The Wetland Intrinsic Potential tool: Mapping wetland intrinsic potential through machine learning of multi-scale remote sensing proxies of wetland indicators

Halabisky, Meghan; Miller, Dan; Stewart, Anthony J.; Lorigan, Daniel; Brasel, Tate; Moskal, L. Monika

doi:https://doi.org/10.5194/egusphere-2022-665

Preprints

https://doi.org/10.5194/egusphere-2022-665

Preprints

11 Oct 2022

| 11 Oct 2022

The Wetland Intrinsic Potential tool: Mapping wetland intrinsic potential through machine learning of multi-scale remote sensing proxies of wetland indicators

Meghan Halabisky, Dan Miller, Anthony J. Stewart, Daniel Lorigan, Tate Brasel, and L. Monika Moskal

Abstract. Accurate, un-biased wetland inventories are critical to monitor and protect wetlands from future harm or land conversion. However, most wetland inventories are constructed through manual image interpretation or automated classification of multi-band imagery and are biased towards wetlands that are easy to detect directly in aerial and satellite imagery. Wetlands that are obscured by forest canopy, occur ephemerally, and those without visible standing water are, therefore, often missing from wetland maps. To aid in detection of these cryptic wetlands, we developed the Wetland Intrinsic Potential tool, based on a wetland indicator framework commonly used on the ground to detect wetlands through the presence of hydrophytic vegetation, hydrology, and hydric soils. Our tool uses a random forest model with spatially explicit input variables that represent all three wetland indicators, including novel multi-scale topographic indicators that represent the processes that drive wetland formation, to derive a map of wetland probability. With the ability to include multi-scale topographic indicators, the WIP tool can identify areas conducive to wetland formation and provides a flexible approach that can be adapted to diverse landscapes. For a study area in the Hoh River Basin in Western Washington, USA, classification of the output probability with a threshold of 0.5 provided an overall accuracy of 91.97 %. Compared to the National Wetland Inventory, the classified WIP-tool output increased areas classified as wetland by 160 % and reduced errors of omission from 47.5 % to 14.1 %, but increased errors of commission from 1.9 % to 10.5 %. The WIP tool is implemented using a combination of R and python scripts in ArcGIS.

Received: 28 Jul 2022 – Discussion started: 11 Oct 2022

Download & links

Meghan Halabisky et al.

Status: closed

RC1:
'Comment on egusphere-2022-665', Anonymous Referee #1, 12 Feb 2023
The manuscript presents a novel method called Wetland Intrinsic Potential, which based on a set of indicators selected by the authors aims to identify wetlands that are often missing from inventory due to presence of forest canopy, due to their ephemerality or because they have no visible standing water. The authors applied their method in a specific study area (Hoh River Basin), showing that they were able to increase the classification of wetland overall area with lower omission errors but higher commission errors.

As the authors mentioned in their introduction, wetland mapping is clearly one of the most interesting challenges for the wide array of functions of these ephemeral water bodies at landscape scale. For this reason, I commend the authors for their contribution in this research direction, however I am a little bit skeptical about the flexibility of their approach (that they mentioned both in the abstract and in the conclusion), and therefore the extension to other landscapes. I have listed below few suggestions and/or questions that might help the authors in improving their work.

The authors did a good job in the introduction listing several of the ongoing challenges and comparing them with accurate research. However, I suggest the authors to look at more recent manuscript that implement machine learning, random forest, or satellite images to extensively map wetlands both at higher and lower spatial resolution such as:

Lane, C. R., D'Amico, E., Christensen, J. R., Golden, H. E., Wu, Q., and Rajib, A.: Mapping Global Non-Floodplain Wetlands, Earth Syst. Sci. Data Discuss. [preprint], https://doi.org/10.5194/essd-2023-3, in review, 2023.

Xiang et al., 2023. GWL_FCS30: a global 30 m wetland map with a fine classification system using multi-sourced and time-series remote sensing imagery in 2020

Wu et al., 2019. Integrating LiDAR data and multi-temporal aerial imagery to map wetland inundation dynamics using Google Earth Engine

Mullen, C., Bertassello, L. E., Rao, P. S. C., & Müller, M. F. (2022). From wetlands to wetlandscapes: Remote sensing calibration of process-based hydrological models in heterogeneous landscapes. Hydrological Processes, 36( 11), e14739. https://doi.org/10.1002/hyp.14739

In particular, it would be nice to compare the estimate of your model with other sources since the limitation of the NWI are quite well known. I believe that the recent data from Lane et al. 2023 are freely available and one of the datasets they use for validation is close to the region used by the authors. At least a comment on that would be appreciated.

Extension to other case studies. I feel this is a crucial point when it comes to wetland classification and preparation of new inventories as the authors stated at line 140. However, I am not sure how this can be done in places where training data (section 3.2) are not available. I would be curious to see – if possible – the application of the model in a different case study, or at least in the same case study without starting from a preliminary model and see how it compares with the current estimates.

A more technical question for the authors is about the threshold 0.5 they used for their binary classification. Can you comment a little bit more why they chose 0.5 and not a different number? Something that would be useful is a sensitivity analysis on the threshold value to strengthen the sentence at the end of section 5.1 “If users want to lower…selected”. I feel the sentence is a little bit weak and would need more quantitative estimates.

Minor comments:

Section 2.1. Can you add more specific of the case study? Such as the size of the watershed, the number of wetlands identified by the NWI?

Figure 1. Can you add the location of the case study also in the map in the inset just above the legend? Either a dot or the boundary of the watershed would be nice to have a sense on where to locate it for readers that are not familiar with the region.

Table 1. I am not sure I understand what some of the numbers in the table are. Is 85 the number of identified wetlands? The percentage? Please improve the caption of the table so a reader can understand what is going on.
Citation: https://doi.org/10.5194/egusphere-2022-665-RC1
- AC1: 'Reply on RC1', Meghan Halabisky, 13 Apr 2023
  
