Course Project- Eastern Redcedar

Quantifying Redcedar in Payne County, Oklahoma


Introduction

    Eastern Redcedar (Juniperus virginianai L.) is a native evergreen tree species of Oklahoma and the Great Plains, and the frequency of its occurrence has changed drastically over the past 100 years. In 1912, Harper (1912, p. 145) noted that it was “conspicuous by its absence” in the region, but by 2013, it was projected to cover 8.6 million acres in Oklahoma (Starks, Venuto, Eckroat, & Lucas, 2011). Due to poor land management practices, suppression of natural wildfires, and the prohibitive cost of removal, Eastern Redcedar has become an invasive nuisance for the state of Oklahoma and has caused millions of dollars in economic damages (Craige, 2013). This damage occurs in the form of depletion of water resources, reduction of cattle forage, hunting, and wildlife acreage, and fire damage. However, several marketable products can be produced from Eastern Redcedar, such as mulch, lumber, biofuels, pharmaceuticals, cedar oil, animal bedding, particleboard, and wood flour (Drake et al., 2002; Gawde, Cantrell, & Zheljazkov, 2009; McNutt, 2012).

    Craige (2013) developed an economic model to estimate costs and advise facility location for the harvest, transport, and processing of Eastern Redcedar in Oklahoma and for the distribution of the final products. This model relied on location-allocation that used “a highway network, impedance limit, biomass supply [points], and a list of potential facilities” to determine the best location for a processing facility (Craige, 2013, p. 66). The biomass supply points consisted of geographic locations of Eastern Redcedar with an associated approximate canopy cover. Canopy cover maps were provided by the U.S. Department of Agriculture (USDA) Natural Resources Conservation Service (NRCS). In these maps, canopy cover was divided into three classes: 10 – 30%, 30 – 70%, and greater than 70% canopy cover (Starks et al., 2011). Figure 1 is the Eastern Redcedar canopy cover map for Payne County. These maps were developed for 18 Oklahoma counties (only 17 were contiguous, central Oklahoma counties) using Landsat-5 and 7 satellite imagery that was collected between 2002 and 2005.

 

Figure 1. NRCS Eastern Redcedar canopy cover map for Payne County.


https://sites.google.com/site/tmooregis/redcedar/NRCS_Redcedar_Payne.png

    Because the model developed by Craige (2013) relies so heavy upon the Eastern Redcedar location and stand density, it is important that this information is accurate and up-to-date. Data collected in 2005 for a rapidly-spreading, invasive species is less than ideal for informing decisions about an industry based on that species. The dated and coarse nature of the data currently being used for Craige's (2013) economic model makes it necessary to collect more recent and higher quality data on the occurrence of Eastern Redcedar in Oklahoma.

Objectives

    The objectives of this project were to:
  1. Identify Eastern Redcedar occurrence in Payne County, Oklahoma using up-to-date spatial data. 
  2. Assess the accuracy of the Eastern Redcedar identification process. 
  3. Establish a methodological framework for identifying Eastern Redcedar in Oklahoma

Methods

    The U.S. Geologic Survey’s (USGS) The National Map viewer was used to download orthoimagery for Payne County, Oklahoma. The selected imagery came from the USDA Farm Service Agency’s (FSA) National Agriculture Imagery Program (NAIP) in 3.75 x 3.75 minute quarter quadrangles with a 300 meter buffer on all sides. This imagery had 1-meter resolution and was collected during the growing season of 2013 in the red (R), blue (B), green (G), and near infrared (NIR) bands (USDA FSA, 2013). To completely cover Payne County, 63 NAIP images were downloaded (partially displayed in Figure 2). All further analyses, unless otherwise noted, were performed in ArcMap 10.2 (ESRI, 2013).


Figure 2. A map of portions of the NAIP imagery within the boundaries of Payne County.

https://sites.google.com/site/tmooregis/redcedar/Pieces%20of%20Payne%20County%20NAIP%20(with%20reference%20map).png

    The mosaic to new raster tool was used combine the 63 NAIP images. Afterwards, the extract by mask tool, along with a shapefile of Payne County, was used to remove portions of the image that fell outside the county. Figure 3 was the result of these processes.


