This dataset was created for The Nature Conservancy (TNC) by the Desert Research Institute (DRI) under a contract entitled Climate Passages (BLM L22AC00234-00). The purpose of this effort was to generate a 500 meter raster grid that would serve as a basemap for species dispersal modeling. The grid represents the distribution of wetland areas whose properties were similar to sites that TNC expects to remain resilient under climate change pressures due to persistent groundwater sources. TNC identified 37 resilient terrestrial wetland areas corresponding to 127 grid cells. These were contrasted with 160 grid cells from other wetlands in the study area that demonstrated interannual drought sensitivity based on climate data and the normalized difference vegetation index (NDVI) from Landsat satellite imagery since 1985. The persistence of larger water bodies since 1985 was mapped using a threshold of Landsat near infrared reflectance and the modified normalized difference wetness index (MNDWI). A logistic regression based on the mean July NDVI from 2015 to 2024, the correlation between July NDVI and the Standardized Precipitation Evapotranspiration Index since 1985, and distance from persistent water bodies was able to back-classify the 287 sites with an overall accuracy of 88.8 %, producer’s accuracy of 85.8 %, and user’s accuracy of 89.3 %. Rasters are listed in the README.

Study Area and Raster Model Grid

The study area of the Climate Passages Project covers a large region that includes all of Nevada, and portions of California, Oregon, Idaho, Montana, Utah, and Arizona. An Albers equal area projection was selected to properly balance the distortions of representing such a large region with a two-dimensional raster grid. The projection parameters are provided in Table 1. A 500-meter spatial resolution was selected for this study based on computer memory limitations in the dispersal modeling software. The raster grid for the study area was 2,204 rows by 2,665 columns.

Table 1: ArcMap projection parameters for the modeling grid.

Wetland Area Delineations

The analysis leading to the estimation of resilience was performed only within areas that already have been mapped as wetlands, not to non-wetland locations. Source maps for delineating wetland areas within the overall study area were assessed on a state-by-state basis and were obtained from various online sources. The most common source of data was from the National Wetlands Inventory (NWI) that is produced and updated by the U.S. Fish and Wildlife Service. NWI wetland polygons in this study are the version that was available for download on March 28, 2024. The accuracy of NWI varies greatly, and NWI data in portions of the study area that date back to the 1980s are known to have serious positional and attribute errors. A state-by-state search of online map resources provided an improved wetland basemap compared to sole use of NWI, and data sources for each state are indicated in Table 2.

Table 2: Wetland map data sources for each state.

Several other data sources were considered, but none provided additional value for this effort. For example, online data for Oregon includes Local Wetland Inventories, Oregon’s Greatest Wetlands and streamlines associated with the Oregon Forest Practices Act, but data from such sources were outside the study area, duplicative, or point data instead of polygons. The two states where alternatives to NWI provided significant added value were California and Nevada. The California Aquatic Resources Inventory and the DRI Wetland Map of Nevada (Saito et al., 2020) are both composites of multiple wetland data sources with attention paid to selecting the best sources for different subregions.

Attribute codes from each data source were assessed to separate palustrine and riparian wetlands from lacustrine, and to remove features that had codes indicating artificial environments such as pipelines, ditches, and excavations. NWI polygons in the Intermountain West can cover areas that have very low vegetation density that are only ephemerally or intermittently wetted, and inclusion of these areas in the analysis would dilute the predictive ability of variables that are associated with resilient wetlands. Prior work has shown that terrestrial vegetation associated with wetlands rarely has a Landsat-based normalized difference vegetation index (NDVI) less than 0.2 (McGwire, 2019; Jeong et al., 2016; Jones et al., 2008), so portions of wetland polygons where the mean July NDVI from the last decade (2015–2024) was less than 0.2 were removed. Where the data existed, an additional map product developed by McGwire (manuscript in preparation) for parts of Nevada was used to limit riparian zones to the most hydrologically connected portion of the valley bottom.

Persistent Lacustrine Environments

A Landsat-based analysis of the persistence of inundation for water bodies was performed using historical late summer imagery from 1985 to 2023. Landsat data was screened using data quality and cloud cover flags, and median values from July 1 through September 30 were composited for each year. The modified normalized difference wetness index (MNDWI; Xu 2007) was calculated using green and mid-infrared reflectances. Using a geospatial tool called WetBar (McGwire, 2021) inundation was determined by an MNDVI > 0.2884. There are known problems with the minerology of some bright playa surfaces providing a false water signal with wetness indices, so an additional step used a near infrared reflectance greater than 0.1 to remove these false positives. The resulting binary inundation map for each year was summed to create a raster indicating the persistence of inundation. Two types of error that should be considered in using this product are that the detection of persistent inundation is dependent on the water body being large relative to the 30-meter spatial resolution of Landsat. A second error is that in the largest water bodies, persistent summer winds can create sun glint in some years which will reduce the count from a potential maximum of 39 years. This effect is prominent in the southeastern portion of Lake Tahoe.

