answers are correct or incorrect and why. No third “Unsure” option was included, to avoid many images being categorized as “Unsure.” Though individual participants may be unsure if a pattern is present, when viewed in aggregate, disagreement among multiple participants indicates ambiguous images that may need expert review.

2.4 Image retrieval and processing

Healthy Gulf initially identified 387 locations within the study area, using LDNR's Strategic Online Natural Resources Information System (SONRIS) tool, which allows the viewing of imagery at 3000-m in height. Bounding box coordinates were assigned to each location, and used to fetch the corresponding CIR aerial photograph from SONRIS by choosing the South LA Photography CIR layer. Each image was georeferenced based on the center of its bounding box.

2.5 Recruitment

Participants were recruited through citizen science platforms including SciStarter (popular citizen science project hub), Public Lab (DIY environmental science community), and Foldit (popular citizen science game for protein folding). The project was sent to SciStarter members via email and covered on SciStarter's blog, podcast, and social media. Additionally, the project circulated via Healthy Gulf's blog and social media and by a Public Lab webinar on October 5, 2020.

2.6 Getting started

Upon visiting the LLL main page, participants were presented with a brief project description, and an optional introductory video. Participants clicked “Get Started” to begin. A popup with the terms and conditions appeared to inform participants of the research study. Northeastern University's Institutional Review Board reviewed and approved the study as Exempt Category 2 (IRB# 16-05-17). After acknowledging the terms, participants were randomly assigned to one of the six subprojects.

After the tutorial, participants began classifying images. A progress bar informed participants of their progress and a help menu enabled them to review the tutorial. Participants could enlarge an image by clicking on it. Participants could classify any number of images, and leave by selecting “I'm Done” at any time. To protect participants' privacy, we did not gather participants' demographic, geographical, or other personal identifiable information. Participants were assigned a randomly generated unique identification number linked to their contributions. A new number was generated every time a participant started a new session. To analyze how long participants spent on certain projects and images, timestamps were stored with participant contributions.

2.7 Exit survey

Once participants finished classifying images, they were directed to an optional survey. The survey asked about participants' knowledge of wetland loss before and after the project, how important they feel that it is to measure effects of climate change and wetland loss, how responsible companies should be for assessing how their greenhouse gas pollution affects the land, whether they enjoyed the project, and how the project could be improved. Participants answered questions on a Likert scale, or in short responses in text boxes.

2.8 Results and contributions page

After the survey, participants were shown their and others' contributions for their assigned subproject on an interactive map. Participants could sign up for Healthy Gulf updates and navigate to an overall results page for results from all project categories. On the overall results page participants could click on any of the six study categories to map aggregate results. The study patterns were visualized in colored map points, which participants could click on to see the corresponding image and participants' percent majority confidence to date.

2.9 Expert comparison classifications

To validate the nonexpert participants' classifications, three experts from Healthy Gulf and five independent experts classified images. Experts used the same interface as the volunteers and classified the same images, however their classifications (labels) were stored separately. Experts were offered $250 in compensation and some are authors on this paper. We anonymously compared inter-expert agreement (described as expert confidence) defined as the percent of experts agreeing on the classification of an image and expert-participant agreement, defined as how frequently the majority classification of an image is the same for experts and participants (Table 2).

TABLE 2. LLL project descriptive statistics by project for both participants and experts
Project Farming Oil and gas Restoration Sea level rise Shipping Shoreline erosion Overall
Participants 104 97 120 134 108 109 672
Avg. number of labels 127.18 170.23 138.48 133.64 167.1 135.37 144.44
Avg. time per label (s) 3.76 2.98 3.42 3.22 3.6 4.28 3.52
Avg. no. of labels per image 38.11 35.77 32.59 42.18 40.93 29.94 36.59
Avg. no. of expert labels per image 5.78 7.12 6.57 5.5 6.36 6.99 6.39
Avg. participant confidence (%) 88% 81% 80% 86% 80% 79% 82%
Avg. expert confidence (%) 95% 87% 93% 93% 90% 85% 90%
Participant —expert agreement (%) 91% 83% 92% 94% 88% 84% 88%
Pattern identification (%) 19% 62% 12% 61% 14% 45% 36%


3.1 Participation and spamming

Throughout our study period (September 23, 2020 to May 13, 2021), 672 unique participants created 97,066 labels (or classifications), an average of 144.44 labels per participant of the 387 available per category. On average, participants spent 8.46 minutes labeling images.

