By Leila Donn, Tim Beach, Cody Schank, Mike Mallner Dept. of Geography, University of Texas at Austin
Computer programs that can automatically identify landscape features are extremely useful for environmental study, conservation and resource management, and natural hazard identification. I am developing a machine-learning computer program to find unmapped cave entrances under forest canopy using Python, GIS, and LiDAR imagery. Such a method can be applied to any landscape feature with distinct morphologic characteristics, even features obscured by vegetation.
I trained the first version of this cave-finding program with a Python machine-learning library using a dataset that included a shapefile of known cave locations to which I added randomly selected locations where there are not caves. I then used LiDAR imagery and ArcGIS to generate values for the morphologic characteristics of these points, including slope, standard deviation of elevation as a measure of slope roughness, land surface curvature, distance from nearest stream, and fill difference. Next, I ran this trained program on LiDAR imagery from an area of northwestern Belize that is under dense tropical forest canopy, where there are no mapped caves. The program identified a number of potential cave entrance locations across 140 km2, concentrated in four areas.
In June 2019, ground-truthing completed by me and a fellow caver confirmed the presence of a number of predicted unmapped caves, sinks, and cave-like features, including a 60‑meter-long by 30‑meter-wide by 35‑meter-deep sink feature with a collapsed cave entrance at the bottom. Predication accuracy will be improved in the next version of the program by incorporating point cloud low points, local relief modeling, and canopy height modeling. I also plan to acquire a larger training dataset; run the re-trained program over imagery from Belize, Guatemala, and Mexico; and to ground-truth some of these areas next summer.