Landsat-based monitoring of annual wetland change in the Willamette Valley of Oregon, USA from 1972 to 2012

Abstract

In Oregon’s Willamette Valley, remaining wetlands are at high risk to loss and degradation from agricultural activity and urbanization. With an increased need for fine temporal-scale monitoring of sensitive wetlands, we used annual Landsat MSS and TM/ETM+ images from 1972 to 2012 to manually interpret loss, gain, and type conversion of wetland area in the floodplain of the Willamette River. By creating Tasseled Cap Brightness, Greenness, and Wetness indices for MSS data that visually match TM/ETM+ Tasseled Cap images, we were able to construct a complete and consistent, annual time series and utilize the entire Landsat archive. With an extended time series we were also able to compare annual trends of net change in wetland area before and after the no-net-loss policy established under Section 404 of the Clean Water Act in 1990 using a Theil-Sen Slope estimate analysis. Vegetated wetlands experienced a 314 ha net loss of wetland area and non-vegetated wetlands experienced a 393 ha net gain, indicating higher functioning wetlands were replaced in area by non-vegetated wetland habitats such as agricultural and quarry ponds. The majority of both gain and loss in the study area was attributed to gains and losses of agricultural land. After 1990 policy implementations, the rate of wetland area lost slowed for some wetland categories and reversed into trends of gain in wetland area for others, perhaps representative of the success of increased regulations. Overall accuracy of land use classification through manual interpretation was at 80 %. This accuracy increased to 91.1 % when land use classes were aggregated to either wetland or upland categories, indicating that our methodology was more accurate at distinguishing between general upland and wetland than finer categorical classes.

Appendices

Appendices

Appendix 1

Here we describe in detail our methods for converting tasseled cap coefficients for MSS data.

We used TC coefficients for TM reflectance factor data (Crist and Cicone 1984) for both the TM and ETM+ data (Cohen et al. 2003) after first deriving surface reflectance for each TM and ETM+ image using LEDAPS (Masek et al. 2006). There is no similar set of coefficients for MSS reflectance factor data, requiring a calibration of the MSS data to the TM/ETM+ time series. To integrate MSS data with TM/ETM+, we used random forest regression (Breiman 2001) to predict TM TC brightness, greenness, and wetness (y-variables) directly from the MSS band 1–4 spectral data (x-variables) with coincidently acquired Landsat 4 MSS and Landsat 5 TM images from 1987 (Main-Knorn et al. 2013). The derivation of pseudo-wetness for MSS data is new, and was chosen over the derivation of TC angle (Gómez et al. 2012; Pflugmacher et al. 2012) for our study because we wanted to maintain the wetness index from TM/ETM+ for our time series. When evaluating the goodness of fit for the predicted TC coefficients, we examined scatterplots for general correlation, but put most emphasis on image appearance consistency both within the MSS images themselves and compared to the TM/ETM+ images through time. Although imperfect, the results were more than sufficient for consistent visual interpretation across the time series. Once the time series was compiled, we applied a two-standard deviation stretch to each image to enhance visual image contrast.

Appendix 2

Here we describe our workflow and methods to manually interpret annual wetland losses and gains associated with land use change.

During interpretations of a given subarea, three software applications were open: TimeSync (Cohen et al. 2010), ArcMap, and Google Earth. TimeSync and ArcMap were complementary in their use. Within TimeSync the 41 annual Landsat TC image “chips” for the subarea were simultaneously displayed, and was thus best for interpretation of wetland gain and loss and land use change (given ArcMap does not have this functionality). Additionally, within ArcMap were two data layers that assisted in the interpretation process by spatially limiting the interpretations, including the 1 m lidar inundation map and a shapefile of the main stem of the Willamette River and its tributaries within the floodplain created from the Pacific Northwest Hydrogaphy Framework’s Oregon Water Courses shapefile. Within Google Earth were historic image snapshots of each subarea and the inundation map (defined by kmz files), which assisted in confirming and identifying wetland gain and loss and land use change within the relevant interpretation area. The availability and temporal resolution of historical imagery in Google Earth varied depending on location.

We interpreted the study area one subarea at a time moving east to west from south to north. Wetland gain was interpreted as new wetland land use derived from a different land use category and loss was any wetland area that was converted to a different land use during the time period defined by the Landsat time series. Wetland-to-wetland conversion was also interpreted and classified as wetland area that was converted to a different wetland type (e.g. riparian to emergent). When interpreting change polygons, three attributes were selected from drop down menus in the attribute table: date when change was first detected, starting land use, and ending land use.

Area and perimeter of each drawn polygon were calculated within ArcMap. Our workflow required an average of 39 min per subarea, ranging between 10 and 90 min, with the biggest factors being the geometric area of the floodplain located within the subarea and the number and complexity of wetland losses and gains.