Figure 3. Mosaicked NAIP imagery for Payne County.

https://sites.google.com/site/tmooregis/redcedar/Payne%20County%20NAIP%20(Clipped).png

    Once the imagery was combined and masked to Payne County, training samples were made to direct image classification. Two sets of training samples were created: one using only a single quarter quadrangle image, and the other using the entire county. To prevent missing landcover types that were not Eastern Redcedar, training sample classes were collected for urban, water, vegetation that was not cedar, and Eastern Redcedar. Because the NAIP imagery was collected during the growing season, Eastern Redcedar was difficult to visually distinguish from other types of vegetation within the imagery. To improve this process, the bird’s eye viewing feature in Bing Maps was used to verify Eastern Redcedar location. This tool provided oblique imagery of the area and was taken in the winter, which made distinguishing Eastern Redcedar from other vegetation easier (Figure 4). Signature files were created from the two sets of training samples and used to run the Maximum Likelihood Classification in the Classification Toolbar.

 

Figure 4. Bing Maps’ bird’s eye viewing feature was used to inform creation of Eastern Redcedar Training samples.

https://sites.google.com/site/tmooregis/redcedar/Bing%20Birds%20Eye.png

    A confusion matrix was used to evaluate the accuracy of the two resulting classified images. This was done by creating a shapefile of reference points that were placed on specific landcover classes and labeled appropriately. Several reference points (15 – 25) were created across the entire image for each of the classes. The point to raster tool was then used to convert the vector reference points into single pixels. The combine tool was used to combine the raster of reference points with the classified image (this was done once for each classified image using the same reference points). The attribute table produced by the combine was exported as a dbase table and used in the pivot table tool. This produced a confusion matrix that was exported to an Excel spreadsheet. The confusion matrix was used to calculate overall accuracy, kappa coefficient, errors of commission and omission, and producer’s and user’s accuracy. The confusion matrix process followed the steps provided by TAMU (2013).

Results and Discussion

    Figure 5 is the resulting classified image of Payne County for the classification performed using training samples from a single NAIP quarter quadrangle. This classification placed all pixels within the image into one of the four classes, and the majority of the county was classified as vegetation/non-cedar. A quick look at the county as a whole reveals that there were some issues with the classification: most of Lake Carl Blackwell (on the west side of the county) was classified as urban and portions of the Cimarron River (runs along southern side of county) appear to be missing as it was classified as veg/non-cedar and urban. However, these were not major issues as the goal of the classification was to differentiate between Eastern Redcedar and non-Eastern Redcedar.

 

Figure 5. First classification of Payne County performed with training samples from a single NAIP quarter quadrangle.

https://sites.google.com/site/tmooregis/redcedar/First%20Classification%20(Clipped).png

    An issue that did conflict with that goal can be seen in Figure 6. In this image, it can be seen that the classification placed a large number of pixels that were in the middle of a lake (Stillwater Creek Site 46 Reservoir, southeast portion of enlarged image) in the Eastern Redcedar class. Additionally, the classification put a nearly solid band of Eastern Redcedar around most of the border of the lake, which was not supported by the Bing Map’s bird’s eye imagery. A closer look at the rest of the county revealed that both of these misclassifications were common throughout the image. These misclassifications were most likely due to pixels that were in shadow being captured in the training samples for Eastern Redcedar. Water bodies that were choppy due to high winds would explain the patchy classification of Eastern Redcedar within those water bodies. Shadow cast around the edges of water bodies by surrounding vegetation would explain the mostly uniform ring of Eastern Redcedar classification around many water bodies.


Figure 6. A portion of the first classified image, zoomed into an area of Payne County just south of Lake Carl Blackwell.
https://sites.google.com/site/tmooregis/redcedar/First%20Classification%20Zoom%20into%20Area.png

    Figure 7 is the resulting classified image of Payne County for the classification performed using training samples that were distributed about the entire county. Unlike the first classified image, this image contains a large number of unclassified pixels. The two classifications were performed using the same settings, so it is unclear why the second one had unclassified pixels and the first did not. This classification also had issues with correctly identifying water (a large portion of Lake Carl Blackwell is still classified as urban), but, as in the first classified image, this does not negatively affect the goal of identifying Eastern Redcedar.


Figure 7. Second classification of Payne County performed with training samples distributed across the entire county.

https://sites.google.com/site/tmooregis/redcedar/Second%20Classification%20(Clipped).png

    Figure 8 show the same area as Figure 6. The second classification was much less likely to classify pixels within water bodies as Eastern Redcedar. In addition to the obvious change of having unclassified pixels, a visual inspection of the second classified image reveals that a much greater portion of the county was classified as Eastern Redcedar than in the first classification. This is confirmed in Table 1, which shows that the second classification contained over 65 million (29%) more Eastern Redcedar pixels than the first classification.

Figure 8. A portion of the second classified image, zoomed into an area of Payne County just south of Lake Carl Blackwell.

https://sites.google.com/site/tmooregis/redcedar/Second%20Classification%20Zoom%20into%20Area.png

Table 1. The pixel counts for each of the landcover classes in each classified image.