  RC1 -1 re: “skeptical about the flexibility of their approach”: We appreciate the suggestions and/or questions. We agree that many researchers falsely claim that a method is flexible when it is not and we certainly do not want to overstate the flexibility. To clarify, we believe the WIP tool, not our specific model for the Hoh developed here, is flexible as it can be adapted for new and different geographies. The flexibility of the tool is that the approach is based on a wetland indicator framework realizing that the predictor variables in the random forest model will change based on the wetland types in a selected area of interest. Here, we tested out many predictor variables used in the literature as well as several new multiscale terrain indices. Framing our model around a wetland indicator framework (and building a tool with this in mind) helped us ground the approach in wetland ecology and integrate domain knowledge (wetland ecology) into our remote sensing solution, which we feel could be helpful for others. We set out to build an adaptable approach and tool, instead of a static map or model, which was flexible from the onset and one that can be improved upon over time realizing that users would continue to find new input datasets.
  For this study, we tested the WIP tool out in one study area that is considered especially difficult to map. However, the WIP tool has been applied to several new and distinct geographies since. For example, as part of the NASA Develop program the WIP tool was adapted to map wetlands in the Big Island of Hawaii, where geology was a key predictor due to the influence of the volcano (Than et al, 2022). The hope is that as the tool is applied to more areas, new variables can be tested out and identified and can be added as input variables for future applications. The tool in ArcGIS allows for additional input variables to be added. While global and national level inventories will always be important, in many areas like the Washington State local jurisdictions want iterative approaches that can be updated and improved over time. This tool was an attempt to move away from simply producing static maps, but providing tools that can be used and improved through time. We have edited the text to make these points more clear.
  RC1-2: We have included the new citations to the manuscript. Thank you. I have compared the Lane et al. 2023, Xiang et al 2023, NWI and the WIP tool (our project) results in this comment. The WIP outperforms all of these datasets substantially in our study area. However, it is perhaps an unfair comparison as the goals of a global dataset are much different than our high resolution watershed specific approach. They are different products and visually the differences are very clear. For these reasons we chose not to include a comparison in the main text of the manuscript. The Lane et al. 2023 dataset aims to improve global wetland maps and uses a much coarser pixel resolution. It does not map wetlands very well across our study area (fig. 1b). As a comment - the Lane et al. 2023 appears to misclassify a large area in the Olympic mountains with few wetlands as a large wetland and seems to pick up riparian area – but not clearly delineating riparian wetlands. Again, this is likely due to the coarse resolution. The Xiang et al 2023 dataset (fig. 1a) misses the majority of the wetlands in the Hoh watershed. The NWI (fig 1c) does an adequate job of mapping wetlands, but misses many of the more difficult to identify wetlands under canopy. All of these datasets seem to illustrate our point that without multiscale terrain metrics it is difficult to identify forested wetlands.
  RC1-3 re: “Extension to other case studies.” - I agree this approach may not be suitable for areas with no training data. Training data is an essential component of machine learning. However, it is possible to create training data for any area with or without an existing classification. Training data can be created by a simple random sample or stratified random sample, perhaps using a dataset like slope index and then labelling these points as wetland or upland using interpretation of high resolution aerial imagery and topography. In the example in this paper we used an existing wetland dataset (NWI) as training data for a preliminary model for more efficient stratified sampling and labelling. A global wetland dataset, such as the ones mentioned above, could be used to create a preliminary model like the one we developed using the NWI. The purpose of the preliminary model is to reduce the amount of labelling needed to create a robust training dataset. Once a preliminary model is created the user then samples along the model probability, which indicates areas where the model is less certain if a pixel is a wetland or not. This allowed us to collect fewer data samples for training our random forest model – reducing overall effort. Several tools exist to make labelling data more efficient such as Collect Earth Online, Google Earth, and ArcGIS. I have added some text to explain how this can be applied to areas with no wetland inventories. This is being tested out in the Digital Earth Africa platform.
  The model has been tested in several watershed across Washington State, British Columbia, Canada, Alaska, Hawaii, and is currently being tested in several watersheds across Africa. However, all of these applications were for management applications and none of these models have yet been published, except the report from the NASA develop team. We have added some text to provide some qualitative information on usability in other areas and cited Than et al, 2022 report for the Big Island of Hawaii. Because a primary focus of this research was to improve errors of omissions, especially forested wetlands, we decided instead of reporting broadly on all of these projects to focus on an intensive validation of our most challenging study area, a densely forested watershed in an old-growth temperate rainforest in the Pacific Northwest where we were able to spend additional time in the field.
  RC1-4 re: “selection of a 0.5 threshold”
  The goal of this project was not to create a binary classification. We only selected a threshold to create a binary classification to validate the model as a continuous probability estimate cannot be validated. The threshold of 0.5 was used because the model results for that pixel (location on the ground) predicted that it was more likely to be a wetland than not. Because we wanted to test model accuracy we wanted to adhere to using what the model predicted to have a higher likelihood of being a wetland than not. Our goal was not to create a binary classification as we believe probability classification has utility as a standalone product. However, if users want to create a binary classification they can select a threshold to reduce errors of omission or commission. In some cases they may want to minimize false positives and select a higher probability. In other instances, users may be interested in identifying moist forest that does not necessarily meet the criteria of a wetland. We can add an additional output from the WIP tool to figure 4 that shows how overall accuracy changes with different cutoffs (Supplement Fig. 2). This cutoff will look different in other watersheds, but is one of the outputs from the WIP tool that may provide guidance for those wanting to create a binary classification.
  RC1-5 Minor comments:
  “Section 2.1. Can you add more specific of the case study? Such as the size of the watershed, the number of wetlands identified by the NWI?”
  Thanks for this suggestion. We have added more description to the text. In addition, we have created an ArcGIS online map for readers (and reviewers) to explore the study area, training and validation data, and the pre-existing National Wetland Inventory and our model results. It can be found here. https://uw.maps.arcgis.com/apps/mapviewer/index.html?webmap=46889ad0fda44662a95efe1559d3f32c
  “Figure 1. Can you add the location of the case study also in the map in the inset just above the legend? Either a dot or the boundary of the watershed would be nice to have a sense on where to locate it for readers that are not familiar with the region.”
  Sorry about that. Somehow the location in the inset map was left off. We have corrected the inset map to include the study area location.
  “Table 1. I am not sure I understand what some of the numbers in the table are. Is 85 the number of identified wetlands? The percentage? Please improve the caption of the table so a reader can understand what is going on.”
  We have improved the caption. 85 is the number of identified wetlands for the validation.
  
  Citation: https://doi.org/10.5194/egusphere-2022-665-AC1
RC2:
'Comment on egusphere-2022-665', Anonymous Referee #2, 27 Feb 2023
The authors developed a wetland mapping tool for identifying wetlands through machine learning algorithms (e.g., random forest) applied to remote sensing datasets. Three wetland indicators are considered in the proposed framework, including hydrophytic vegetation, hydrology, and hydric soils. NAIP imagery, Lidar data, and SSURGO data correspond to each wetland indicator, respectively. The proposed framework was applied to mapping potential wetlands in the Hoh River Basin in Washington, USA. The results show that the proposed framework reduced the omission error, but increased the commission error when compared to the National Wetlands Inventory (NWI).
Mapping wetlands is a challenging task in remote sensing. This wetland mapping tool developed by the authors can potentially be useful to the wetland mapping community. See below my comments for the authors’ consideration:
There is a Wetland Identification Model (WIM) that has been available through Arc Hydro since 2020. The WIM methodology is similar to the proposed method in this manuscript, except that WIM only considers DEM data. This manuscript also considers vegetation and soil data. However, these two wetland indicators have also been widely studied in the literature. What’s new in the proposed method compared to what has been available in the literature? The O'Neil et. al (2018) has been cited in this manuscript, but the O'Neil et. al 2019 and 2020 papers on the WIM models are not. Why not build upon WIM rather than starting from scratch?

