Executive Summary

Increases in trans-oceanic commerce, terrestrial transportation systems, and varying land use practices have ensured that many nonnative, invasive plant species have either accidentally or deliberately been introduced into places that are well beyond their native range. Once established, invasive species have the ability to displace native plant and animal species (including threatened and endangered species), disrupt nutrient and fire cycles, and alter the character of the community by enhancing additional invasions.

The consequences of nonnative invasions in our National Parks have not gone unnoticed.
Approximately 2.6 million acres (1.1 million ha) of national park lands are effected. Big Bend National Park (BIBE) in southwestern Texas is an example. Over 65 exotic plant species were documented in BIBE in 1998. In the region around BIBE, the National Biological Information Infrastructure reports over 200 exotic plants that could potentially influence the Park’s environmental quality, biotic health, and biotic integrity

Developing early detection methods for invasive species could result in substantial economic savings and circumvent negative ecological impacts. Components of early detection protocols include: 1) knowledge on the current presence or absence of a given plant species in or near management units, 2) knowing vectors and pathways of plant dispersal, and 3) understanding the likelihood of the plant establishing or spreading inside management units. This knowledge can then be used to plan a rapid response to remove or control the invasive plant.

Predictive modeling using remotely sensed and Geographic Information System (GIS) data can provide resource managers with a tool for early detection of invasive plants, especially over large landscapes. Once areas of invasive species occurrence are predicted, ground reconnaissance can be more effectively used for verification and control. Over the last decade, there has been a substantial increase in the use of remotely sensed and GIS data to model invasive species distributions or potential habitats. This increase coincides with improved remote sensors, computer technology, and classification techniques.

We evaluated the efficacy of using remotely sensed and GIS data as tools to support early detection of invasive plants in BIBE. To accomplish this evaluation, we identified four objectives for this study: 1) identify a subset of invasive plant species that occur within BIBE that may be useful for landscape scale modeling, 2) create spatially explicit models of predicted distributions of targeted species using remotely sensed and GIS data, 3) conduct a risk assessment of areas for invasive plant incursion, and 4) conduct computer simulations of common field sampling designs to evaluate efficient ground survey methods to detect invasive species in BIBE.

We choose eight invasive plant species for our predictive modeling: giant reed (Arundo donax), Bermuda grass (Cynodon dactylon), Lehmann’s lovegrass (Eragrostis lehmanniana), horehound (Marrubium vulgare), buffelgrass (Pennisteum ciliare), Russian thistle (Salsola kali), Johnson grass (Sorghum halepense), and saltcedar (Tamarix sp.). These species are a concern because they rapidly colonize disturbed areas, outcompete native species, and alter natural processes that further promote invasive species invasions and establishment. We also built a predictive model of all eight species combined to investigate areas where management efforts could impact multiple species.

Spatial locations on existing known occurrences in BIBE were provided by Park staff. This dataset included present and historic plant locations from opportunistic sightings and planned roadside surveys from 2001 to present. We conducted additional ground surveys in June of 2006 throughout BIBE to ascertain additional locations of target species populations. We used a systematic (opportunistic) sampling approach to optimize our chance of detecting additional invasive plant populations. We systematically surveyed roads and trails throughout BIBE. Arroyos, springs, and cross-country surveys were also conducted.

We used Landsat 7 ETM+ data for this study because of the large size (roughly 320,000 ha (790,734 ac) subjected to analyses. We used three seasons of spectral data across five years (Fall 1999, Summer and Fall 2000, Spring and Fall 2001, Spring 2002, and Spring 2003) to account for dynamic habitat conditions that occur in the study area due to discrete rainfall events and to provide a means to capture phenological differences among the targeted species and the surrounding landscape. We used MaxEnt software to generate predictive habitat models for each target species. MaxEnt uses the principle of maximum entropy to estimate the target probability distribution that has the broadest distribution compatible with the information available. Outputs include a probability surface with values from 0-100, tabular and graphical representations of model performance, and model variable contribution. We randomly withheld 10% of the occurrence points for each target species to assess model performance.