    The results of the confusion matrix for the first classification are shown in Table 2. This confusion matrix indicated that the first classification correctly classified only 20% of the Eastern Redcedar reference points. Additionally, 100% of the reference points that should have been classified as vegetation/non-cedar were classified as Eastern Redcedar. The overall accuracy of this classification (number correctly classified divided by total number of reference points) was only 38%, and the kappa coefficient was 0.203. Overall, the first classification did a very poor job of determining Eastern Redcedar location.

 

Table 2. Results of the confusion matrix for the first classified image.


    The results of the confusion matrix for the second classification are shown in Table 3. This confusion matrix indicated that the second classification correctly classified 60% of the Eastern Redcedar reference points. Additionally, it only incorrectly classified 5% of the vegetation/non-cedar reference points as Eastern Redcedar. The overall accuracy of this classification was 53%, and the kappa coefficient was 0.441. While this classification is still not ideal, it is a vast improvement over the first classification.

Table 3. Results of the confusion matrix for the second classified image.


Conclusions

    Using the high spatial resolution, 4-band NAIP imagery, a maximum likelihood classification was able to identify Eastern Redcedar from other landcover types. While training samples for the classification could be gathered from a relatively small area, confusion matrix analyses indicated that collecting test samples from across the entire study area greatly improved the accuracy of the classification. This should be the baseline standard for all future work that might be associated with this project. For this reason, the image produced by the second classification should be used in future iterations of Craige's (2013) Eastern Redcedar economic model. This classification identified approximately 290 km2 of Eastern Redcedar in Payne County and had an overall accuracy of 53%. Figure 9 is the final map of Eastern Redcedar occurrence in Payne County, Oklahoma.

 

Figure 9. Final map of Eastern Redcedar occurance in Payne County, Oklahoma.

https://sites.google.com/site/tmooregis/redcedar/Final%20Redcedar%20Map%20for%20Payne%20County%20(clipped).png

Future Work

    This project went a long ways towards developing usable Eastern Redcedar coverage data for Payne County, Oklahoma, for use in Craige's (2013) biofeedstock economic model. However, additional data transformations may have to be made to the final output before it can be used directly in the model. Additionally, updated classification for Eastern Redcedar need to be performed for the 17 counties already included in the model and for the rest of Oklahoma. This process could possibly be written into an automated Python script to run classification on the individual NAIP quarter quadrangles without having to mosaic them together. This would greatly reduce the time and storage space needed to process the imagery. The limiting component to an automatic script would be having to gather representative training samples for the classification process.

    The confusion matrix used in this project consisted of fewer than optimum reference points. Developing a confusion matrix with a greater number of reference points would be more informative of the classification accuracy. Additionally, gathering reference points in the field with a global positioning system (GPS) would further improve the validation potential of the confusion matrix.

References

Craige, C. (2013, May). Biofeedstock Supply Chain Logistics Dynamic Modeling: Eastern Redcedar. Oklahoma State          University, Stillwater, OK.

Drake, B., Todd, P., England, D., Atkinson, K., Gerondale, G., Hart-Berton, D., & Hiziroglu, S. (2002). A Strategy for           Control and Utilization of Invasive Juniper Species in Oklahoma: Final Report of the “Redcedar Task Force.” Oklahoma          City, OK: Oklahoma Department of Agriculture, Food and Forestry.

ESRI. (2013). ArcGIS Desktop (Version 10.2). Redlands, California: Environmental Systems Research Institute.

Gawde, A. J., Cantrell, C., & Zheljazkov, V. D. (2009). Dual Extraction of Essential Oil and Podophyllotoxin from Juniperus         virginiana. Industrial Crops and Products, 30(2009), 276 – 280.

Harper, R. M. (1912). The diverse Habitats of the Eastern Redcedar and Their Interpretation. Torreya, 12(7), 145–154.

McNutt, M. (2012, September 2). Oklahoma Looks at Ways to Curb Spread of Redcedar Trees. The Daily Oklahoman.

Starks, P. J., Venuto, B. C., Eckroat, J. A., & Lucas, T. (2011). Measuring Eastern Redcedar (Juniperus virginiana L.) Mass    With the Use of Satellite Imagery. Rangeland Ecology and Management, 64(2), 178 – 186.

TAMU. (2013). Accuracy Assessment of an Image Classification in ArcMap [YouTube]. Retrieved from                       https://www.youtube.com/watch?v=FaZGAUS_Nlo

USDA FSA. (2013). National Agriculture Imagery Program. Digital Quarter Orthoquad.

ć
Thomas Moore,
May 7, 2015, 12:52 PM
Comments