| Despite advances in remote sensing technology and computational power, automatic land use classification has remained a difficult process, and successful approaches require the incorporation of ancillary GIS data. This study seeks to address the issues which have held back land use classifications making exclusive use of remotely sensed data by using buildings as an alternative to parcels for classification and using ontology to identify existing classification problems and delineate avenues for improving classification methodology. LIDAR point cloud data from the City of Austin is used exclusively for the initial building extraction process, and an impervious surface classification layer derived from DOQQ aerial photography is joined with this layer for the building classification process. The building extraction employs segmentation in Definiens Professional and Neural Network classification in MATLAB and achieves a segment level of accuracy of 92% and kappa of 0.7, with an adjusted building detection rate of 74%, after the exclusion of buildings under 50 square meters in area. A land-use classification scheme using classes derived from the City of Austin's Urban Land Use codes is applied to ground truth buildings and results in a classification accuracy of 80.4% with a kappa of 0.35. When this classification methodology is applied to the extracted buildings the result is overall accuracy is 78.6% with a kappa of 0.33. Further avenues for improving these results are suggested by the ontology and an exploration of the errors. |
