Modifications of imaging spectroscopy methods for increases spatial and temporal consistency: a case study of change in leafy spurge distribution between 1999 and 2001 in Theodore Roosevelt National Park, North Dakota

Abstract

From the early 1800's, when leafy spurge (Euphorbia esula L.) was first introduced into North America, it has spread throughout the northern Great Plains where it is currently a significant management concern. Accurate, rapid location is critical for economical control of this noxious weed, while repeatable measurements are necessary for successful temporal monitoring of infestations. Leafy spurge has been located using imaging spectroscopy (hyperspectral remote sensing) in the past, but the development and dissemination of standardized mapping procedures that produce consistent multi-temporal maps has been absent. In this study Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) data, collected in 1999 and 2001 over Theodore Roosevelt National Park, North Dakota, were used to locate leafy spurge. Mapping strategies were developed that improved the consistency of maps over time and space, increasing their suitability for monitoring change in spurge. Commercial software and algorithms were used to process the AVIRIS imagery. Published image-preprocessing methods, mapping algorithms, as well as field and image-derived spectral libraries were tested to determine which were most successful for consistent, repeatable mapping of leafy spurge over time and space. Best results were obtained using: 1) NDVI masking; 2) cross-track illumination correction; 3) an image-derived spectral library; and 4) the mixture-tuned matched filtering (MTMF) algorithm. The application of the algorithm was modified to standardize processing and to eliminate threshold decisions; the image-derived spectral library was also refined to eliminate additional variability in the spurge maps. The image-processing protocol appears to be consistent and reliable both temporally and spatially. Primary (spurge dominant), secondary (spurge non-dominant), and fraction maps were produced, as well as area-wide vegetation maps. Map accuracies were analyzed with three independent reference data sets (points, polygons, and grids) using standard confusion matrices as well as regression between field-measured percent spurge and image-derived matched filter ("abundance") scores. To accommodate offset between imagery and reference points, accuracies were recalculated after applying a majority filter, and after applying buffers ranging from 1-5 pixels wide around the classified data. The similar accuracy values of identical ground areas mapped from adjacent flight lines, as well as good visual correspondence, indicated that the mapping method produced consistent maps. Overall accuracy of area-wide spurge maps varied from 39% to 82%. Registration problems accounted for some of this variability. Some validation data and methods were more appropriate for assessing the detailed maps of small, fragmented patches of leafy spurge as well. Different mapping strategies produced very different spurge maps, another source of accuracy variation. Validation data for hyperspectral remote sensing should be carefully selected. Accuracy based on standard confusion matrices was sensitive to: 1) registration offsets between field and image locations; 2) modification of analytical methods; and 3) quality of the reference data. The vegetation maps of identical areas produced from adjacent flight lines varied when the difference in sensor viewing angles was at a maximum. Less spurge was classified at the highest viewing angles, regardless of viewing direction. Initial corrections for cross-track illumination differences were insufficient to entirely eliminate the effects of viewing geometry on the spurge classifications. Higher order corrections appeared to model the cross-track variation more closely. The first order correction may account for the discrepancies in the classification between high and low view angles. Overall accuracy was only 1% higher with the second and third order corrections; user's accuracy was 13% higher, however, and the visual correspondence between maps and reference data improved with the higher order corrections. Better modeling of cross-track variation improved the consistency of the spurge classifications where differences were related to sensor viewing angle. Regional patterns of change in spurge, as well as some localized trends, were evident based on differenced primary, present/absent maps. Post-classification change detection indicated a 41% decrease in leafy spurge within the park, and 36% in the surrounding National Grasslands, private range, and agricultural land between 1999 and 2001. The change in spurge fraction was calculated, but pixel-by-pixel change detection was difficult to interpret due to misalignment between multi-temporal images and slight shifts in the position of the flight lines between 1999 and 2001. The leafy spurge change maps produced from consistent multi-temporal maps will be valuable for land managers tasked with monitoring leafy spurge, as well as assessing the effect of different biological and chemical controls in order to optimize the management of leafy spurge on a regional scale.