Can the authors make the resulting data products (overlaid on NWI layers) available to the public? Maybe through ArcGIS Online and an Earth Engine App so that readers can visually compare the authors’ wetland mappings to NWI. Although the commission and omission errors seem reasonable, I am more interested in how the resulting products align with NWI at a fine scale. I am always a bit skeptical about new wetland products unless I can visualize them on an interactive map and compare them with well-known wetland products such as the NWI.

The proposed wetland tool produced an increased wetland area by 160% compared to NWI. Why? This needs an in-depth discussion. As a reader, I am interested in knowing when the tool works best, and when it fails.

What is the minimum mapping unit used in this study? Did the authors do any post-processing to reduce the salt-and-pepper effect of the resulting wetland maps? How would that affect the omission and commission error calculations?

In terms of the accuracy assessments, did the authors perform both pixel-based and object-based accuracy assessments?

The data used in this study are mostly available at the national scale. For example, NAIP and SSURGO data are available at the national scale, and LiDAR data are also available for the majority of the US through the USGS 3DEP program. The training data are derived from NWI, which is also available nationally. I would hope that the proposed tool can be applied to other areas. However, the authors stated in Section 5.2 that their intention was not to develop a model that could be extended to new areas without the collection of new training data. This greatly reduces the transferability of the method and usability of the tool.

Lastly, here are two recently published papers on multi-scale geomorphometric analysis that might be of interest to the authors.

References:
Wetland Identification Model: https://community.esri.com/t5/water-resources-blog/the-wetland-identification-model-wim-a-new-arc/ba-p/884298

O'Neil, G. L., Goodall, J. L., Behl, M., Saby, L. (2020). Deep Learning using Physically-Informed Input Data for Wetland Identification. Environmental Modelling and Software. 104665. https://doi.org/10.1016/j.envsoft.2020.104665.

O'Neil, G. L., Saby, L., Band, L. E., Goodall, J. L. (2019). Effects of LiDAR DEM Smoothing and Conditioning Techniques on a Topography-Based Wetland Identification Model. Water Resources Research, 55. https://doi.org/10.1029/2019WR024784.

O'Neil, G. L., Goodall, J. L., Watson, L. T. (2018). Evaluating the potential for site-specific modification of LiDAR DEM derivatives to improve environmental planning-scale wetland identification using random forest classification. Journal of Hydrology, 559, 192-208.

https://doi.org/10.1016/j.jhydrol.2018.02.009.

Lindsay, J. B., Newman, D. R., & Francioni, A. (2019). Scale-optimized surface roughness for topographic analysis. Geosciences, 9(7), 322. https://doi.org/10.3390/geosciences9070322

Newman, D. R., Cockburn, J. M., Drǎguţ, L., & Lindsay, J. B. (2022). Evaluating Scaling Frameworks for Multiscale Geomorphometric Analysis. Geomatics, 2(1), 36-51. https://doi.org/10.3390/geomatics2010003
Citation: https://doi.org/10.5194/egusphere-2022-665-RC2
- AC2: 'Reply on RC2', Meghan Halabisky, 14 Apr 2023
  
  RC2-1: “There is a Wetland Identification Model (WIM) that has been available through Arc Hydro since 2020. The WIM methodology is similar to the proposed method in this manuscript, except that WIM only considers DEM data. This manuscript also considers vegetation and soil data. However, these two wetland indicators have also been widely studied in the literature. What’s new in the proposed method compared to what has been available in the literature? The O'Neil et. al (2018) has been cited in this manuscript, but the O'Neil et. al 2019 and 2020 papers on the WIM models are not. Why not build upon WIM rather than starting from scratch? “
  Building upon the WIM tool is a fantastic suggestion. Development of the WIP model was started before the WIM paper came out and delayed because of COVID. However, since this manuscript was submitted for review (July 2023) we have been put in contact with O’Neil et al (2018, 2019, 2020) and are currently working to integrate the components of the WIP into the WIM tool. This will greatly improve the sustainability of the WIP toolbox within ArcGIS as updates are made by ESRI. We have applied the WIM tool ‘as is’ to the Hoh watershed and were not able to produce adequate results likely because the curvature metrics used for the WIM do not adequately capture the complex multi-scale terrain associated with wetlands. The most novel part of our method and the finding that has helped us to finally be able to map wetlands in the PNW is our inclusion of multi-scale terrain indices that help identify wetlands of multiple shapes and sizes. We have since shared all our data with ESRI and have met several times to identify ways we can integrate the multi-scale indices into the WIM. In addition we are working with O’Neil to allow for points within the WIM and not just polygons as training data. The updates do far to the WIM have been published on this blog by ESRI https://community.esri.com/t5/water-resources-blog/wim-updates-for-arcgis-pro-3/ba-p/1233973. We will add this update to the manuscript and a link to this blog describing the WIM. We will add the additional O’Neil citations from 2019 and 2020.
  To reiterate, the big breakthrough for our research was the inclusion of the multi-scale terrain indices (plan curvature, profile curvature, deviation from local elevation, gradient) as a complement to the other wetland indicator variables. To our knowledge we don’t believe that any GIS software can produce these multi-scale terrain indices at this time in the way we have, including the WIM. The inclusion of these multi-scale terrain indices are important complements to other existing datasets like TWI, imagery, soils, etc.. In our watershed these common input datasets did not have as much model importance as our multi-scale terrain indices and may explain why many automated approaches in the PNW fall short. We will emphasize the novelty of our approach more clearly and how it builds on other research more clearly in the discussion.
  Despite our enthusiasm at integrating the WIP into the WIM, we still see value in a stand-alone open source tool for those without access to ESRI products. We are currently working with Digital Earth Africa to develop an open source python based tool to map wetland intrinsic potential using the Open Data Cube and have plans to release an R package as soon as time permits.
  
