Download PDF Download Document

Final Report on High Resolution Change Detection Project

Category: Habitat - Research

Date Published: August 09, 2011

Number of Pages: 13

Author(s): Ken Pierce


The High Resolution Change Detection project was designed to explore the feasibility of using high-resolution aerial imagery (1 m resolution National Agriculture Inventory Program data) to detect changes in land-cover from 2006 to 2009 in selected WRIAs of the Puget Sound region. Early in the project we defined land cover change as the transition from forest landcover to a human dominated landcover, i.e., developed areas. High resolution imagery is preferable to medium resolution imagery (30m Landsat pixels) for mapping change because important areas such as riparian vegetation and marine shorelines cannot be accurately delineated at the resolution of 30-m pixels. However, high resolution imagery is difficult to work with in an automated manner due to the volume of information it contains (large file size per unit ground area), the effect of solar position on illumination (shadows) and high local variability of imagery within single land cover classes (e.g., forest of different age can look very different in aerial photography). With the help of relatively new software and computing power, we used a combination of supervised classification to isolate shadows and areas devoid of vegetation, image segmentation to create homogenous regions for statistical analysis, and high-efficiency methods for analyst review of sampled change locations.

The "high-efficiency methods for analyst review" were crucial to the success of this project as over 30,000 individual polygons were reviewed in the accuracy assessment portion of the study. Because change occurs at very low rates on an annual basis, finding change locations against an enormous backdrop of nonchanging data is an arduous task. For example, WRIA 7 (the Snohomish River Basin), which is 480,000 ha in size, is represented by 4.8 billion pixels and 514,293 polygons or cover polygons after analysis. The initial predictive model that we developed labeled only 5,121 polygons as having changed between 2006 and 2009. That is, the model successfully separated the initial 514,293 polygons into the 1% predicted as change and 99% as non-change. To improve accuracy we developed high-efficiency methods for analyst review that automates the individual review of all computer indentified change polygons (5,121 for Snohomish WRIA). That is, all polygons predicted as change are qued up and individually displayed for both time periods, which allows the analyst to determine if the change is real. Of the 5121 predicted change polygons, the review process verified that 3165 polygons had actually changed and 1956 did not change. Of the 3,165 change locations mapped between 2006 and 2009, 2,670 (84%) were smaller than 1 ha and thus undetectable by coarser imagery (Landsat). By reviewing all predicted change polygons, we effectively eliminated commission error such that every mapped polygon has been verified as change, i.e., 50% of the vegetation in the polygon (estimated visually) was changed into a developed land cover class. Errors of omission were assessed by randomly selecting a sample of polygons predicted as non-change and subjecting those to analyst review. In WRIA 7 we reviewed 3,834 polygons and found 29 (0.7% of polygons) that actually changed. By expanding the percentage of the changed area that the model missed to the entire WRIA, we estimated that 4,848 acres of change took place beyond the 10,678 acres we mapped. Of the mapped change areas, 2.9 acres were within 60.9 m (200-ft) of the Washington Department of Ecology’s (DOE) marine shoreline delineation and 73 acres were within 100 m of a WDFW documented fish-bearing streams.

The final product shows that 100% of the areas mapped as change exhibited real change over the time period examined here. Of the areas not mapped as change, we estimate that 1 out of every 246 acres mapped as no-change was actually change. These results were consistent over the four WRIAs that we analyzed in this project.