A global map opens. In the Contents pane, there are four feature classes:

• Global ocean measurements—Ecological Marine Units point data that contains ocean measurements up to a 90-meter water depth.
• USA seagrass—Polygon data for seagrass occurrence. Every polygon in USA seagrass is an identified seagrass habitat.
• USA shallow waters—Shallow bathymetry polygon for the continental United States used as the study area for model training.
• Global shallow waters—Global shallow bathymetry polygon used to predict seagrass globally.

The data layers are in the Equal Earth projected coordinate system, which is appropriate for global analysis.

The light blue areas represent the shallow bathymetric zones around the world where the water depth may permit seagrass habitat.

The bright green areas are where seagrass habitat has been identified. You will use information about known seagrass presence locations around the continental United States to make predictions about where else around the world seagrass habitat might exist. Since this will be a prediction at a global scale, it will not be as appropriate for identifying seagrass habitats for smaller areas—for example, identifying the locations within a specific bay where seagrass is most likely to be present. Later, you will learn how to repurpose a model for other prediction scenarios.

These Global ocean measurements points show Ecological Marine Units (EMU) decadal average, 50-year mean, data values. Most of the data points lie outside of the seagrass observation layer. To develop a good prediction model using the Presence-only Prediction tool, you need many points in known seagrass areas with corresponding ocean measurement data. If you use only the subsample of EMU_Global_90m points within the seagrass polygon, you'll have too few observations.

To solve this problem, you will create a set of random points within the known seagrass habitat to train the model. You will also interpolate surfaces from the Global ocean measurements variables and use the random seagrass habitat points to sample the values of the interpolated measurements. The Global ocean measurement variables are: temp (temperature), salinity, dissO2 (dissolved Oxygen), nitrate, phosphate, silicate, and srtm30 (depth).

You will first dissolve the U.S. seagrass polygons into a single multipart feature and create a set of 5,000 random points within the areas of known seagrass presence.

When the tool completes, the new feature layer, USAseagrass_Dissolve, is added to the map and is listed in the Contents pane. You will be using this version of the seagrass layer, so it’s best to remove the original layer to keep your workspace clean and reduce confusion.

Now you will generate the random seagrass presence location points.

The random points appear on the map.

Now you have a new feature class with 5,000 points that fall within the known seagrass habitat around the United States coastline, which you will use in training your presence-only prediction model. At the moment, there are no environmental variables associated with these locations. That information is stored at the locations of the Global ocean measurements points. To address this, you'll create continuous interpolation surfaces for the environmental variables sampled at the Global ocean measurements points.

The Batch Empirical Bayesian Kriging page appears.

A list of the fields appears.

When you click Add, the fields are added to the tool pane.

This will create a raster with the name EBK_ plus the name of the field, for each of the fields.

These settings help increase the speed of the EBK prediction by limiting the number of points in each model and the number of simulated semivariograms. Increasing these values may enhance the precision of the predictions but will also increase the tool’s processing time. To better understand these parameters, see the What is empirical Bayesian kriging? help page.

Using a Standard Circular search neighborhood reduces the tool processing time. Limiting the minimum neighbor requirement ensures that values at unknown locations will be estimated even when there are only a few neighbors. Review the Empirical Bayesian Kriging tool documentation for more information about these and other parameters.

Because this tool will run in batch mode to generate seven separate global interpolation rasters, it will take some time to run (approximately five minutes).

The tool will complete with warnings indicating that NODATA values were ignored for several features. This is not a problem.

Once the Batch Empirical Bayesian Kriging tool completes, each ocean measurement surface is added to the map. They should look similar to the one below, which shows the EBK model for nitrate concentration.

A list of the raster layers appears.

When you click Add, the rasters are added to the tool pane.

These are all continuous measurement rasters, so the Categorical check boxes are not checked. The tool will also accept training variables that are categorical, for which you would need to check the box.

Basis functions transform (or expand) the explanatory variables to incorporate more complex relationships between seagrass presence and the variable of interest into the model. Selecting multiple basis functions includes all transformed versions of the variables into the model, from which the best performing variables are selected using regularization. In this case, you select all but the Discrete step option because Smoothed step and Discrete step are relatively similar and selecting only one will save processing time. Review the tool documentation for more information about each basis function.