Prior to analyzing the labels, an exploratory data analysis found no evidence of “spamming,” the act of clicking through the task without properly analyzing images. This analysis was conducted by looking at “high confidence images”: images with >90% of the participants voting for a certain label. A hypothetical spammer would have high disagreement with their fellow participants, even when looking at high confidence images. Our analysis showed that no participant performed poorly enough on high confidence images to warrant suspicion of spamming.

3.2 Analysis

For our analysis, we examined the following metrics (Table 2):
  • Average Number of Labels: The mean number of labels each participant contributed.
  • Average Time per Label: The mean time that it took participants to complete one label.
  • Average No. of Labels per Image: The average number of times images in the various categories received a participant label.
  • Average No. of Expert Labels per Image: The average number of times images received an expert label.
  • Average Participant Confidence: What percent of participants who agreed with the majority answer for a projects' set of images. Majority answer for an image is defined as the answer upon which at least 50% of participants agreed.
  • Average Expert Confidence: On average, what percent of experts agreed with the majority answer for a projects' set of images. Majority answer for an image is defined as the answer upon which at least 50% of participants agreed.
  • Participant - Expert Agreement: Did the participants and experts agree about whether there was evidence of a wetland morphology? For a given category and image, this was 0 or 1; this metric was averaged across each category.
  • Pattern Identification Percentage: The percentage of images for which the majority (>50%) of labels submitted indicated the presence of a wetland morphology (e.g. evidence of Farming etc.).

Participants labeled more images in some categories than others. For example, participants assigned to the Shoreline Erosion category labeled 127.18 images on average, while participants in Farming labeled 170.23 images on average. These two categories had the lowest and highest participation respectively.

3.3 Intra-Participant and Intra-Expert confidence

While overall agreement was high, there was still some expected disagreement among participants. To assess agreement, we analyzed average participant confidence, which is the fraction of participants whose labels agreed (e.g. if 90% of participants labeled an image as having shoreline erosion, participant confidence was 0.9). On average participants agreed with each other 82% of the time, though confidence varied between categories: Farming had the least intra-participant conflict (0.88), whereas Shoreline Erosion had the most (0.79) (Table 2).

At least five experts also examined each image. On average, each image received 6.38 expert labels. Experts had higher average confidence than participants (90% compared to 82%). Among experts, Farming and Shoreline Erosion were also the least and most controversial categories (0.95 & 0.85 expert confidence respectively).

3.4 Participant-Expert agreement

We compiled participant-expert agreement, a simple percentage of images for which the majority answer for participants was the same as the majority answer for experts (Table 2). Agreement was defined as more than 50% of the participants and more than 50% of experts voting that a pattern was (or was not) present in an image. Generally, participant-expert agreement was strong. Participants and experts agreed the most in identifying sea level rise (94%) and the least in identifying oil and gas (83%).

As some images were more difficult to categorize than others, we examined how agreement with experts varied based on participant confidence (Figure 1). Of the 2322 images available across all 6 projects, we found that participants agreed with experts 88.5% of the time (2054 images). Almost half of the time, participants agreed with experts (990 out of 2054 images) when participant confidence was extremely high, in the 90–100% range. However, there were still many images (161) where only a little over half of participants agreed on a label (50–60%). More controversial images, where only 50–60% of participants agreed on a certain label, had a high likelihood of disagreeing with the average expert opinion. However, images with greater than 70% participant confidence were overwhelmingly likely to agree with the majority of expert classification.

