Object-based target search using remotely sensed data: A case study in detecting invasive exotic Australian Pine in south Florida

This study developed an object-based geographic image retrieval (GIR) approach for detecting invasive exotic Australian Pine in south Florida, USA. To filter out non-tree image objects, a hierarchical multi-resolution segmentation and filtering approach was first adopted to segment remote sensing images (DOQQs) into image objects (image regions) of irregular shape, compared to a regular square shape used in the literature. The study then computed object-level spectral, texture, and three-dimensional information for image object content representation using NDVI-based spectral, wavelet transform-based texture, variogram -based texture, and canopy surface height information. The effectiveness of content representation was evaluated using these different properties and their combinations in 10 sets of replica retrieval experiments with 5% random sample fractions of ground-truth identified Australian Pine image objects as query templates. The set of features providing the best fit was found to be a combination of canopy surface height and wavelet transform-based texture. These variables were selected for further tests to determine the similarity threshold beyond which retrieval is regarded as irrelevant. A series of regression tree models were built based on replica retrieval experiments with sample fractions of 1%, 5%, 10%, 15%, and 20%. The predicted results were analyzed to examine the sensitivity of retrieval performance (precision and recall) to the sample fraction and similarity threshold. A moderate retrieval performance was achieved in detecting Australian Pine in the study area. The study suggested that GIR with target search as its major objective by design could be an important supplement to image classification for invasive exotic plant species detection from remotely sensed images.