First Advisor

Adam Booth

Term of Graduation

Winter 2021

Date of Publication


Document Type


Degree Name

Master of Science (M.S.) in Geology






Landslides -- Maps -- Automation, Remote sensing, Geomorphology



Physical Description

1 online resource (xi, 110 pages)


Recent advances in remote sensing data and technology have allowed for computational models to be designed that successfully extract landforms from the landscape. The goal of this work is to create one such semi-automated model to extract deep-seated landslides located in complex geomorphic terrain. This is accomplished using geographic object-based image analysis (GEOBIA) techniques, considered by leaders in the field of image analysis to have an advantage over traditional automated classification methods. GEOBIA methods can mimic human visual interpretation by including more characteristic features used to assess the relationship between image data and the ground surface such as color reflectance (spectral), texture, shadow, location, pattern, height, tone, context, size, and shape.

The standard method for identifying and mapping landslides in the Pacific Northwest is for professional geologists to manually delineate landform features using remote sensing data, referred to as remote mapping. The method is currently employed by United States Geological Survey (USGS), Washington State Department of Natural Resources (WA DNR), and Oregon Department of Geology and Mineral Industries (DOGAMI). The question remains if semi-automated models can perform as well as independent manual mappers when identifying landslides, while reducing bias due to interpretation discrepancies between mappers.

To test this hypothesis, two modeled landslide datasets are created. The first, using a model design that was not influenced by manual mapping efforts, and the second created using manually-mapped landslides for visual reference. These two modeled datasets are then compared to a manually-mapped landslide inventory, created with input from four professional geomorphologists. Differences in landslide numbers, densities, geometries, and extents, that were delineated by the geologists, reflected the range of professional backgrounds. The spatial area of landslides that was delineated by all geologists (i.e., overlapping spatial area where manual-mappers delineated terrain as a landslide) is used as reference to verify landslide areas from the first modeled dataset. Agreement statistics (i.e., accuracy) suggest 81% of the modeled landslide area are appropriately delineated by the model in this study. The second set of modeled landslides are verified by comparing the spatial dataset to all landslides inventoried by the geologists (i.e., any terrain delineated as a landslide). Using all landslides eliminated data filtering, that could introduce bias in the reference inventory. Agreement statistics (i.e., accuracy) suggest 78% of the modeled landslide area are appropriately delineated. Perhaps more interesting, agreement statistics for recall, emphasizing correct identification for all landslides, suggest 69% of the area is correctly identified as a landslide. This is compared to landslide area recall between the manual mappers which ranges from 35-99% within the study area.

The data suggest the model is objectively using a set of morphometric characteristics to map the landslides, while the professional geomorphologists have developed interpretation style biases that lead to a large range in area mapped as a landslide. Incorrectly identifying terrain as stable could have negative impacts on public safety, suggesting more research is necessary to determine the true population of landslides that exist on the landscape. Automated models can be useful with that effort.


© 2021 Tiffany E. Justice

In Copyright. URI: This Item is protected by copyright and/or related rights. You are free to use this Item in any way that is permitted by the copyright and related rights legislation that applies to your use. For other uses you need to obtain permission from the rights-holder(s).

Persistent Identifier