Details are in the caption following the image
Images, split up by aggregate participant confidence in the classification and also by whether the experts agreed with the participant label. Participant confidence is defined as the percent of participants who viewed an image and agreed that a wetland morphology was (or was not) visible

Table 2 and Figure 1 show that participants agreed with each other's classifications for the majority of images, participants agreed with experts at least 83% of the time in each project, and disagreement between experts and participants increased as disagreement among participants increased (Figure 1).

3.5 Wetland loss morphology patterns identified by participants

The ~100 participants in each category identified that some wetland morphologies were more prevalent than others. Restoration was the least frequently identified wetland morphology by both experts and participants. Experts identified restoration in only 8% of the images, while participants identified restoration in only 12%. Oil and Gas was the most frequent wetland morphology identified by participants (62% of images), and Sea Level Rise was most frequently identified by experts (58% of images) (Table 2 and Figure S2).

3.6 Geography of wetland morphologies identified in this study

The most frequently identified wetland morphologies identified in our study were Oil and Gas and Sea Level Rise. These two wetland morphologies frequently co-occurred in the same photograph. Of the 387 images, 266 were categorized as having either Oil and Gas or Sea Level Rise morphologies, and 205 were categorized as having both. Therefore, 69% of all images contained at least one of these two morphologies and 77% of the subset of images with either pattern contained both, however, as the discussion elaborates, this does not imply that relative sea level rise and waterlogging due to oil and gas canals are the primary processes of wetland loss in these locations (Figure 2).

Details are in the caption following the image
Map of most prominent categories identified show the frequent co-occurrence of “oil and gas” and “sea level rise,” with 69% of images containing at least one of these two categories and 77% of this subset of images with either pattern containing both

A map of classifications per wetland morphology category can be found in Figure 3. Participants identified clusters of Farming and Restoration wetland morphologies. Farming was mostly found in the areas surrounding the Mississippi River (Figure 3c), which generally corresponds with the “Fastlands” designated by LDNR (LDNR, 2020). Concentrated areas of Restoration were found to the south of Jefferson Parish, and around the Grand Isle region, which corresponds with completed restoration projects (LA CPRA, 2017).

Details are in the caption following the image
Maps of locations where the corresponding category was identified by participants. From left to right, top to bottom: (a) sea level rise (b) oil and gas (c) farming/agriculture (d) shipping (e) restoration (f) shoreline erosion

3.7 Survey results

Of the 672 unique participants that participated in the project, 45% answered at least one question on the optional feedback survey at the end of the project. While average response rates of online surveys vary, typically online survey response rates fall between 20% and 30% (How to Increase Survey Response Rates in 2021 // Qualtrics, n.d.). Participants answered the questions “How much did you know about land loss in the Gulf of Mexico before this project,” and “After participating in this project, how much do you know about land loss in the Gulf of Mexico?” on a Likert scale with 4 response options: “Nothing”, “Little”, “Some” and “A Lot.” The percentage of responses “Some” and “A Lot,” increased by 104% on average across all 6 categories after participants completed the project. Conversely, responses “Nothing” and “Little” decreased by 67% (Figure S3).

When asked if they found any of the images that they were trained to find, the majority of respondents (89.9%) answered “yes.” Only 10.1% answered “no.” In response to the statement “I feel this work was valuable,” nearly half (48.9%) of respondents strongly agreed, 46.8% somewhat agreed, 3.9% somewhat disagreed, and 0.3% strongly disagreed. In response to the statement “This project was easy for me to do,” 44% of respondents strongly agreed, 48.2% of respondents somewhat agreed, 7% of respondents somewhat disagreed, and 0.7% of respondents strongly disagreed. In response to the question “I enjoyed this project,” the overwhelming majority of respondents either somewhat agreed (48.2%) or strongly agreed (42.3%). Only 8.8% somewhat disagreed, and 0.7% strongly disagreed. Additionally, the majority of survey respondents (62.5%) felt that companies should be mostly responsible for assessing and measuring how their carbon pollution affects the land. Only 1% felt that companies should be slightly responsible, and 1% felt that companies should not be responsible for accounting for their pollution.


