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.

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.