Model performance was assessed based on threshold-dependent and threshold-independent metrics. Threshold-dependent metrics require a known threshold to classify a response as presence/absence, or suitable/not suitable. We selected a threshold that provided an equal tradeoff between the test data sensitivity and specificity. We also evaluated model performance using threshold-independent receiver operating characteristic (ROC) curves. The advantage of ROC analysis is that area under the ROC curve (AUC) value provides a single measure of model performance, and can be interpreted as the likelihood that habitat quality in the predictive model is correctly classified at a random positive site and a random negative site.

Our risk analysis incorporated a system of numerical ranks and weighting of six spatial variables: 1) occurrence locations, 2) habitat suitability, 3) disturbance, 4) hydrologic features, 5) wind direction, and 6) fire history. Numeric weights and ranks represented each variable’s relative importance in the establishment and spread of our target invasive plant species, and allowed us to combine multiple stressors, habitats, and invasive species occurrences in to a spatial model that represents areas most at risk of invasion by nonnative species. The final risk surface was created using an additive Boolean overlay and resulted in classes of low, moderate, and high risk.

We used program SAMPLE to run sampling simulations that identified the most efficient method for sampling each species and evaluated if adaptive sampling techniques would benefit sampling effort. SAMPLE allows for the evaluation of single-stage, two-stage, and stratified sampling designs in the context of adaptive and non-adaptive strategies using simulations generated from GIS datasets.

A total of 1,464 invasive plant locations were documented for the eight species of interest. The number of occurrence points used to generate individual species models ranged from 26-627. Small sample sizes (< 50 occurrences) occurred with giant reed, horehound, and Russian thistle. The influence of small sample sizes on model performance and variable selection is unknown.

All predictive plant models performed better than a random prediction, with the exception of the Russian thistle model. Test AUC values ranged from 0.77 to 0.99. Confidence intervals (95%) built around test AUC values indicate models did not vary across discrimination classes, with the exception of Russian thistle. Horehound and giant reed models yielded high test AUC values, narrow 95% confidence intervals (0.99 + 0.02), high model gain values (3.4) and low errors of omission (˜ 0%; P<0.01). However, model assessment values for these species may not be robust due to the low number of test samples. Likewise, AUC estimates and 95% confidence intervals for the Johnson grass and saltcedar models yielded high AUC values (˜ 0.92 + 0.02) with moderate errors of omission (˜ 15%; P<0.01).

Model test AUC values indicated moderate discrimination for buffelgrass (0.88 + 0.01), Bermuda grass (0.85 + 0.05), and Lehmann’s lovegrass (0.81 + 0.02). Errors of omission were high for Bermuda grass (27%) and Lehmann’s lovegrass (28%), but were moderate for buffelgrass (19%). We observed similar levels of discrimination when combining all plant locations into one predictive model. The combined target plant predictive model resulted in a test AUC value of 0.82 + 0.004, and a 27% error of omission. Combining data from all target plants likely created a habitat generalist dataset which may have reduced our ability to differentiate suitable habitats (resulting in a moderate AUC value) or detect vegetation phenology differences using the maximum entropy approach. However, the larger sample size resulted in greater model precision (low AUC standard deviation). Our Russian thistle model yielded a moderate AUC value (0.77) with 95% confidence intervals ranging from 0.58 to 0.96 and a high error of omission (33%), effectively reducing the reliability of the model.

We used analyses provided in MaxEnt to guide us toward parsimonious models. However, even with these procedures, the number of variables retained in each model was relatively large and varied between 15 and 35 spectral classes. Our results indicated that models with larger sample sizes (number of plant occurrences) yielded models with greater number of variables. Further, the greater number of variables in the models reduced the likelihood that the model could differentiate plant occurrence locations from random background pixels. These results likely reflect MaxEnt’s attempt to estimate the plants broadest probability distribution given an increase in variability associated with greater numbers of sample locations and the increased variability associated with greater number of model variables. Our giant reed, horehound, and Russian thistle models may be overfit due to their low number of occurrence locations. We recommend additional model selection processes including correlation analyses, or evaluating multiple models using Akaike’s Information Criterion (AIC).

