Kerry McMahon – Classification and Change Detection of Erosion Features in the Ka’a’awa Valley Using Object-Based Image Analysis


[ hide ]

    Classification and Change Detection of Erosion Feature in the Ka’a’awa Valley Using Object-Based Image Analysis

    Author: Kerry Deneene McMahon

    Towson University, Towson, MD

    This research made possible through the National Science Foundation Geospatial Research and Mapping Research Experience for Undergraduates (GRAM REU), in cooperation with California State University Long Beach (CSULB) and University of Hawaii, Manoa


    The Ka’a’awa Valley, while a pristine and protected piece of Oahu’s natural land, is also home to several recreational activities including horseback riding, hiking, biking, and ATV rides. In addition, the Ka’a’awa Valley is home to a working cattle ranch. These activities, combined with over 1700mm of annual rainfall, make this valley subject to accelerated erosion, as evident when entering the valley proper.

    While natural erosion occurs at a rate where the eroded state is replaced at the same rate at which it erodes, accelerated erosion, typically caused by human or other unnatural activities, can exceed 40% of the natural occurrence (WWF.Panda.Org). This erosion can be seen throughout the valley in the form of roadside cuts, hill slides and slumps, deep channels or gullies, and other concentrated flow paths. Accelerated erosion causes the loss or degredation of valuable topsoil, which houses the nutrients and micro-organisms necessary for new vegetation to grow and thrive. In addition, sediment loosened by erosive forces eventually ends up in streams, rivers, and/or the ocean, potentially harming vegetation, sea life, coral, and other organisms (DNR, Australia).

    Erosion works in a positive feedback loop. Once erosion occurs on a surface, the bare surface accelerates the erosive power of flowing water, as well as allows wind to pass over more easily, further accelerating erosion, even without outside influences. This means that erosion features can take tens to hundreds of years to ever repair themselves, even when left undisturbed by anthropogenic impacts.

    This research will classify and measure the extent of erosion patches contained within the Ka’a’awa Valley.

    Study Site


    Using object-based image analysis (OBIA) and eCognition software, a segmentation and subsequent classification of erosion features will be created using imagery obtained from Worldview 2 satellite for the year 2011. In addition, a segmentation and classification will be performed using  Digital Ortho Quad imagery from the year 2000, provided by the USGS.

    After completing both classifications, conversion to a shapefile format will be performed and the resulting files will be opened in ArcGIS for further analysis. The goal is to determine the extent (in acres) of erosion patches for the years 2000 and 2011, whether or not any change has occurred, and  to determine an estimated percentage of potential grazing land that is affected by these patches.

    In the field, GPS coordinates will be recorded for observed erosion patches that are accessible by foot. These GPS points will serve as error control ground truth points to supplement the erosion classifications performed in eCognition. In addition, the location of cattle, cattle excrement, cattle guards, and fences will be recorded, in an attempt to create a “potential grazing area” where it is assumed that cattle can travel.


    Using OBIA in eCognition, command-based classification was used after an initial segmentation was performed. The segmentation consisted of equally weighted bands and an object setting of 30. This ensured that all erosion patches would be gathered as unique objects, and not be blended with neighboring vegetation.

    Portion of Segmentation Using OBIA and eCognition

    Step 1 – Shadow Extraction Term1: Mean Yellow<100

    Step 2 – Water Extraction Term1: Mean NIR1<50, Term2: Mean NIR2<50

    Step 3 – Vegetation Extraction Term1: Mean Red Edge>500

    Step 4 – Residential Extraction Term1: Mean NIR1<150, Term 2: Mean Red<130

    Step 5 – Erosion Extraction Term1: Mean Red>150


    Because of the drastic difference in reflective properties between soil and vegetation, the extraction process required only a few commands to achieve the results The final classification performed in eCognition on the 2011 Worldview 2 imagery yielded the following:


    2011 Classification of Erosion Using eCognition and ArcMap


    Methods, continued

    Once the classification of the 2011 imagery was performed, it was converted into shapefile format and imported into ArcMap, where the 124 collected GPS points were layered in order to determine the accuracy of the classification. These 124 points were collected on the ground in the Ka’a’awa Valley where observed erosion patches exceeded 1.8 meters X 1.8 meters. Patches significantly smaller than 1.8m² would likely not show up in a classification due to the spatial resolution of the source imagery. Of 124 collected points, 121 were accurately classified using eCognition.