Persistent Shoreline Environments

A non-wetland area in proximity to a persistent water body might provide a viable habitat or transit corridor for species dispersal modeling, so a binary shoreline raster was created from the calculation of inundation persistence. This raster was initialized using the edge where more than 35 years of inundation were calculated. The value of 35 years was used to reduce the aforementioned sun glint problem. Some manual editing was performed, and shoreline grid cells were created for two reservoirs with lower inundation counts because they were constructed after 1985, the South Fork Reservoir in Nevada and the Jordanelle Reservoir in Utah.

Palustrine and Riparian Environments

TNC provided the location of 37 sites that were to be considered resilient to climate change. Using those seed locations, 127 proximate 500m grid cells were selected as resilient based on photo interpretation of imagery available through ArcMap. For contrast, Dr. McGwire (Desert Research Institute) selected 160 grid cells in locations across the study area that had less-resilient wetland. Lower resiliency was inferred by low vegetation density or vigor based on photo interpretation and Landsat NDVI, or by sites with more vegetation but greater drought sensitivity based on a high correlation between the interannual variability in summertime NDVI and the 9-month Standardized Precipitation Evapotranspiration Index (SPEI; Vicente-Serrano et al., 2010).

The following variables were tested as predictors of resilient status:

• Mean July NDVI for 2015 – 2024
• Correlation between July NDVI detrended with a 2nd order polynomial and 9-month SPEI
• Distance to spring locations from the Springs Stewardship Institute (SSI)
• Distance from Landsat-derived persistent water bodies (>= 35 years)
• Minimum value of prior two distance variables (SSI & water bodies)
• SSI springs point density (ArcMap function)
• SSI springs kernel density (ArcMap function) default radius
• SSI springs kernel density (ArcMap function) 3500 m radius
• SSI springs kernel density (ArcMap function) 7000 m radius
• Topographic slope – average
• Topographic slope – minimum
• Topographic curvature – average
• Topographic curvature – minimum
• Topographic curvature – maximum

This analysis used logistic regression with L2 regularization to provide the probability of a grid cell containing resilient wetland. In fitting model coefficients, L2 regularization adds a penalty equal to the square of the magnitude of the coefficient to a loss function. This discourages the model from assigning excessively large weights to any feature, promoting simpler and more generalizable models that can handle multicollinearity and that are more robust when applied to new data. The regularized model is fitted using an optimization algorithm and has no P-value associated with the coefficients, so  variable selection relied on P-values from a simple logistic regression. The training grid cells for resilient wetlands were a clustered sampling around 37 seed locations, so they were not fully independent and P-values might overpredict statistical significance. While more predictors showed acceptable P-values, the analysis was limited to three predictor variables to maintain a more conservative approach. Table 3 shows the L2 regularized coefficients and simple logistic regression P-values for the three selected variables: the correlation between July NDVI and SPEI, the longterm mean of July NDVI, and the minimum distance to a spring or persistent water body. The overall back-classification accuracy of the logistic regression with L2 regularization was 88.8 %, with a producer’s accuracy of 85.8 % and user’s accuracy of 89.3 %. Figure 1 provides a box-and-whiskers plot of resilience probabilities. If a binary classification for resilient wetlands is desired, a probability >= 50 % could be used. However, a higher threshold might be used to provide a more conservative model output, or a range of values might be used for testing model sensitivity.

Table 3: Selected variables for the logistic regression.

Model Grids

The modeling grids derived from this analysis are:

• frequency of inundation (CPwetyears_V2.tif): number of years that summertime Landsat imagery indicated inundation
• persistent shoreline (CPshore_V2.tif): edge of waterbodies where CPwetyears_V2.tif >= 35 and some manual editing
• probability of resilient wetlands (CPprobability_V2.tif): probability estimates from the L2 regularized logistic regression
• area of mapped wetlands (CPareaNWI_V2.tif): the area within the grid cell covered by wetland polygons from the sources listed in Table 2
• area of resilient wetlands (CPareaResil_V2.tif): as CPareaNWI_V2.tif above, but area set to 0 if probability of resilient wetlands was < 50 %, and
• percent area of resilient wetlands (CPpctareaResil_V2.tif): as CPareaResil_V2.tif, but area expressed a percentage of grid cell.