  RC2-2: “Can the authors make the resulting data products (overlaid on NWI layers) available to the public? Maybe through ArcGIS Online and an Earth Engine App so that readers can visually compare the authors’ wetland mappings to NWI. Although the commission and omission errors seem reasonable, I am more interested in how the resulting products align with NWI at a fine scale. I am always a bit skeptical about new wetland products unless I can visualize them on an interactive map and compare them with well-known wetland products such as the NWI.”
  Absolutely. Great idea. All of the training and validation data, and the NWI for comparison are included, as well as the model outputs have been included in the ArcGIS online map for easy review by readers and reviewers. They were also included as supplemental data with this manuscript. We have added this to the text and also mentioned that the datasets are free and open for others to use for model development.
  The ArcOnline map is located here: https://uw.maps.arcgis.com/apps/mapviewer/index.html?webmap=46889ad0fda44662a95efe1559d3f32c
  RC2-3: “The proposed wetland tool produced an increased wetland area by 160% compared to NWI. Why? This needs an in-depth discussion. As a reader, I am interested in knowing when the tool works best, and when it fails.“
  I agree with your suggestion above that providing a visual may be the best way to assess the accuracy. The comparison with the NWI is very different and easy to see in the ArcGIS map online and worth an in-depth comparison. The reason the WIP is able to identify substantially more wetlands is likely due to the inclusion of our novel multi-scale terrain indices that complement the additional input variables. We tested other methods for many years without success, until the discovery and development of these multi-scale terrain indices. The multi-scale terrain indices are helpful because most of the Hoh wetlands are under forest canopy and difficult to identify in imagery alone. Hydrologic indices, such as TWI, which are used in other wetland mapping approaches, are helpful in identifying areas where surface water flows accumulate. The Depth-to-Water is also helpful at identifying wetlands that may have groundwater inputs. However, in the Hoh watershed there are several small swales, depressions, and hummocky areas that are more difficult to identify using hydrologic indices. Additionally, many of the wetlands in the Hoh are precipitation driven peat forming wetlands, and for these the TWI & Depth-to-Water are not particularly helpful. Our novel multi-scale terrain indices are useful at detecting wetlands that are under canopy and occur in nested features of different shapes and sizes (depression, gulch, valley). We tried to set this up in our introduction, but will try and tie the results more directly in the discussion to emphasize the important addition of these multi-scale terrain indices in our WIP model and explain the model results. Of note, the math is incorrect and the increase is not 160% but rather ~125% - about 2.25 times more wetlands than the NWI. This aligns with other qualitative observations that in the PNW about 50% of the forested wetlands are missing from the NWI.
  
  RC2 – 4: “What is the minimum mapping unit used in this study? Did the authors do any post-processing to reduce the salt-and-pepper effect of the resulting wetland maps? How would that affect the omission and commission error calculations?”
  The output of the tool is a probability raster at 4m pixel resolution. The minimum mapping unit, therefore, is the pixel resolution (4 meters). We only selected a threshold to create a binary classification to determine the accuracy of the model in predicting wetlands. However, we feel the usefulness in the tool is the probability raster itself. Depending on the application users can decide to implement further steps, such as OBIA, salt-and-pepper post-processing, manual photo delineation or Cowardin classification, use in the field for sampling, forestry management, or as a model input to predict additional wetland characteristics such as carbon stocks. We did not do any salt-and-pepper post-processing as our primary goal was to assess model accuracy as a probability and any clean-up would distort assessment of the model accuracy. As a note, this landscape has few depressional wetlands and many complex, hummocky wetlands that occur along a flat terrain. I would imagine if we did implement some post-processing, errors of omission would go up as it would erroneously remove a proportion of these wetlands. Again, this was not the goal of the binary classification so we did not implement any post-processing that would obscure the accuracy assessment of the WIP tool itself
  RC2-5: “In terms of the accuracy assessments, did the authors perform both pixel-based and object-based accuracy assessments?”
  We did not perform two accuracy assessments. As mentioned above, we did not use an object based approach. The results are a continuous raster of wetland probability. We selected a 0.5 cutoff to test model accuracy. If an object based classification is desired, segmentation may improve results for areas where wetlands have more distinct boundaries. Nevertheless, this tool was not meant to smooth out results, but to be a way to identify all potential wetland area, so we did not implement any post-processing steps such as object-based segmentation or salt and pepper removal. We can add some text to describe how others may improve binary classification if that is their goal through post-processing such as OBIA or filtering.
  RC2-6: “The data used in this study are mostly available at the national scale. For example, NAIP and SSURGO data are available at the national scale, and LiDAR data are also available for the majority of the US through the USGS 3DEP program. The training data are derived from NWI, which is also available nationally. I would hope that the proposed tool can be applied to other areas. However, the authors stated in Section 5.2 that their intention was not to develop a model that could be extended to new areas without the collection of new training data. This greatly reduces the transferability of the method and usability of the tool.”
  Correct. We did not develop a “model” that could be extended to new areas, but rather a method and a tool that could be extended to new areas. Our model was trained on wetland types that occur in the Hoh watershed and therefore the model is inappropriate for other areas, unless of course the watershed is similar to the Hoh (e.g., same ecoregion) or it was used as a preliminary model to help with sampling for further refinement. The downside of machine learning models is that they require a lot of training data. A rule-based approach may be more suitable for areas that have no wetland data at all. However, having created several rule based approaches, it can be difficult identify thresholds without any training data. These thresholds likely are interdependent on other variables and climate making it difficult to implement in complex areas like the Hoh. We will add text and citations for these rule based approaches. However, as mentioned in the response to reviewer #1 a preliminary model can be used to create training data and it is not too difficult to use a preliminary model to develop an efficient sampling scheme to create labelled training points. Several platforms exist making it easier to label data. Our approach is interesting in that we have far fewer training data points than most random forest models and yet we are still able to produce sufficient results. I believe this is because we sampled across the variability of wetlands in an efficient way through the use of a preliminary model classification. A preliminary classification is not needed, sample training points could be created using a random sample or a stratified random sample using a simple layer such as slope index. We will add more text describing methods for areas with low or no training data.
  RC2 -7: “Lastly, here are two recently published papers on multi-scale geomorphometric analysis that might be of interest to the authors.”
  Thank for you these references. We have added them to the manuscript. It is exciting to see the development of these multi-scale geomorphometric indices. Our research has shown these multi-scale indices to be critically important for identifying wetlands in complex forested study areas like the Hoh watershed with variable sized wetland features. We will add some text to the description of how gradient and curvature are calculated:
  "Gradient and curvature were calculated using the methodology described by Zevenbergen and Thorne, 200 (1987) in which the shape of the ground surface at a DEM grid point is interpolated as a smooth polynomial surface that matches elevations of the
  
  grid point and its eight adjacent points. This methodology was modified to use a circular neighbourhood (Shi et al., 2007) of arbitrary radius, with elevations along the circle interpolated from adjacent DEM grid points. This procedure allows estimates
  
  of gradient and curvature for each DEM point measured over any length scale, down to the DEM grid size. This is similar to the "local quadratic regression" described by Newman et al. (2022), but uses a slightly higher-order polynomial with an exact fit to only 9 points, elevation at the current DEM grid point and elevations at 8 equally spaced points on the circumference of a circle of specified radius. This effectively smooths the DEM over the diameter of the circle with no increase in processing time with increasing spatial scale, i.e., with larger circle diameters."
  We did not evaluate surface roughness and therefore, did not cite this paper. It is not clear to us how surface roughness (or texture) would be related to topographic controls on groundwater flow, and we were seeking to characterize those topographic controls. Potentially, this could be something worth exploring.
  
