The opening slide of my Thesis Defense
Pictured above is my opening slide of my Thesis Defense. Pictured on this slide is a slice of LiDAR I extracted in the software package ENVI, this slice is what was captured by the bounding box in the image at the top of my homepage. If you look closely in this side profile you can see a road running parallel to the slice. It was my objective in my thesis research to identify historic logging road networks in Point Reyes National Seashore underneath dense canopy. Since roads under canopy cannot be detected by aerial or satellite imagery, the only source of data I used in analysis was LiDAR. The webmap inserted here shows the LiDAR file I am working with, in the upper left you will see a pulldown to change the visible LiDAR points. With all of the LiDAR points turned on the logging roads are invisible - just like in satellite imagery. If you select just the ground points you can see what has penetrated the canopy and captured the topography of the region. These points were what was used to generate the slope model which is displayed under the points.
Detecting roads in LiDAR is difficult under any conditions, but degraded, eroded roads under canopy is especially difficult. The workflow is first to generate a slope raster of the ground points in the LiDAR point cloud. Next segment and classify this image in eCognition. These image objects were then used to train a convolutional neural network (CNN), which is a specific type of machine learning method. Then the resulting heatmaps were resegmented, smoothed, and converted into road network shapefiles. Historic aerial photography from before re-vegetation were used for an accuracy assessment.
Check out the JavaScript running this webmap at the bottom of the page.
Shown here is a training area containing one logging road. Although to the human eye this looks like a straight forward classification, it was not. Roads have a very similar slope profile to that of the hill tops of the region, and without other sources of data at hand to help distinguish the hilltops from the roads, other methods are needed.
First the raster is segmented into image objects. Errors associated with hilltops were removed through writing a set of rules that defined the difference between a road and a hilltop. There are still a few errors here and there, but generally speaking there are enough correct samples to start building a neural network.
Next I trained the neural network and then applied it to my validation area. You can see that now this area is different than my training area. This is so I'm certain that I'm not just training a model to identify features in only one area. Shown here is the applied model. In green are areas that I have trained the model to identify as a high likelihood of not being roads. In red are areas with a high likelihood of being roads.
Shown here is an aerial photo used in the accuracy assessment, if you compare it to the neural network you can see how the network paths align to the road network. The LiDAR I am using is from 2018 and this historic photo is from 1963. Despite ~60 years of naturalization, the lasting effects of these roads on the morphology of the region can still be seen.
My thesis was accepted in May of 2022 and published in August of 2022. Check out the publication on Scholar Works.
The 3d webmap above was created in CesiumJS. CesiumJS is an opensource library for tiling, visualizing, sharing, and analyzing 3d geospatial data. Unlike other libraries, like Leaflet or Mapbox, Cesium is out-of-the-box 3d. You can see below that in just under 100 lines of code I was able to create a custom 3d webmap with a pulldown to filter my point data.
Any questions at all just drop me a line, I usually respond within a couple days.