Number of Knots is a setting related to the Smoothed step (Hinge) basis function that specifies the number of equal intervals between the minimum and maximum values of the variable, with both forward hinge and reverse hinge transformed variables being created. The Convex hull setting means that the study area will be designated as the convex hull of all input training points. The tool will generate background points, representing potential absence of seagrass, in areas of the study area that do not contain presence points.

These settings help minimize potential sample bias by removing presence and background points that are within the specified distance of each other so that areas are not spatially over-sampled. The distance between background points is impacted by the spatial resolution of the explanatory rasters, so using a 2-kilometer distance in this case will prevent over-sampling of background areas as compared to the seagrass presence areas. Using multiple iterations for thinning allows the tool to make multiple attempts at the thinning process and select the option that retains the most training points.

You will want to save a model file to share your analysis later, but only after you are sure that the model performs well.

The Relative Weight of Presence to Background value of 100 indicates that it is unknown whether seagrass might be present at the tool-generated background point locations.

It is appropriate to use C-log-log for the Presence Probability Transformation in this scenario because seagrass has minimal ambiguity in terms of location (that is, seagrass has no mobility or migration to take into account). The Presence Probability Cutoff of 0.5 indicates that locations with probabilities greater than 0.5 are classified as present.

This will be an output feature class containing the trained features (presence points and background points, in this case) used to generate the model.

The Output Response Curve Table and Output Sensitivity Table are helpful for understanding the performance of the model.

This will be the output raster that will show the model’s predictions of the likelihood of seagrass habitat presence.

Earlier, you designated the explanatory rasters for training the model on the coastal United States data points and here, you use the same rasters to make a global prediction. In some cases, you might want to make a prediction using different explanatory rasters. For example, you might use the same ocean measurement variables, but with projected values for 50 years in the future, to assess how climate change may impact seagrass habitat and range.

Because you are only using data from the coastal United States to train the model, you will need to allow predictions outside of the data ranges to make worldwide predictions.

Keep this in mind later when you look at the prediction results for places such as Antarctica, where conditions are very different from the United States coast.

These parameters instruct the tool to conduct K-fold cross-validation of the model.

The tool is almost ready to run. You'll add an Environments setting to restrict the area that is processed before running it.

Because seagrass grows in shallow waters, it will save time to restrict the processing to areas to shallow water.

The tool will take some time (about two minutes) to run.

The map shows areas of predicted seagrass habitat, symbolized with darker purple representing the areas with the highest probability of seagrass presence. The prediction may not be as accurate in certain areas, such as Antarctica, where the explanatory variables are outside the range that were used for training.

The gray and green points represent background training points created by the tool to gather data on locations where seagrass habitat may or may not exist.

There is a major problem with these data points. The vast majority are over land, which does not make sense for a model that is supposed to predict seagrass habitat. This is a conceptual problem with the model, which highlights the importance of having domain-specific knowledge and understanding each of the tool parameters to ensure the model is correctly specified.

Next, you will check the model diagnostics to see how the model performed.

The Details window provides important information about the model you have created and its performance. It also contains any warnings from the tool run. In this case, the warnings are not a problem for your analysis.

This table shows the model’s omission rate under the given presence probability cutoff (0.5, in this case) and AUC value. The AUC is the Area Under the ROC (Receiver Operating Characteristic) Curve, which measures the performance of the model by comparing the rates of true and false positives. Better model performance is indicated by lower omission rates and AUC values approaching 1.

The model’s AUC (close to 1) is very high, which is promising, but the omission rate (greater than 0.15) is also a bit high. You can also review other information in the Details window to better understand the model, including the regression coefficients and cross-validation summary.

The Cross-Validation Summary table shows that the % Presence - Correctly Classified ranged from 82 percent to 86 percent.

The final aspects of the model you will evaluate are the response curve and sensitivity tables.

The Partial Response of Continuous Variables chart visualizes the impact of changes in the value of each explanatory variable on presence probability, holding all other variables constant.

Clicking the smaller charts gives a better view of their variables in the larger chart on the right. The EBK_SALINITY chart shows that the probability of presence of seagrass habitat peaks sharply in a narrow range of salinity values.