  Citation: https://doi.org/10.5194/egusphere-2022-665-AC2

Status: closed

RC1:
'Comment on egusphere-2022-665', Anonymous Referee #1, 12 Feb 2023
The manuscript presents a novel method called Wetland Intrinsic Potential, which based on a set of indicators selected by the authors aims to identify wetlands that are often missing from inventory due to presence of forest canopy, due to their ephemerality or because they have no visible standing water. The authors applied their method in a specific study area (Hoh River Basin), showing that they were able to increase the classification of wetland overall area with lower omission errors but higher commission errors.

As the authors mentioned in their introduction, wetland mapping is clearly one of the most interesting challenges for the wide array of functions of these ephemeral water bodies at landscape scale. For this reason, I commend the authors for their contribution in this research direction, however I am a little bit skeptical about the flexibility of their approach (that they mentioned both in the abstract and in the conclusion), and therefore the extension to other landscapes. I have listed below few suggestions and/or questions that might help the authors in improving their work.

The authors did a good job in the introduction listing several of the ongoing challenges and comparing them with accurate research. However, I suggest the authors to look at more recent manuscript that implement machine learning, random forest, or satellite images to extensively map wetlands both at higher and lower spatial resolution such as:

Lane, C. R., D'Amico, E., Christensen, J. R., Golden, H. E., Wu, Q., and Rajib, A.: Mapping Global Non-Floodplain Wetlands, Earth Syst. Sci. Data Discuss. [preprint], https://doi.org/10.5194/essd-2023-3, in review, 2023.

Xiang et al., 2023. GWL_FCS30: a global 30 m wetland map with a fine classification system using multi-sourced and time-series remote sensing imagery in 2020

Wu et al., 2019. Integrating LiDAR data and multi-temporal aerial imagery to map wetland inundation dynamics using Google Earth Engine

Mullen, C., Bertassello, L. E., Rao, P. S. C., & Müller, M. F. (2022). From wetlands to wetlandscapes: Remote sensing calibration of process-based hydrological models in heterogeneous landscapes. Hydrological Processes, 36( 11), e14739. https://doi.org/10.1002/hyp.14739

In particular, it would be nice to compare the estimate of your model with other sources since the limitation of the NWI are quite well known. I believe that the recent data from Lane et al. 2023 are freely available and one of the datasets they use for validation is close to the region used by the authors. At least a comment on that would be appreciated.

Extension to other case studies. I feel this is a crucial point when it comes to wetland classification and preparation of new inventories as the authors stated at line 140. However, I am not sure how this can be done in places where training data (section 3.2) are not available. I would be curious to see – if possible – the application of the model in a different case study, or at least in the same case study without starting from a preliminary model and see how it compares with the current estimates.

A more technical question for the authors is about the threshold 0.5 they used for their binary classification. Can you comment a little bit more why they chose 0.5 and not a different number? Something that would be useful is a sensitivity analysis on the threshold value to strengthen the sentence at the end of section 5.1 “If users want to lower…selected”. I feel the sentence is a little bit weak and would need more quantitative estimates.

Minor comments:

Section 2.1. Can you add more specific of the case study? Such as the size of the watershed, the number of wetlands identified by the NWI?

Figure 1. Can you add the location of the case study also in the map in the inset just above the legend? Either a dot or the boundary of the watershed would be nice to have a sense on where to locate it for readers that are not familiar with the region.

Table 1. I am not sure I understand what some of the numbers in the table are. Is 85 the number of identified wetlands? The percentage? Please improve the caption of the table so a reader can understand what is going on.
Citation: https://doi.org/10.5194/egusphere-2022-665-RC1
- AC1: 'Reply on RC1', Meghan Halabisky, 13 Apr 2023
  
  RC1 -1 re: “skeptical about the flexibility of their approach”: We appreciate the suggestions and/or questions. We agree that many researchers falsely claim that a method is flexible when it is not and we certainly do not want to overstate the flexibility. To clarify, we believe the WIP tool, not our specific model for the Hoh developed here, is flexible as it can be adapted for new and different geographies. The flexibility of the tool is that the approach is based on a wetland indicator framework realizing that the predictor variables in the random forest model will change based on the wetland types in a selected area of interest. Here, we tested out many predictor variables used in the literature as well as several new multiscale terrain indices. Framing our model around a wetland indicator framework (and building a tool with this in mind) helped us ground the approach in wetland ecology and integrate domain knowledge (wetland ecology) into our remote sensing solution, which we feel could be helpful for others. We set out to build an adaptable approach and tool, instead of a static map or model, which was flexible from the onset and one that can be improved upon over time realizing that users would continue to find new input datasets.
  For this study, we tested the WIP tool out in one study area that is considered especially difficult to map. However, the WIP tool has been applied to several new and distinct geographies since. For example, as part of the NASA Develop program the WIP tool was adapted to map wetlands in the Big Island of Hawaii, where geology was a key predictor due to the influence of the volcano (Than et al, 2022). The hope is that as the tool is applied to more areas, new variables can be tested out and identified and can be added as input variables for future applications. The tool in ArcGIS allows for additional input variables to be added. While global and national level inventories will always be important, in many areas like the Washington State local jurisdictions want iterative approaches that can be updated and improved over time. This tool was an attempt to move away from simply producing static maps, but providing tools that can be used and improved through time. We have edited the text to make these points more clear.
  RC1-2: We have included the new citations to the manuscript. Thank you. I have compared the Lane et al. 2023, Xiang et al 2023, NWI and the WIP tool (our project) results in this comment. The WIP outperforms all of these datasets substantially in our study area. However, it is perhaps an unfair comparison as the goals of a global dataset are much different than our high resolution watershed specific approach. They are different products and visually the differences are very clear. For these reasons we chose not to include a comparison in the main text of the manuscript. The Lane et al. 2023 dataset aims to improve global wetland maps and uses a much coarser pixel resolution. It does not map wetlands very well across our study area (fig. 1b). As a comment - the Lane et al. 2023 appears to misclassify a large area in the Olympic mountains with few wetlands as a large wetland and seems to pick up riparian area – but not clearly delineating riparian wetlands. Again, this is likely due to the coarse resolution. The Xiang et al 2023 dataset (fig. 1a) misses the majority of the wetlands in the Hoh watershed. The NWI (fig 1c) does an adequate job of mapping wetlands, but misses many of the more difficult to identify wetlands under canopy. All of these datasets seem to illustrate our point that without multiscale terrain metrics it is difficult to identify forested wetlands.
  RC1-3 re: “Extension to other case studies.” - I agree this approach may not be suitable for areas with no training data. Training data is an essential component of machine learning. However, it is possible to create training data for any area with or without an existing classification. Training data can be created by a simple random sample or stratified random sample, perhaps using a dataset like slope index and then labelling these points as wetland or upland using interpretation of high resolution aerial imagery and topography. In the example in this paper we used an existing wetland dataset (NWI) as training data for a preliminary model for more efficient stratified sampling and labelling. A global wetland dataset, such as the ones mentioned above, could be used to create a preliminary model like the one we developed using the NWI. The purpose of the preliminary model is to reduce the amount of labelling needed to create a robust training dataset. Once a preliminary model is created the user then samples along the model probability, which indicates areas where the model is less certain if a pixel is a wetland or not. This allowed us to collect fewer data samples for training our random forest model – reducing overall effort. Several tools exist to make labelling data more efficient such as Collect Earth Online, Google Earth, and ArcGIS. I have added some text to explain how this can be applied to areas with no wetland inventories. This is being tested out in the Digital Earth Africa platform.
  The model has been tested in several watershed across Washington State, British Columbia, Canada, Alaska, Hawaii, and is currently being tested in several watersheds across Africa. However, all of these applications were for management applications and none of these models have yet been published, except the report from the NASA develop team. We have added some text to provide some qualitative information on usability in other areas and cited Than et al, 2022 report for the Big Island of Hawaii. Because a primary focus of this research was to improve errors of omissions, especially forested wetlands, we decided instead of reporting broadly on all of these projects to focus on an intensive validation of our most challenging study area, a densely forested watershed in an old-growth temperate rainforest in the Pacific Northwest where we were able to spend additional time in the field.
  RC1-4 re: “selection of a 0.5 threshold”
  The goal of this project was not to create a binary classification. We only selected a threshold to create a binary classification to validate the model as a continuous probability estimate cannot be validated. The threshold of 0.5 was used because the model results for that pixel (location on the ground) predicted that it was more likely to be a wetland than not. Because we wanted to test model accuracy we wanted to adhere to using what the model predicted to have a higher likelihood of being a wetland than not. Our goal was not to create a binary classification as we believe probability classification has utility as a standalone product. However, if users want to create a binary classification they can select a threshold to reduce errors of omission or commission. In some cases they may want to minimize false positives and select a higher probability. In other instances, users may be interested in identifying moist forest that does not necessarily meet the criteria of a wetland. We can add an additional output from the WIP tool to figure 4 that shows how overall accuracy changes with different cutoffs (Supplement Fig. 2). This cutoff will look different in other watersheds, but is one of the outputs from the WIP tool that may provide guidance for those wanting to create a binary classification.
  RC1-5 Minor comments:
  “Section 2.1. Can you add more specific of the case study? Such as the size of the watershed, the number of wetlands identified by the NWI?”
  Thanks for this suggestion. We have added more description to the text. In addition, we have created an ArcGIS online map for readers (and reviewers) to explore the study area, training and validation data, and the pre-existing National Wetland Inventory and our model results. It can be found here. https://uw.maps.arcgis.com/apps/mapviewer/index.html?webmap=46889ad0fda44662a95efe1559d3f32c
  “Figure 1. Can you add the location of the case study also in the map in the inset just above the legend? Either a dot or the boundary of the watershed would be nice to have a sense on where to locate it for readers that are not familiar with the region.”
  Sorry about that. Somehow the location in the inset map was left off. We have corrected the inset map to include the study area location.
  “Table 1. I am not sure I understand what some of the numbers in the table are. Is 85 the number of identified wetlands? The percentage? Please improve the caption of the table so a reader can understand what is going on.”
  We have improved the caption. 85 is the number of identified wetlands for the validation.
  
