The use of remote sensing for weed mapping has well-known tradeoffs between spectral and spatial resolution using current remote sensing platforms. Often restricted spectral capabilities limit the ability to discriminate particular plant types that are often spectrally similar, for example, similar shades of ‘green’, particularly in pastoral settings. Regardless of the spectral capabilities of any sensor, the discrimination of a single plant species remains challenging. Linking particular plant traits to a specific parameter of an image can assist towards discrimination. For example, targeting temporal patterns related to a plant’s growth cycle, or, similarly, acquiring imagery at a time when the background can be separated from the target plant. In addition, the use of an object-based image analysis (OBIA) approach to image classification, instead of the more typical pixel-based approach, has been shown to make significant traction with mapping individual species. OBIA seeks to identify spatially and spectrally homogeneous ‘objects’ which are created by grouping pixels of similar traits together, then allows a combination of ancillary data (morphology, contextual) to be incorporated to assist the classification results. Archived World-View 2 satellite imagery was selected at a time to maximise phenological contrast between the plant and the background pastoral setting. Plots of African Lovegrass were established for four density categories across the study area. The classification approach, using eCognition software, included manual threshold tests with subsequent segmentation. This resulted in a hierarchical classification rules set to separate the target African lovegrass from trees and other landscape features present. A standard supervised classification algorithm was applied, with validation plots used to assess classification success for each density class.
Results/Conclusions
Independent validation using the established plots located across the study site and for four density classes, indicate the classification was 76% accurate across all density categories inclusive. Based on the same validation dataset, high and low density distributions were well distinguished, but the medium categories not well discriminated. Accuracy for an aggregated density class was highest, followed by 52% for the lowest category. Confusion was found with the distribution of a native poa grass, which existing techniques were unable to resolve. These results indicate that the OBIA approach, based upon the characteristics of the plant architecture and assisted by a tonal difference in the pastoral background, provides a much more detailed and accuracy representation of the distribution of African lovegrass.