Data mining to identify optimal spatial aggregation scales and input features: Digital image classification with topographic LIDAR and LIDAR intensity returns.

摘要

Choices about modifiable spatial aggregation scales and optional input features are two important decision making areas in digital image classification. Limited research has shown that, with all other parameters being equal, both the inclusion of LIDAR-derived height and the spatial aggregation of input features (i.e. via multi-resolution image segmentation) can increase classification accuracy. Two data mining experiments examined the potential to optimize impervious surface classification accuracy in both these dimensions. Emerge 0.3 x 0.3 m digital aerial imagery acquired over Hilton Head Island, SC was segmented at seven alternate scales of spatial aggregation. Optional LIDAR-derived features, including height, intensity, scan angle, and density, were fused at each of the alternate scales. In the first experiment, all possible combinations of optional input features and scales of spatial aggregation were modeled using the C5.0 machine learning algorithm within the image as a whole. In the second experiment, all of the same combinations were modeled using C5.0 within each of ten ISODATA clusters (calculated using the baseline Emerge imagery). Through cross-validation trials and heuristic rules of thumb based on Ockham's Razor, an “optimal” model was selected for each of the ten spectral clusters, with all ten models being integrated for classification inference. A total of 3,498 reference pixels, acquired by stratified random sampling, formed the standard against which accuracy measures (e.g. K-hat, impervious producer's accuracy, etc.) were obtained. Results from the first experiment indicated that the use of an alternate spatial aggregation scale significantly improved impervious surface classification accuracy, thus confirming previous studies. Results from the second experiment showed that alternate spatial aggregation scales caused an increase in impervious producer's accuracy but a decrease in user's accuracy. Interestingly, neither experiment confirmed accuracy increases through data fusion. It is possible that the availability of alternate spatial aggregation scales overshadowed the benefit of the LIDAR-derived inputs. Future research should continue to examine potential advantages in approaching digital image classification as a data mining process. Spatial database management systems, largely through their adherence to the relational data model, promise to afford increased capability and efficiency in this area.