Detection of invasive plant species is vital for the management of natural resources and ecosystem protection. In this study, we demonstrate the utility of remotely sensed data, in the newly developed Software for Assisted Habitat Modeling, and predicting invasive species occurrence on the landscape. Hi, I'm Tom Stohlgren, a Senior Scientist at the Natural Resource Ecology Lab, at Colorado State University.
This is tamarisk. It's a native of South Africa, Asia, Europe, and parts of the Middle East, but it's not from around here. That is, it's not a native species to the Southwestern portion of the United States, or near LaJuna, Colorado, where we are today.
Tamarisk is unusual, in that it's a facultative phreatophyte, which means, it can live far away from a creek, or right along a creek. It has a taproot that goes down, upwards of 30 meters. So imagine taking 30 steps from here, and that's how far this taproot can go down.
So it can withstand drought conditions a little bit better than some of the native species in the area. Without the tamarisk, there would be plenty of more room for native species, that attract native butterflies and pollinators, as we see here, on the rabbitbrush. Here we have the tamarisk shrubs, directly competing with native vegetation.
In this case, a native Cottonwood. Behind me, we have a very dense stand, a very large stand. Maybe a square kilometer of nothing but tamarisk.
By dropping leaves, and with dead and dying branches, this is a real fire hazard. On the other side of the street, we have Salix, we have a Willow, which is very green and moist, and much less prone to fire than the tamarisk on this side. We're here at a site where tamarisk is invading on range land.
Now, this is important because it actually reduces the area of range land that those cows can graze on. Tamarisk, like many invasive plants, exhibits phenological variation throughout the growing season, that differs from native Riparian species phenology. In some areas, for example, tamarisk leaf out, is before some native Riparian plants, as tamarisk retains its foliage longer than other native species.
By using a time series of satellite data throughout the growing season, we can use these phenological differences to help distinguish tamarisk from native plants. Landsat satellites have been orbiting the earth since 1972, and are the ideal image source for detecting tamarisk distribution, and phenology on the landscape. With a spatial resolution of 30 meters, and a temporal resulting of 16 days, Landsat is a joint program of NASA and USGS.
Our objectives in this study were to test and evaluate, five different species distribution models, in the Software for Assisted Habitat Modeling, using Landsat 5 imagery and tamarisk presence points, acquired from an intensive field mapping campaign conducted by the Tamarisk Coalition, along the Arkansas River, in Colorado. And, to create an accurate map of tamarisk distribution in the study area, based on model outputs. This conceptual diagram provides an overview of our methodology in this study.
Field data for tamarisk were derived from a vector polygon data set, collected by the Tamarisk Coalition, in 2005 and 2006. Landsat 5 Thematic Mapper data were acquired from Earth Explorer, for the years corresponding to the tamarisk field data. At least one scene from each month of the growing season was collected.
Using the Remote Sensing Indices Derivation Tool, we derived Spectral Indices from the Landsat imagery, to distinguish the spectral signature of tamarisk from other species on the landscape. These Indices, and the tamarisk field data, were inputs, in five Species Distribution Models, within the Software for Assisted Habitat Modeling. Model outputs were tested with an independent data set, and an ensemble approach was employed to create Species Distribution Maps of tamarisk in the study area.
To map a large stand of tamarisk, I'll have a starting location, nearest the shrub on the boundary, and then I'll keep picking my locations around the entire patch, like so. This was the methodology, employed by the Tamrarisk Coalition to collect xy locations of tamaraisk. Using GPS, select Mark, then scroll up to name the point.
You will see the latitude and longitude of the point on the screen. Select Done, when finished. We downloaded Landsat 5 Thematic Mapper scenes, from earthexplorer.usgs.gov.
If you already have an account, log in. Otherwise, register for an account to download data. First, type in the Path and Row of the Landsat scene, encompassing the study area.
Our study used Path 32, Row 34. Select the Date Range corresponding to the scenes you are searching for. We selected April 2005, through November 2006, when the tamarisk data were collected.
Then, select Data Sets, and scroll down to Landsat Archive. Select the Landsat Surface Reflectance Product. Scroll down, and select Additional Criteria.
