Automated Earthwork Detection Using Topological Persistence

For thousands of years, humans have altered the movement of water through construction of earthworks. These earthworks remain in landscapes, where they continue to alter hydrology, even where structures have long since been abandoned. Management of lands containing earthworks requires an understanding of how the earthworks impact hydrology and knowledge of where the structures are located in the landscape. Various methods for detecting topographic features exist in the literature, including a set of rule and threshold-based techniques and machine learning methods. These tools are either labor-intensive or require special pre-processing or a priori assumptions about structures that limit generalizability. Here, we test a topological analysis tool called "persistence" to determine if it is useful for earthwork detection in rangelands. We found that persistence can be used to detect earthworks with 83% precision and 64% accuracy. Breached berms and berms with significant upslope sedimentation are most likely not to be detected using persistence. These results indicate that persistence can be useful for terrain analysis, and it has the potential to substantially reduce manual effort in feature detection by identifying regions where berms may be found. The shape of landscapes controls how water moves over the surface. Humans have modified landscapes by building earthen structures to direct water for thousands of years. The legacy of human water management systems remains in many places around the world, and land managers need to understand how these structures impact hydrology and where they are in order to support management decisions. Earthen dams, berms, and stock ponds dot the southwestern United States. High-resolution elevation maps and imagery are increasingly available, and although man-made structures are readily visible to the eye, automatic detection remains a challenge. In this study, a method from the mathematical field of topology called "persistence" was applied to automatically detect berms and stock ponds from elevation arrays. This method identifies features from a series of binary images generated from the same elevation map at consecutive threshold values. Results demonstrate that persistence is able to detect well-defined berms. While not all berms are detected, these results are still promising since berms are generally found in groups. Thus, detection of even two thirds of the berms substantially narrows down regions for manual inspection. Persistent cohomology can be used to summarize topographic information A persistence threshold is adequate to identify berms and stock ponds from focused regions A set of criteria calculated from persistence can detect 64% of observed berms