4.1 Participant accuracy and utility of crowdsourcing

When validated against expert classifications for accuracy, participant-expert agreement was very strong. Smaller subsets of controversial images, like those found in this study, could be passed onto experts for manual verification. Based on our finding that high participant confidence images reliably correlate with expert assessments, future work could expand this crowdsourcing project geographically. Additionally, crowdsourced assessments could provide training data for machine learning algorithms. Multiple algorithmic approaches to analyzing land cover have been developed, such as random forest, support vector machines, and recently, deep learning approaches. However, there are no algorithmic approaches to categorizing wetland loss according to wetland morphologies, and algorithmic approaches have limitations. Machine learning algorithms used for land cover categorization require large amounts of data and are only as good as their training data (Adadi, 2021; Vali et al., 2020). Images analyzed by citizen scientists can serve as additional training data to improve algorithm performance and can benefit the further processing and correcting of algorithms' outputs. While the most advanced land/water categorization algorithms incorporating LiDAR can reach over 95% accuracy, it is still helpful to have human experts examine categorizations for errors that can arise in ambiguous cases. Recent projects using automated analysis of land area change have relied on expert analysis to manually recode categorization errors (Couvillion et al., 2017). Studies examining change over time also rely on registration of images to accurately determine the correspondence between multiple images. Accurate automated registration techniques do exist, however it can be useful to augment them with analyses using manual image registration assistance, such as manual identification of control points (Oommen et al., 2012; Pairman et al., 2011). Crowdsourcing identification of wetland morphologies could expand the speed and frequency of wetland loss assessments and support agencies and experts.

4.2 Participant engagement

Participants were highly engaged and categorized many images; on average, participants labeled: 144.44 images out of the 387 that were available in a single session (37%) and on average 42 participants labeled each image. These findings are similar to another crowdsourced image classification project where each image circulated to, on average, 27 volunteers (Swanson et al., 2016).

After completing LLL, 56 participants signed up to receive project updates from Healthy Gulf. Participants' self-reported knowledge of wetland loss on the Gulf Coast increased after participating in LLL. These results suggest potential for LLL and similar crowdsourced projects as educational and engagement tools and that citizen science can be a novel approach to building nonprofit organizational membership. Future work will investigate whether LLL improved and increased existing members' engagement with Healthy Gulf and wetland loss conservation, and the sense of collaboration and community among participants.

4.3 Geography of identified wetland morphologies

Participants and experts identified wetland morphologies at expected frequencies in the expected locations. Restoration and Farming were expected to be least frequent in this data set as to-date restoration efforts have been limited and largely focused on the barrier islands (Louisiana's Comprehensive Master Plan for a Sustainable Coast, 2017). Additionally, industrial agriculture (described as Farming) is infrequent (<3% by acreage) in the marshes of the study area (Penland et al., 2000b). Farming was identified on Grand Isle (Figure 3c), this is likely because livestock ranches there have similar rectangular shapes to farmed fields.

Shipping was identified in expected locations. Participants identified the 4 major shipping channels in this region: Barataria Waterway, Intracoastal Waterway, the Empire to the Gulf and the Port Sulfur Canal. Gaps in the shipping channels are expected when identifying shipping channels in aerial images as, where land is eroded away, the channel can be hard to discern. Barataria Waterway, in particular, is hard to discern where land is eroded.

