Harnessing the temporal dimension to improve object-based image analysis classification of wetlands

Research related to object-based image analysis has typically relied on data inputs that provide information on the spectral and spatial characteristics of objects, but the temporal domain is far less explored. For some objects, which are spectrally similar to other landscape features, their temporal pattern may be their sole defining characteristic. When multiple images are used in object-based image analysis, it is often constrained to a specific number of images which are selected because they cover the perceived range of temporal variability of the features of interest. Here, we provide a method to identify wetlands using a time series of Landsat imagery by building a Random Forest model using each image observation as an explanatory variable. We tested our approach in Douglas County, Washington, USA. Our approach exploiting the temporal domain classified wetlands with a high level of accuracy and reduced the number of spectrally similar false positives. We explored how sampling design (i.e., random, stratified, purposive) and temporal resolution (i.e., number of image observations) affected classification accuracy. We found that sampling design introduced bias in different ways, but did not have a substantial impact on overall accuracy. We also found that a higher number of image observations up to a point improved classification accuracy dependent on the selection of images used in the model. While time series analysis has been part of pixel-based remote sensing for many decades, with improved computer processing and increased availability of time series datasets (e.g., Landsat archive), it is now much easier to incorporate time series into object-based image analysis classification.