  Citation: https://doi.org/10.5194/egusphere-2022-665-AC1
RC2:
'Comment on egusphere-2022-665', Anonymous Referee #2, 27 Feb 2023
The authors developed a wetland mapping tool for identifying wetlands through machine learning algorithms (e.g., random forest) applied to remote sensing datasets. Three wetland indicators are considered in the proposed framework, including hydrophytic vegetation, hydrology, and hydric soils. NAIP imagery, Lidar data, and SSURGO data correspond to each wetland indicator, respectively. The proposed framework was applied to mapping potential wetlands in the Hoh River Basin in Washington, USA. The results show that the proposed framework reduced the omission error, but increased the commission error when compared to the National Wetlands Inventory (NWI).
Mapping wetlands is a challenging task in remote sensing. This wetland mapping tool developed by the authors can potentially be useful to the wetland mapping community. See below my comments for the authors’ consideration:
There is a Wetland Identification Model (WIM) that has been available through Arc Hydro since 2020. The WIM methodology is similar to the proposed method in this manuscript, except that WIM only considers DEM data. This manuscript also considers vegetation and soil data. However, these two wetland indicators have also been widely studied in the literature. What’s new in the proposed method compared to what has been available in the literature? The O'Neil et. al (2018) has been cited in this manuscript, but the O'Neil et. al 2019 and 2020 papers on the WIM models are not. Why not build upon WIM rather than starting from scratch?

Can the authors make the resulting data products (overlaid on NWI layers) available to the public? Maybe through ArcGIS Online and an Earth Engine App so that readers can visually compare the authors’ wetland mappings to NWI. Although the commission and omission errors seem reasonable, I am more interested in how the resulting products align with NWI at a fine scale. I am always a bit skeptical about new wetland products unless I can visualize them on an interactive map and compare them with well-known wetland products such as the NWI.

The proposed wetland tool produced an increased wetland area by 160% compared to NWI. Why? This needs an in-depth discussion. As a reader, I am interested in knowing when the tool works best, and when it fails.

What is the minimum mapping unit used in this study? Did the authors do any post-processing to reduce the salt-and-pepper effect of the resulting wetland maps? How would that affect the omission and commission error calculations?

In terms of the accuracy assessments, did the authors perform both pixel-based and object-based accuracy assessments?

The data used in this study are mostly available at the national scale. For example, NAIP and SSURGO data are available at the national scale, and LiDAR data are also available for the majority of the US through the USGS 3DEP program. The training data are derived from NWI, which is also available nationally. I would hope that the proposed tool can be applied to other areas. However, the authors stated in Section 5.2 that their intention was not to develop a model that could be extended to new areas without the collection of new training data. This greatly reduces the transferability of the method and usability of the tool.

Lastly, here are two recently published papers on multi-scale geomorphometric analysis that might be of interest to the authors.

References:
Wetland Identification Model: https://community.esri.com/t5/water-resources-blog/the-wetland-identification-model-wim-a-new-arc/ba-p/884298

O'Neil, G. L., Goodall, J. L., Behl, M., Saby, L. (2020). Deep Learning using Physically-Informed Input Data for Wetland Identification. Environmental Modelling and Software. 104665. https://doi.org/10.1016/j.envsoft.2020.104665.

O'Neil, G. L., Saby, L., Band, L. E., Goodall, J. L. (2019). Effects of LiDAR DEM Smoothing and Conditioning Techniques on a Topography-Based Wetland Identification Model. Water Resources Research, 55. https://doi.org/10.1029/2019WR024784.

O'Neil, G. L., Goodall, J. L., Watson, L. T. (2018). Evaluating the potential for site-specific modification of LiDAR DEM derivatives to improve environmental planning-scale wetland identification using random forest classification. Journal of Hydrology, 559, 192-208.

https://doi.org/10.1016/j.jhydrol.2018.02.009.