    Erosion Feature Collected by GPS

    Screenshot of Ground Truth Points

    In addition to checking for classification errors, a polygon had to be designated as “potential cattle grazing” in order to determine how much land in the study area was being affected by erosion. The initial classification included areas that were outside designated grazing areas, including roads, yards and beaches. It was important, to ensure accuracy, that these areas NOT be included in an acreage calculation.

    A slope layer was included to assist in this determination. When erosion features were collected by GPS unit, evidence of cattle presence was always present, either by cattle sightings or excrement sightings. Cattle or excrement were never sighted above where the pasture met the tree line, so a polygon representing an estimated cattle grazing potential region was created based on these facts. Using the slope layer generated from the DEM, it was assumed that cattle would not travel beyond a slope exceeding 55 degrees. This was partially based on the fact that no cattle were sighted beyond this point.

    While this was of course and unscientific estimation, it was the best or at least only available method of determining what percentage of grazing land was disturbed by erosion.

    When testing larger or smaller cattle grazing potential polygons, the resulting percentage was always within 1%, so some leeway is allowable.

    Polygon of Estimated Cattle Potential Layered Over Slope

    Grazing Potential Layered Over 2011 Classification




    Calculations Using Calculate Geometry Feature in ArcMap – Classification Layer Clipped to Cattle Potential Layer

    With a calculated grazing potential area of 633 acres,  56 acres was determined to be low erosion, meaning that the ground in that area was beginning to erode, but was not completely void of vegetation. 43 acres were classified as high erosion, meaning that no vegetation (only bare soil) was present. Totaling 99 acres, approximately 16% of all grazing potential has been eroded or is subject to high erosion.


    Change Detection

    The same process was conducted using the DOQ imagery for the year 2000, however, a combination of low spectral resolution (only three bands) as well as nearly 50% of the image covered in shadow, it was not possible to create a classification that could be accurately used in change detection from 2000 to 2011. Nuances detected between high and low erosion features in the 2011 Worldview imagery were not apparent in the 2000 imagery. It was only possible to create one class of erosion which would further distort and change detection analysis.

    2000 Classification of Erosion Features


    Sources of Error

    Inevitably, sources of error are part of the research process. It cannot be assumed that GPS coordinates or classifications are 100% accurate. This research was limited to available imagery and other source layers than have the potential to affect scale and output. In addition, for each layer used, accuracy must be multiplied to where subsequent accuracy is the product of the accuracy of each layer. So a layer with .80 accuracy over another layer of .90 accuracy, would result in a product with .72 accuracy. This must be taken into account in GIS. Some sources of error and bias in this research are as follows:

    •Time constraints
    •Lack of a subject matter expert (erosion specialist, soil specialist)
    •Lack of or access to temporal data
    •Processing time of UAV imagery – was not available for classification for 2013
    •Access to information about cattle locations other than anecdotal
    •Quality of older imagery, low spectral resolution compared to newer imagery
    •Better DEM would have resulted in accurate slope for each erosion feature


    Future Research 

    The inability to accurately detect change between 2000 and 2011, lends itself to future investigation of the study site with newer and high resolution imagery, or obtaining of historic imagery that contains either more spectral bands, less shadows or both. It is recommened, because erosion is a slow process, that imagery be examined, classified, and quantified every five years, in order to determine the extent of erosion and how erosion features have changed over time, either by growing, shrinking, or new formations.

    To ensure accuracy in classification, similar resolution imagery should be used for both time periods and ground truth GPs points must be collected to assess the accuracy of the classification.

    If it determined that erosion has accelerated based on a change detection analysis, erosion mitigation is recommended in the form of:

    • Paddock Rotation
    • Fencing of highly damaged areas
    • Creation of vegetation buffers along highly eroded areas
    • Consultation of erosion expert



    First and foremost, I’d like to extend a sincere thank you to Dr. Suzanne Wecshler, Dr. Christopher Lee, Dr. Matt Becker, Dr. Carl Lipo and Dr. Terry Hunt, the professors who made this research opportunity possible. 

    In addition, graduate students Paul Nesbit, Michelle Baroldi, Briton Voorhees, Emily Allen, Mike Ferris, and Scott Winslow were instrumental in making this program and my personal research a success. 

    Ted Ralsten, Dave Thielen, Cheryl Thielen, Dave Morgan, John Morgan, Williams Aerospace,  and an immense supporting cast of others also made this trip a success. Thank you to all of you who participated and offered your support in our work.