These two charts give additional context to the omission rate and AUC diagnostics you looked at earlier.

The 0.5 cutoff value is the default value that you used in the model.

You can investigate how changing the presence probability cutoff would impact the classification of background points by clicking and dragging to select points on the Omission Rates chart.

Lowering the cutoff value increases the proportion of background points classified as potential presence.

You've reviewed the model results and examined some of the contextual diagnostic data. Now you'll adjust the model to deal with the conceptual problem of having training points on land.

Opening the tool this way opens it with all of the previous parameters still filled in.

You will only change a few of the tool parameters.

After tool validation completes, a new parameter will appear.

This will restrict the area of possible seagrass habitat presence and absence test locations to the shallow water coastal areas around the continental United States.

After a few moments, the path within your project folder structure will be populated and the .ssm file extension is added to the model name.

You will work with this model file in the next section of the tutorial.

• Output Trained Features: trainfeatures2
• Output Response Curve Table: rc2
• Output Sensitivity Table: sensitivity2

• Output Prediction Raster: seagrass_predict2

The tool will take some time (about two minutes) to run.

When the tool completes, layers will be added to the Contents pane.

The training features (presence and background locations) are located appropriately in nonland, coastal areas.

Check the Omission Rate and AUC values. Note that the AUC is similar to the previous model but the omission rate is much lower, indicating better model performance.

The Cross-Validation Summary table shows that the % Presence - Correctly Classified ranged from 95 percent to 96 percent.

You could also explore the Sensitivity and Response Curve charts for this new model and compare them to the previous model.

The Raster Layer contextual tab appears on the ribbon. This tab is available when a raster layer is selected in the Contents pane.

The Swipe tool interactively hides the selected layer and reveals the layer beneath it. You can use this tool to explore differences between your first and second predictions.

Note the differences around the Baltic Sea. With the initial model, predicted probability of seagrass habitat presence was very low in the Baltic Sea, especially around Copenhagen, Denmark, for example. The predicted probability increased in this region in the second model. Seagrass meadows are important carbon hot spots in the Baltic Sea, particularly so in certain protected bays around Denmark, so this helps increase confidence in the newer model's performance.

You would normally continue to explore the model predictions and compare them to other known seagrass locations outside of U.S. coastal waters, but for the purposes of this tutorial, you are ready to move on to sharing the model.

An information message appears, providing basic information about the model, including the model type and variable to predict.

Many details about the model are provided, including the date the model was created, the type of model, predictors and response, and model characteristics and diagnostics, including the AUC and Omission Rate.

Importantly, input locations and values are not disclosed in a model file, so you can share a model even if the input data is sensitive, such as nesting locations for an endangered bird species.

There is no information for the Description and Unit fields for the variable to predict and explanatory training rasters. Without understanding what each variable represents and its units, another user will not be able to make use of this model file. Imagine if a user assumed that temperature was measured in Fahrenheit for this model when it is actually in Celsius—their predictions would be incorrect.

Next, you will fill in this missing information.

The variable and raster names used in the model are listed. The Description and Unit boxes allow you to add information to the model documentation.

• For: EBK_DISSO2, Description: Dissolved oxygen, Unit: ml/l
• For: EBK_NITRATE, Description: Nitrates, Unit: μmol/l
• For: EBK_PHOSPHATE, Description: Phosphates, Unit: μmol/l
• For: EBK_SALINITY, Description: Salinity, Unit: None
• For: EBK_SILICATE, Description: Silicates, Unit: μmol/l
• For: EBK_SRTM30, Description: Depth, Unit: Meters
• For: EBK_TEMP, Description: Temperature, Unit: °C

This should trigger validation of the variables you entered. Sometimes these values are lost if tool validation is not triggered before running the tool.

The tool reports that the fields have been updated.

The details are updated.

You have confirmed that the variable descriptions and units are now correctly documented and the model file is ready to be shared by email, on a shared drive, or online. You can keep this model file to run a different prediction in the future or share it with others who may want to run additional predictions. For example, this prediction used Ecological Marine Units (EMU) decadal average (50-year mean) data, but another researcher may want to predict using projected ocean measurements to understand how seagrass distributions might change under warming ocean conditions.