Lindsay, J. B., Newman, D. R., & Francioni, A. (2019). Scale-optimized surface roughness for topographic analysis. Geosciences, 9(7), 322. https://doi.org/10.3390/geosciences9070322

Newman, D. R., Cockburn, J. M., Drǎguţ, L., & Lindsay, J. B. (2022). Evaluating Scaling Frameworks for Multiscale Geomorphometric Analysis. Geomatics, 2(1), 36-51. https://doi.org/10.3390/geomatics2010003
Citation: https://doi.org/10.5194/egusphere-2022-665-RC2
- AC2: 'Reply on RC2', Meghan Halabisky, 14 Apr 2023
  
  RC2-1: “There is a Wetland Identification Model (WIM) that has been available through Arc Hydro since 2020. The WIM methodology is similar to the proposed method in this manuscript, except that WIM only considers DEM data. This manuscript also considers vegetation and soil data. However, these two wetland indicators have also been widely studied in the literature. What’s new in the proposed method compared to what has been available in the literature? The O'Neil et. al (2018) has been cited in this manuscript, but the O'Neil et. al 2019 and 2020 papers on the WIM models are not. Why not build upon WIM rather than starting from scratch? “
  Building upon the WIM tool is a fantastic suggestion. Development of the WIP model was started before the WIM paper came out and delayed because of COVID. However, since this manuscript was submitted for review (July 2023) we have been put in contact with O’Neil et al (2018, 2019, 2020) and are currently working to integrate the components of the WIP into the WIM tool. This will greatly improve the sustainability of the WIP toolbox within ArcGIS as updates are made by ESRI. We have applied the WIM tool ‘as is’ to the Hoh watershed and were not able to produce adequate results likely because the curvature metrics used for the WIM do not adequately capture the complex multi-scale terrain associated with wetlands. The most novel part of our method and the finding that has helped us to finally be able to map wetlands in the PNW is our inclusion of multi-scale terrain indices that help identify wetlands of multiple shapes and sizes. We have since shared all our data with ESRI and have met several times to identify ways we can integrate the multi-scale indices into the WIM. In addition we are working with O’Neil to allow for points within the WIM and not just polygons as training data. The updates do far to the WIM have been published on this blog by ESRI https://community.esri.com/t5/water-resources-blog/wim-updates-for-arcgis-pro-3/ba-p/1233973. We will add this update to the manuscript and a link to this blog describing the WIM. We will add the additional O’Neil citations from 2019 and 2020.
  To reiterate, the big breakthrough for our research was the inclusion of the multi-scale terrain indices (plan curvature, profile curvature, deviation from local elevation, gradient) as a complement to the other wetland indicator variables. To our knowledge we don’t believe that any GIS software can produce these multi-scale terrain indices at this time in the way we have, including the WIM. The inclusion of these multi-scale terrain indices are important complements to other existing datasets like TWI, imagery, soils, etc.. In our watershed these common input datasets did not have as much model importance as our multi-scale terrain indices and may explain why many automated approaches in the PNW fall short. We will emphasize the novelty of our approach more clearly and how it builds on other research more clearly in the discussion.
  Despite our enthusiasm at integrating the WIP into the WIM, we still see value in a stand-alone open source tool for those without access to ESRI products. We are currently working with Digital Earth Africa to develop an open source python based tool to map wetland intrinsic potential using the Open Data Cube and have plans to release an R package as soon as time permits.
  
  RC2-2: “Can the authors make the resulting data products (overlaid on NWI layers) available to the public? Maybe through ArcGIS Online and an Earth Engine App so that readers can visually compare the authors’ wetland mappings to NWI. Although the commission and omission errors seem reasonable, I am more interested in how the resulting products align with NWI at a fine scale. I am always a bit skeptical about new wetland products unless I can visualize them on an interactive map and compare them with well-known wetland products such as the NWI.”
  Absolutely. Great idea. All of the training and validation data, and the NWI for comparison are included, as well as the model outputs have been included in the ArcGIS online map for easy review by readers and reviewers. They were also included as supplemental data with this manuscript. We have added this to the text and also mentioned that the datasets are free and open for others to use for model development.
  The ArcOnline map is located here: https://uw.maps.arcgis.com/apps/mapviewer/index.html?webmap=46889ad0fda44662a95efe1559d3f32c
  RC2-3: “The proposed wetland tool produced an increased wetland area by 160% compared to NWI. Why? This needs an in-depth discussion. As a reader, I am interested in knowing when the tool works best, and when it fails.“
  I agree with your suggestion above that providing a visual may be the best way to assess the accuracy. The comparison with the NWI is very different and easy to see in the ArcGIS map online and worth an in-depth comparison. The reason the WIP is able to identify substantially more wetlands is likely due to the inclusion of our novel multi-scale terrain indices that complement the additional input variables. We tested other methods for many years without success, until the discovery and development of these multi-scale terrain indices. The multi-scale terrain indices are helpful because most of the Hoh wetlands are under forest canopy and difficult to identify in imagery alone. Hydrologic indices, such as TWI, which are used in other wetland mapping approaches, are helpful in identifying areas where surface water flows accumulate. The Depth-to-Water is also helpful at identifying wetlands that may have groundwater inputs. However, in the Hoh watershed there are several small swales, depressions, and hummocky areas that are more difficult to identify using hydrologic indices. Additionally, many of the wetlands in the Hoh are precipitation driven peat forming wetlands, and for these the TWI & Depth-to-Water are not particularly helpful. Our novel multi-scale terrain indices are useful at detecting wetlands that are under canopy and occur in nested features of different shapes and sizes (depression, gulch, valley). We tried to set this up in our introduction, but will try and tie the results more directly in the discussion to emphasize the important addition of these multi-scale terrain indices in our WIP model and explain the model results. Of note, the math is incorrect and the increase is not 160% but rather ~125% - about 2.25 times more wetlands than the NWI. This aligns with other qualitative observations that in the PNW about 50% of the forested wetlands are missing from the NWI.
  