Giant reed and horehound had the most limited predicted distribution in BIBE, with approximately 2% of the park surface modeled as potential habitat. Giant reed predicted habitat was fairly restricted to the Rio Grande corridor while horehound predicted habitat was primarily around developed areas near the Chisos Mountains and small patches near Rosillos Mountains. We estimated approximately 14% of BIBE was suitable for saltcedar. Saltcedar occurs in large stands along the Rio Grande with smaller stands fairly common in arroyos and springs that have adequate seasonal water for establishment. Buffelgrass is a potentially increasing problem for BIBE. Approximately 20% of BIBE was predicted to have suitable conditions for this plant. Large stands of buffelgrass already occur in the southern part of the Park, but this grass can be found throughout BIBE. Bermuda grass is another large problem in BIBE. We estimated approximately 27% of BIBE to be suitable for Bermuda grass. This grass is already well established in developed areas, campgrounds, and along the Rio Grande where it has displaced native vegetation. Johnson grass had the largest predicted amount of potential suitable habitat, approximately 40% of BIBE. Johnson grass is well established along the North Rosillos Road, but potential habitats occur throughout the Park.

Based on model performance, our proof-of-concept approach worked well for giant reed, horehound, Johnson grass, and saltcedar, and moderately well for buffelgrass, Bermuda grass, and Lehmann’s lovegrass. The efficacy of using remotely sensed data to model the potential distribution likely was influenced by invasive patch size, habitat generality, and similar environmental conditions surrounding invasive plant populations. Modeling approaches that employ presence/absence analytical approaches may help to differentiate these communities. Further, integrating other spatial information (e.g., disturbances, and soil moisture regimes), high spatial resolution data, or larger sample sizes may improve model performance.

Our risk assessments were designed to evaluate the risk of landscape parcels being further invaded or established by our target invasive plants. Areas at risk included Panther Junction west to Croton Springs, Chisos Basin, Tornillo Flat near U.S. 385, Tornillo Creek near State Highway 118, Rio Grande Village, Boquillas Canyon, drainage between Dominguez Spring and the Rio Grande, riparian habitats from Cottonwood Campground to Santa Elena Canyon, East Corazones Draw, and paved roadways, especially Chisos Basin Road, Rio Grand Village Road, Boquillas Canyon Road, State Highway 118, Ross Maxwell Scenic Drive, and Persimmon Gap Road (U.S. 385).

The amount of BIBE at high risk to invasion varied from approximately 188 ha (465 ac) to 1,957 ha (4,836 ac). Assessments using all eight target species indicated approximately 3,976 ha (9,825 ac) are at high risk of invasion and 25,027 ha (61,843 ac) are at moderate risk in BIBE. Moderate risk landscapes may develop into high risk landscapes if invasive plant species are found in close proximity, disturbance regimes enhance habitat suitability, or suitable vectors and pathways are developed linking landscapes. Although the acreage at risk represents < 10% of the entire Park, habitats are not equally at risk. For example, habitats that host greater human infrastructures (roads, trails, campgrounds, and developed areas), are at greater risk to incursion by invasive plants. Likewise, our analyses indicated that riparian habitats may be 3-times more likely to become invaded than other general habitat types.

We evaluated six sampling designs using simulations derived from known plant occurrences and potential distributions in BIBE. Our results indicate that for single-stage designs, an adaptive simple random sampling procedure could provide reliable estimates for six (75%) of the species evaluated. We were unable to evaluate complex sampling designs with our datasets. However, sampling procedures should consider areas at risk of invasive plant incursion. Including areas at risk as a stratum or using the risk assessment to alter plant probabilities may increase the likelihood of detecting rare species and increase the utility of field sampling procedures for early detection protocols.