Select Less than 10%Cloud Cover, to ensure the best quality images. When you select Results, a list of available Landsat scenes will appear for download. To derive Indices from the Landsat remotely sensed imagery, we downloaded the Remote Sensing Indices Derivation Tool from github.com.
We ran the Python script, and selected the appropriate satellite sensor, desired Indices, and set the input image file, in the output folder for files to be stored. In our study, we exported the individual bands, and used NDVI, SAVI, and Tasseled Cap Brightness, Greenness, and Wetness Indices. To develop the Tamarisk Species Distribution Models, we used the we used the SAHM software package within the VisTrails program.
For our study, we opened the SAMH tutorial 2.0 VT file, that came with the package download, and selected the Independent locations workflow example, within the History view of this tutorial. Other examples are also available in this tutorial. We then went to the Pipeline view, to set up the models.
First, we selected Packages, to change the session folder. Next, we selected the Template Layer Module, and navigated to the Template Layer, that would define the projection, cell size, and extent of the study. Next, we selected the Field Data Module, and navigated to the CSV file, containing xy coordinates for tamerisk, that were collected.
Next, we selected the Predictor List file, and navigated to the CSV file containing our predictor list for this study. Next, we selected the Field Data Query to define the Response column, x column, and y column within our Field Data CSV file. Next, we selected the MDS Builder Module, and defined out background point count as 10, 000.
A Background Probability Surface option is also availabel in this location. We used a Background Probability Surface with values of 100, within a 5, 000 meter buffer of the Arkansas River, and 0, for areas outside this buffer. This was based on the areas sampled by the Tamerisk Coalition in our study.
Next, we added the Maxent Module to our workflow, and connected it to the Covariate Correlation and Selection Module. The Boosted Regression Tree, Generalized Linear Model, Multivariate Adaptive Regression Splines, and Random Forest Modules, were already located in the workflow. Next, we added a Model Output Viewer Module to the workflow, and changed the Column and Row to match the other Modules.
Next, we selected a unique output name as a sub folder name within the workflow. Next, we added an Ensemble Builder Module to the workflow, and connected it to all five models. We set our Threshold Metric, and Threshold Value for the Ensemble.
This can be changed based on study objectives. Next, we navigated to a CSV file, containing our independent test data set of tamarisk. Once again, with the Field Data Query Module selected, we define the Response, x, and y columns within the independent file.
We added an Apply Model Module to the workflow, and connected it to the MDS Builder Module for the independent data set, and the Maxent Module. We also added another Model Output Viewer, and connected it to the Supply Model Module, changing the Column and Row to match the other models. Next, we selected Packages, to change the Processing Mode to single model sequentially, allowing more than one core to be used during model execution.
The first screen to appear, is a Covariate Correlation Viewer, which indicates the correlation between any two variables. Our study objectives were based on dropping variables that were highly correlated, or greater than point seven, based on a generalized additive model. We used percent deviants explained for each variable, to decide which variable to keep, in the instance that two variables were highly correlated.
When we have made a decision on the number of covariants to keep, we selected OK.After the models are complete, a VisTrail spreadsheet will appear. This spreadsheet can be used to compare model results, including AUC Plots, Text Outputs, Response Curves, Calibration Plots, Confusion Matrices, and Residual Plots. For our results, there was very little difference between the five models, based on threshold independent, and threshold dependent evaluation metrics.
Based on these metrics, and after comparing probability surfaces produced by each model, we decided an ensemble of the five models was an appropriate approach for these data. Ensemble Mapping, aims to combine the strengths of several correlative methods, while minimizing the weaknesses of any one model. However, we caution that models which under perform, can weaken overall results.
Our results demonstrate fitting Boosted Regression Tree, Generalized Linear Model, Multivariate Adaptive Regression Splines, Random Forest, and Maxent, with presence points for tamarisks, and a time series, of remotely sensed Landsat satellite imagery, can distinguish tamarisk on the landscape, and is an effective alternative to traditional, single scene classification methods. Maps produced from these models, will provide an important management tool for targeted tamarisk control efforts in the study area.