  RC2 – 4: “What is the minimum mapping unit used in this study? Did the authors do any post-processing to reduce the salt-and-pepper effect of the resulting wetland maps? How would that affect the omission and commission error calculations?”
  The output of the tool is a probability raster at 4m pixel resolution. The minimum mapping unit, therefore, is the pixel resolution (4 meters). We only selected a threshold to create a binary classification to determine the accuracy of the model in predicting wetlands. However, we feel the usefulness in the tool is the probability raster itself. Depending on the application users can decide to implement further steps, such as OBIA, salt-and-pepper post-processing, manual photo delineation or Cowardin classification, use in the field for sampling, forestry management, or as a model input to predict additional wetland characteristics such as carbon stocks. We did not do any salt-and-pepper post-processing as our primary goal was to assess model accuracy as a probability and any clean-up would distort assessment of the model accuracy. As a note, this landscape has few depressional wetlands and many complex, hummocky wetlands that occur along a flat terrain. I would imagine if we did implement some post-processing, errors of omission would go up as it would erroneously remove a proportion of these wetlands. Again, this was not the goal of the binary classification so we did not implement any post-processing that would obscure the accuracy assessment of the WIP tool itself
  RC2-5: “In terms of the accuracy assessments, did the authors perform both pixel-based and object-based accuracy assessments?”
  We did not perform two accuracy assessments. As mentioned above, we did not use an object based approach. The results are a continuous raster of wetland probability. We selected a 0.5 cutoff to test model accuracy. If an object based classification is desired, segmentation may improve results for areas where wetlands have more distinct boundaries. Nevertheless, this tool was not meant to smooth out results, but to be a way to identify all potential wetland area, so we did not implement any post-processing steps such as object-based segmentation or salt and pepper removal. We can add some text to describe how others may improve binary classification if that is their goal through post-processing such as OBIA or filtering.
  RC2-6: “The data used in this study are mostly available at the national scale. For example, NAIP and SSURGO data are available at the national scale, and LiDAR data are also available for the majority of the US through the USGS 3DEP program. The training data are derived from NWI, which is also available nationally. I would hope that the proposed tool can be applied to other areas. However, the authors stated in Section 5.2 that their intention was not to develop a model that could be extended to new areas without the collection of new training data. This greatly reduces the transferability of the method and usability of the tool.”
  Correct. We did not develop a “model” that could be extended to new areas, but rather a method and a tool that could be extended to new areas. Our model was trained on wetland types that occur in the Hoh watershed and therefore the model is inappropriate for other areas, unless of course the watershed is similar to the Hoh (e.g., same ecoregion) or it was used as a preliminary model to help with sampling for further refinement. The downside of machine learning models is that they require a lot of training data. A rule-based approach may be more suitable for areas that have no wetland data at all. However, having created several rule based approaches, it can be difficult identify thresholds without any training data. These thresholds likely are interdependent on other variables and climate making it difficult to implement in complex areas like the Hoh. We will add text and citations for these rule based approaches. However, as mentioned in the response to reviewer #1 a preliminary model can be used to create training data and it is not too difficult to use a preliminary model to develop an efficient sampling scheme to create labelled training points. Several platforms exist making it easier to label data. Our approach is interesting in that we have far fewer training data points than most random forest models and yet we are still able to produce sufficient results. I believe this is because we sampled across the variability of wetlands in an efficient way through the use of a preliminary model classification. A preliminary classification is not needed, sample training points could be created using a random sample or a stratified random sample using a simple layer such as slope index. We will add more text describing methods for areas with low or no training data.
  RC2 -7: “Lastly, here are two recently published papers on multi-scale geomorphometric analysis that might be of interest to the authors.”
  Thank for you these references. We have added them to the manuscript. It is exciting to see the development of these multi-scale geomorphometric indices. Our research has shown these multi-scale indices to be critically important for identifying wetlands in complex forested study areas like the Hoh watershed with variable sized wetland features. We will add some text to the description of how gradient and curvature are calculated:
  "Gradient and curvature were calculated using the methodology described by Zevenbergen and Thorne, 200 (1987) in which the shape of the ground surface at a DEM grid point is interpolated as a smooth polynomial surface that matches elevations of the
  
  grid point and its eight adjacent points. This methodology was modified to use a circular neighbourhood (Shi et al., 2007) of arbitrary radius, with elevations along the circle interpolated from adjacent DEM grid points. This procedure allows estimates
  
  of gradient and curvature for each DEM point measured over any length scale, down to the DEM grid size. This is similar to the "local quadratic regression" described by Newman et al. (2022), but uses a slightly higher-order polynomial with an exact fit to only 9 points, elevation at the current DEM grid point and elevations at 8 equally spaced points on the circumference of a circle of specified radius. This effectively smooths the DEM over the diameter of the circle with no increase in processing time with increasing spatial scale, i.e., with larger circle diameters."
  We did not evaluate surface roughness and therefore, did not cite this paper. It is not clear to us how surface roughness (or texture) would be related to topographic controls on groundwater flow, and we were seeking to characterize those topographic controls. Potentially, this could be something worth exploring.
  
  Citation: https://doi.org/10.5194/egusphere-2022-665-AC2

Meghan Halabisky et al.

Data sets

WIP training and validation data and input datasets Meghan Halabisky, Dan Miller, Anthony Stewart https://terrainworks.sharefile.com/d-s88888b342d3f49d09b067683a73a916e

Model code and software

WIP tool code Dan Miller , Meghan Halabisky , Anthony J. Stewart , Daniel Lorigan , Tate Brasel https://github.com/TerrainWorks-Seattle/ForestedWetlands

Meghan Halabisky et al.

Viewed

Total article views: 1,372 (including HTML, PDF, and XML)

HTML	PDF	XML	Total	BibTeX	EndNote
742	605	25	1,372	33	31

HTML: 742
PDF: 605
XML: 25
Total: 1,372
BibTeX: 33
EndNote: 31

Views and downloads (calculated since 11 Oct 2022)

Month	HTML	PDF	XML	Total
Oct 2022	99	35	5	139
Nov 2022	41	34	1	76
Dec 2022	38	27	1	66
Jan 2023	28	52	0	80
Feb 2023	105	101	3	209
Mar 2023	40	71	2	113
Apr 2023	48	45	4	97
May 2023	56	59	2	117
Jun 2023	53	42	1	96
Jul 2023	68	41	3	112
Aug 2023	79	60	1	140
Sep 2023	87	38	2	127

Cumulative views and downloads (calculated since 11 Oct 2022)

Month	HTML	PDF	XML	Total
Oct 2022	99	35	5	139
Nov 2022	41	34	1	76
Dec 2022	38	27	1	66
Jan 2023	28	52	0	80
Feb 2023	105	101	3	209
Mar 2023	40	71	2	113
Apr 2023	48	45	4	97
May 2023	56	59	2	117
Jun 2023	53	42	1	96
Jul 2023	68	41	3	112
Aug 2023	79	60	1	140
Sep 2023	87	38	2	127

Viewed (geographical distribution)

Total article views: 1,364 (including HTML, PDF, and XML) Thereof 1,364 with geography defined and 0 with unknown origin.

Country	#	Views	%

Latest update: 01 Oct 2023

Short summary

Accurate wetland inventories are critical to monitor and protect wetlands. However, in many areas a large proportion of wetlands are unmapped because they are hard to detect in aerial and satellite imagery. We developed a machine learning approach using spatially mapped variables of wetland indicators (i.e., vegetation, hydrology, soils) to predict wetland probability across a landscape. Our tool provides a flexible approach that can be adapted to diverse landscapes to improve wetland detection.


Total:	0
HTML:	0
PDF:	0
XML:	0