Shoreline erosion was expected to be a common morphology. Experts identified this morphology somewhat more frequently than participants, we hypothesize that experts are better trained at identifying erosion occurring on a smaller scale within the images. Erosion is expected to be geographically extensive, but limited to areas with larger bays, and thus, shorelines. Gaps in shoreline erosion are expected around large marsh areas without much shoreline, such as near Bay Sanbois, Ile de Barataria, near Three Bayou Bay, and Alombro Cemetery.

Oil and Gas wetland morphologies are also identified at expected locations and levels. As of April 2021, there were 7667 registered oil and gas wells within the study area, these well locations overlap well with photos identified in the Oil and Gas subproject (Morton et al., n.d., p. 36; LDNR, 2021, see Figure S4). It is possible, but unlikely, that some of the canals identified in the oil and gas canal assessment were made for fishing rather than oil and gas infrastructure. This is unlikely however due to canals' uniform widths and the “key holes” at the end of the canals where well heads are located. The locations where Oil and Gas morphologies were identified, but Sea Level Rise was not correspond with elevation. Photos with Oil and Gas but not Sea Level Rise were largely north of Lafitte and around Lake Salvador where elevation is higher. In areas of lower elevation, Oil and Gas wetland morphologies strongly co-occurred with Sea Level Rise. In the lowest lying regions of our study area, Sea Level Rise morphology was identified in many places that Oil and Gas was not.

4.4 Hydrology

This study is solely based on aerial photography and cannot distinguish the relative importance of different hydrological drivers of water level. Instead, our study area was chosen to avoid known ambiguities (Penland et al., 2000b), and to avoid areas where subsidence is likely the dominant driver of the water level rising relative to the marsh platform (Belhadjali, 2016; Freeman et al., 2021; Schleifstein, 2020). Additionally, though subsidence and reduced sediment deposition are likely also contributing to ponding observed in both the Oil and Gas and Sea Level Rise subprojects, current computer models and LA CPRA's Master Plan forecast that relative sea level will be the leading contributor to wetland loss in the near future particularly at lower elevations (Törnqvist et al., 2020; Visser & Duke-Sylvester, 2017). Rainfall can also contribute to waterlogging, however because rainfall is relatively uniform across the region (Perica et al., 2013), it is unlikely to affect waterlogging as significantly as impacts to drainage.

Both impounding by canals and relative sea level rise can create circular ponds examined in the subprojects Oil and Gas and Sea Level Rise via waterlogging. Impounding from oil and gas canals raises water levels in interior areas and produces a visually similar morphology to relative sea level rise. Relative sea level rise also drowns interior marshes (Swenson & Turner, 1987). Future studies could examine the distance of circular ponds from oil and gas canals. Waterlogging far away (>2 km, Swenson & Turner, 1987) from oil and gas canals could suggest that relative sea level rise, subsidence, reduced sediment deposition or rainfall are driving waterlogging in those areas. This, combined with systematic field work, could shed significant light on evaluating waterlogging in coastal Louisiana.

4.5 Nonprofit citizen science can differ from academic science

Cartoscope aims to build capacity for nonprofit citizen science projects because community and advocacy research questions and aims can differ from academic aims, particularly regarding environmental and justice issues like climate change (Hendricks et al., 2022; Tuck, 2009). Scientific uncertainty regarding causes of environmental issues has been strategically manipulated by industrial actors, particularly the oil and gas industries, with a vested interest in maintaining uncertainty around environmental and human health harms associated with economically valuable industries (Grasso, 2019; Oreskes & Conway, 2011). It can be difficult, if not impossible, to disentangle which wetland loss cause is primary: sea level rise, subsidence or impounding from oil and gas canals (among others). However, these three processes are synergistic and all associated with the production and burning of fossil fuels. Thereby, transitioning from fossil fuels could reduce wetland loss by reducing oil and gas extraction, GHG emissions and sea-level rise (Day, 2022).

