Improving Landslide Inventories by Limiting Land Classification to Drainage Areas of Debris Flow-Dominated Channels

Landslide inventories, frequently created by aerial photograph interpretation (API), are often used in the production of hillslope hazard maps to characterize past landslides or to evaluate a hazard model. In the former application of inventories, potential landslides in hazard maps are delineated as areas that have similar morphometrics as past landslides at locations of modeled hillslope instability. Therefore, the accuracy of the inventory has a strong influence upon hazard extent. In the latter application, the partial inventories that sometimes result from API, due to the subjectivity of interpretation and revegetation of landslides, likely results in incorrect evaluations. A more complete, less subjective technique is needed to not only better characterize past landslides and improve evaluation of hazard models, but also to assess the extent of areas prone to significant mass wasting in mountainous regions due to the evolution of landscapes. Inventory accuracy continues to improve with new technology and automated techniques, though rarely is the form of a channel's topography incorporated into the inventory process despite the growing evidence of a topographic signature of debris flows. This signature demarcates the transition between the dominant channel erosional process: fluvial or debris flow. These process transitions are often observed at scaling breaks in log-log plots of a channel's drainage area versus slope (DS plot). The scaling breaks, above which the effects of fluvial power laws upon channel topography are not observed and below which debris flow scars are not found, may signify the lowest point in the watershed where debris flows occur. We present an inventory technique that limits a land classification algorithm to areas that are upstream from this scaling break determined from DS plots of five streams in the Great Smoky Mountains National Park (GSMNP) region of the southern Appalachians. Topographic data for the DS plots and the classification inputs is extracted from a lidar-derived digital elevation model with a resolution of 4 m. A normalized difference vegetation index map created from an aerial photograph is also used as a classification input as a surrogate of vegetation canopy thickness. A map of debris flow candidates is produced by multi-resolution, contextual segmentation and classification and is compared to a previously published API landslide inventory of GSMNP. 67% of debris flows are identified and 79% of the land surface is correctly classified when the algorithm is limited to areas that drain to debris flow dominated channels. Candidates in this map include numerous features that have similar characteristics as debris flows delineated in the API inventory. When the entire study area is included, 60% of debris flows are identified and 61% of the land surface is correctly classified. This inventorying method, while limited to one landslide type, allows for a focused approach to statistically characterize the land surface and results in more accurate landslide identification in this study area of the automated techniques. Future work will expand this technique's utility to landslide terrain in different physiographic settings and include field verification of the previously unmapped landslide areas.