Using UAS to Identify Red Spruce in the Monongahela Forest

Revitalized efforts from the United States Forest Service have brought about a desire to restore the historic Red Spruce forest that once existed on Cheat Mountain, a range deep inside of the Monongahela National Forest. Much of this forest was clearcut by the Mower Mining Company in the early 1900's to gain access to the coal, giving the area the nickname of the "Mower Track". The Mower Track has not recovered in the time since; tree species were planted to "rejuivinate" the forest but the planted trees were non-native, creating an ecosystem mostly void of its natural functions. These misplaced trees have been removed, leaving the desirable Red Spruce standing; the focus now turns to identifying existing stands of Red Spruce and performing a canopy release, so that the Red Spruce can spread and thrive again. This project was brought about by foresters at the Monongahela National Forest in the form of a federal grant for Shepherd University. The foresters simply want to know if a drone can be used to identify Red Spruce trees. They hope that a healthy Red Spruce forest will bring environmental prosperity to the Monongahela Forest and serve as a tourist attraction to bring revenue to the state of West Virginia.

Over the course of the project, I brushed shoulders with foresters both on campus and in the field. Lots of information can be gathered on how foresters operate in the field by observation but there is some basic terminology that needs to be understood for proper communication. The methodology is simple in design, but becomes complex when practiced in the field, since nature has a way of not wanting to be categorized into simple groups.

We collected 228 images over two flights, below are some of the images take by the Mavic. Images were taken at 250 above ground level with a 33% fromtal image overlap (North/South) and 67% side image overlap (East/West) at equal interval snap periods.

Imagery was then uploaded into Agisoft Metashape Pro. Running through the standard workflow with the highest quality settings, were are able to generate a orthomosaic showing this portion of the Mower Track.

Structurally composed of 218 images, 51 million tie points, and 17 million different faces, the orthomosaic covers 51 acres of land.

At its tallest and widest extent, the orthomosaic is 13,976 x 20,804 pixels with a resolution of 3.5 cm per pixel, an incredible level of spatial quality for a piece of data this size.

When viewed from a distance the orthomisaic can be underwhelming, it is at first glace just another aerial image. Lets take a tour of the data to better understand it and see what is really on display here.

To meet the requirements of a capstone we need to declare a hypothesis and test it, therefore...

In order to properly test the hypothesis three main steps were taken.

• Data Collection - where we returned to the Mower Track, set up plots, and counted the Red Spruce in each plot
• Analysis - where Arc Pro's analysis through classification was used to identify Red Spruce in the orthomosaic
• Results - JUPYTR notebook was used for statistical analysis to find the quantifiable results

Here we can see where the 15 plot locations fall onto the orthomosaic.

Unsupervised Classification

The unsupervised classification was broken down into four classes, here are the results of the classification with the symbology altered to make the data presentable.

The resolution of the classification were higher than expected when mapped (turned out more distinct and usable then I thought it would be. The symbology was altered to allow for clear distinction between the trees and surrounding areas. The Red Spruce are highlighted in green and visually stand out from there surroundings as clusters of matching pixels. The shadows, the darkest values in the orthomosaic, are shown as brown, making identifying Red Spruce among the established forest impossible. There is also a lot of noise in this data. Green pixels are speckled through the orthomosaic, are these small groups of pixels Red Spruce seedlings or just areas of shadows?

Supervised Classification

The supervised classification was also organized into four classes using the Nearest Neighbor classifier. Each class had 20 samples for training and the final results uses the same symbology as the unsupervised classification.

The supervised classification shows more definitive results using the same sybology as the unsupervised classification. Vegetation, roads, and bare land can be discerned from each other with better clarity when compared to the unsupervised classification. The Red Spruce are clearly visible, highlighting the tree and the shadow that they cast. There is much less noise throughout the orthomosaic, showing possible seedling locations that were not visible in the unsupervised classification.

To better visualize the differences between the two classifications we can look at the map below with the plot locations. Neither map is superior over the other, both offer there own interpretation of the data.

With the plots marked and the classification done, the Red Spruce were then counted within the plots. The orthomosaic, unsupervised and supervised classifications were manually counted then put into a table alongside the ground truthing count. The subjectivity of an individual's distinction between a tree and its surrounding on the plot creates an expected degree of inaccuracy with this method, though a visual understanding of the data aids in this distinction.

The hypothesis questions how accurate a drone is at identifying conifers. In the data presented above, there are three types of data resulting from the drone imagery: the orthomosiac, the supervised classification, and the unsupervised classification.  All three need to be compared to the ground truthing in order to determine accuracy. This is done by testing for the variance between the four groups. The data is numerical, but we don't know if the data is normally distributed or if the variance between the four groups is parametric. A Shapiro-Wilks test is used first to determine if the data is normally distributed. 

Since only three of the four p-values are above 0.05, the data is not normally distributed. Next a Levene test is used to test if the groups are parametric. The results are a p value of p=0.067 telling us that the data is parametric since the p values is greater than 0.05.

Since the data is parametric but not considered normal, a Kruskal Wallis test was performed as the alternative to ANOVA to determine if any groups are statistically different from the others. This is followed by a post hoc-Dunn test to determine the statistical difference between the groups of data. The Dunn test can be thought of as a test for median difference. The result of the Dunn test is a z-score for comparison between all groups within a data set.

The table displays the results from the Post-Hoc Dunn test. The P-values in green display the three UAS based counts when put against the ground truthing. The P-values for any of the UAS based counts are all less than 0.05 when compared to the ground truthing, indicating that the counts by UAS are all significantly different from ground truthing.

For the purposes of this project, it can be concluded that using UAS as a form of cruising does not yield results to a statistically accurate extent when compared to ground truthing. Therefore, we reject the null hypothesis. There are many variables in the process, including the type of the UAS imagery used, method of counting, and interpretation of data that would give different results. This project shows that there is ample room for further testing on the subject and can serve as proof of concept that UAS can have potential as a replacement for traditional cruising.