The academic aims of identifying primary mechanisms of wetland loss differs from the advocacy and policy aims of preventing and reversing wetland loss as rapidly and equitably as possible (Tuck, 2009), particularly when the window to prevent worst case scenarios of climate warming above 2°C is rapidly closing (Allan et al., 2021). Specific to this study region and the issue of climate change, environmental managers for Venture Global Plaquemines LNG, a recently permitted terminal for the export of liquid natural gas in Plaquemines Parish, assert that climate change impacts of LNG on wetlands are “speculative” as there are “no scientifically accepted methods for evaluation of such impacts” even though it is expected to export 13.33 million metric tons of LNG per year (Venture Global, 2021, p10, Environmental Resource Management, 2019; DiSavino, 2022). Importantly, the proposed site for their facility was flooded by Hurricane Ida in 2021 (Yoder, 2021).

In response, it is important to build the public's capacity to research industrially related environmental and human health issues (Wylie, 2018). The LLL survey asks participants their perception of whether companies should be responsible for the impacts of their pollution, to evidence public support for corporate responsibility for GHG emissions. Evidence of public support for corporate accountability supports Healthy Gulf's nonprofit aims to advocate against building of new fossil fuel infrastructure in the region because while each individual facility's (i.e. Venture Global's Terminal) harms cannot yet be readily entirely modeled; the aggregate harm of continued dependence on fossil fuels to wetlands is already readily apparent (Allan et al., 2021).

4.6 Limitations

Shoreline erosion was consistently the most controversial category for experts (85% confidence) and participants (79% confidence) to classify. This is expected as shoreline erosion has no defined size, and is difficult to identify if the erosion is small compared to the size of the photo. It also involves accretion, the addition of eroded materials, particularly sand, further down-current. Accretion is not likely to be a common process here as it occurs along sandy shorelines and not amidst muddy wetlands. Wetlands account for the majority of land in this pilot area except for barrier islands. Further attention to accretion will be required if this approach is expanded to sandy shoreline areas and addition of time-series comparisons may improve assessment of erosion for both experts and participants.

Although we analyzed co-occurrence of the morphologies, we did not analyze which morphology is primary in a given image or otherwise determine the relative proportions of morphologies identified in an image. These analyses are not technically feasible in our present system. To make the analysis more fine grained we could divide each tile into smaller images, although this would require more volunteers and time.

4.7 Next steps

To assist with disentangling wetland morphologies from one another, next stages of this research will analyze datasets from previous years, and compare findings to prior studies. Building on the present findings, future work could focus more on the fossil fuel industry. We could overlay oil and gas well locations, along with ownership and production data to analyze wetland loss from oil and gas wells, canals, pipelines and other infrastructure. This method could also be used to analyze whether dry/closed wells and pipelines have been properly revegetated as is required by law.

In conclusion, we demonstrate that volunteers can be trained relatively easily online to identify characteristic wetland morphologies. We find a high level of agreement between volunteers with little or no training and experts with years of experience. Crowd identified patterns are consistent with expectations, although more work is needed to directly compare these patterns with previous studies. This provides a foundation for extending crowd-based wetland analysis to speed up land surveys and potentially train machine learning algorithms.


SE, SW, SC and SS conceived of the project; SS, EB, SW, SE and MR developed the project and led outreach, SS, KG and SC led data analysis. All authors contributed to the writing, editing and revision of the paper.


This work is supported by the National Science Foundation under grant no. 1816426. The authors thank the following experts that assisted the project: K. Castagno, R. Lloyd, J. Visser, E. Nost, S. Scyphers, Naomi,Y., S.E., and M.R. The authors thank C. Nickerson, SciStarter, and Public Lab for promoting the project.


    The authors declare no conflicts of interest.


    All the data for the project is accessible for download as a csv from the results tab of the Land Loss Lookout Project 2016: https://cartosco.pe/kioskProject.html#/resultsHub/landloss2016. The Cartoscope platform is open source and available on github: https://github.com/crowdgames/cartoscope.