Peregrine Falcon (Falco peregrinus) Nesting Habitat Suitability Model
for the Shasta-Trinity National Forest

---

Dale Jackan
American River College, Geography 350: Data Acquisition in GIS; Fall 2017

There have been little research on Northern California populations of Peregrine Falcons (Falco peregrinus) in National Forest. This project was an attempt to help focus survey efforts in the Shasta-Trinity National Forest (STNF). I was provided with Peregrine Falcon nest locality data by a raptor biologist at the Shasta-Trinity National Forest. A Maximum Entropy Model and the Maxent program to create the habitat suitability model (HSM) due to the small presence sample. Ten environmental variable as potential predictors of Peregrine Falcon nesting habitat. The variables were slope, roughness index, distance from roads, wind exposure index, distance from railroads, heat load index, distance from trails, distance from lakes, elevation, and aspect. The resulting model showed that slope, roads, and terrain heterogeneity (roughness) were the strongest indicators of Peregrine falcon nesting habitat. Unfortunately this model has low predictive value due to the elimination of many locality data points as a result of quality control standards.
Preferences by Peregrine Falcons may vary with region, but a few known factors appear to be fairly consistent across their range. They prefer to have open habitat near nesting sites, they nest on cliffs, and are near water (Barnes 2011, Comrack et al. 2008). Some of these associations may be related to their hunting behavior and water may attract prey. I will note that peregrines do not only nest on cliff faces and worldwide in some places tree nesting was more widespread prior to DDT contamination (Comrack et al. 2008) (Brambilla et al. 2006). Given my small presence sample size a Maximum Entropy Model was used to create a suitable nesting habitat model for the Shasta-Trinity National Forest and some of the surrounding area.


Kevin Cole from Pacific Coast, USA (https://commons.wikimedia.org/wiki/File:Peregrine_Falcon_in_flight.jpg), „Peregrine Falcon in flight“, https://creativecommons.org/licenses/by/2.0/legalcode

---

Figures and Maps

---

Variable	Percent contribution	Permutation importance
slope	34.3	42.1
rough	31.3	0
roads	21.2	55.2
exposure	6.4	0.1
rxr	2.7	2.4
heatld	2.1	0.1
trails	2	0
lakes	0	0
dtm	0	0
aspect	0	0

Figure 1. Represents the contribution each variable makes to the model. The permutation importance represents the gain as the model is fitted.
Figure 2. These are the models response curves to each variable. More or less it gives you an idea if the response was positive or negative and how quickly the response was.

Variable	Fitted-lambdas	Minimum	Maximum
Aspect	0.0	-1.0	359.6556091308594
dtm	0.0	30.299999237060547	4029.403076171875
Exposure	5.649085624345769	0.7818626165390015	1.3125859498977661
Heatld	0.5116726733240735	0.5014203786849976	1.0279760360717773
lakes	0.0	0.0	23151.560546875
roads	-51.68423083717352	0.2838197946548462	65234.0390625
rough	0.0	0.0	105.21299743652344
rxr	3.4626300584649505	30.27614974975586	149058.90625
slope	12.409922976119857	0.0	61.11526107788086
trails	-1.3919254364894016	0.26538679003715515	123160.796875

Figure 3. This is a jackknife chart essentially where the blue informs you of how important the variable to the model by itself. The teal color informs you of the model response if that variable is removed. If there is not much of a reduction in gain with it's removal most likely there is some correlation among variables.
Figure 4. The most important part of this chart are the fitted_lambdas. Essentially they inform you of how the model responds to the variables. If there are negative values there is a negative response and if positive values the response are positive. These values also give some indication of the weight given to each variable.


Of the ten variables used in this model seven appear to have some contribution to the model (fig.1). Slope has the highest contribution followed by the roughness index and roads. The three variables that had no contributive value to the model were lakes, elevation, and aspect.

The slope had the highest contribution and strongest predictive value among all of the variables with a moderate positive sigmoidal response as slope increased, and the largest decrease in gain after removal from the model (fig. 2,3). Given the preferred nesting habitat is on cliff faces this makes sense (Barnes 2011, Comrack and Logsdon 2008).

The roughness index showed a strong positive sigmoidal response as roughness increased( fig. 2, fig.4). Basically this is a measure of topographic heterogeneity and given that cliff faces are atypical for the area I would expect areas with higher heterogeneity to include greater slope variation. The removal of the variable from the model had little affect on the gain and is therefore highly correlated with other variables (fig. 3).

For roads there was a moderate negative sigmoidal response curve as distance from roads increased (fig.2, fig. 4). Roads do show a strong positive response to being in the model and show some response to their removal (fig.3). Though that may mean it is an important variable this is most likely the result of sample bias and more presence samples are needed to clarify the response. This is also where the understanding of Peregrine falcon behavior becomes important and an understanding of how the data was collected. The bias is most likely due to roads giving easier access to remote areas and therefore more sightings near roads. The subject matter expert (SME) who provided the presence samples did inform me that some raptors actually do have a positive response to roads due to greater availability of perches near an open area, and may even scavenge road kill.

For the wind exposure variable there was a weak positive sigmoidal response as wind exposure increased (fig.2, fig 4). This is most likely due to correlation between slope and wind exposure, there was only a small decrease in gain as the wind exposure variable was removed from the model (fig 3 jack). Peregrine falcons prefer very steep sloped nesting sites, generally on cliff faces, this habitat choice generally indicates a greater wind exposure (Barnes 2011, Comrack and Logsdon 2008).

For the railroads variable there was a positive effect that did not change over distance (fig.2). After examining the jackknife test the variable on it's own has no value and should be removed from the model (fig. 3).

The model has a very weak positive sigmoidal response to an increasing heat load index (fig.2. fig.4). Heat load index is a relative measure of how much heat a particular area may retain based on hours of sunlight while also taking into account aspect, as in southwestern aspects will generally retain more heat than all others (Evans et. al. 2014).

For the trails variable there was a very weak negative sigmoidal response as distance from trails increase (fig.2, fig. 4). This is most likely a artifact of sampling bias given sample sites are more often near roads and trails given ease of access. Given that the gain does not decrease after exclusion form the model more presence samples are needed to clarify the response (fig. 3).

Unfortunately this model has low predictive value due to the elimination of many locality data points as a result of quality control. However stricter quality control standards are now being implemented for data collection in the Shasta-Trinity National Forest and the variable layers still hold value having been created and processed for habitat modeling. If a protocol or structured survey methodology is implemented in the STNF with modeling in mind then additional locality points will improve the accuracy of this model. This project makes no attempt to assess the accuracy of the model. A lot of models base there data points on presence or observation points of a species, as does this model. The difference for this model is the unit of study is the nest locality, not where a peregrine falcon is observed. What this means is naturally this model will be derived from fewer presence points because there are always going to be fewer nest localities as opposed to observations of individuals anywhere spatially. In the future there are steps that can be taken to fine tune this model such as using environmental variable data for the species over a much larger portion of the range and use a portion of it as background samples. Additional suitable variables will continue to be added as they are found, and others rejected as the analyses dictate. Developing and improving this model will be an ongoing iterative process as I learn new techniques, as additional research is published, and as I develop a greater understanding of the subjects encompassing habitat modeling and prediction.

---

Acknowledgements

---