Object-based methods for landscape delineation with high spatial resolution orthophotography
Abstract
High spatial resolution remotely sensed imagery has been widely used for land-use and land-cover (LULC) mapping. In the meantime detailed and high accuracy land-cover maps are also increasingly desired by land-use researchers and planners. Continental-scaled land cover maps with moderate spatial resolution (30 meter) such as National Land Cover Data Sets (NLCD) and National Agricultural Statistics Service (NASS) Cropland Data Layer (CDL) were employed as primary LULC data and reference maps. However, the differences in land cover nomenclature of NLCD have been a recurring problem. The relatively coarser resolution of NLCD and NASS CDL also reduces the comparability to classification results of high resolution aerial photography and satellite imagery. Scientific research has revealed a strong demand of new methods for accurate and detailed LULC mapping, especially for the large-scale, complicated and spectrally heterogeneous landscapes. This study examined how to combine object-based and pixel-based approaches for the identification of land-cover types from 1-meter resolution aerial orthophotography. Specifically, this study included four major procedures: (1) pre-processing the aerial photography by using convolution filters, IHS, and principal component transformation for spatial and spectral information enhancement; (2) addressing building extraction in urban areas from high spatial resolution imagery by defining object-based decision rules to reconfigure the morphology of the initial approximation of building locations and shapes, derived from a 5-foot Digital Elevation Model (DEM) and Digital Surface Model (DSM); (3) selecting suitable features for image object discrimination according to their spectral, contextual, and spatial information, as well as their morphological characteristics and; and (4) developing object-based methodologies, that employ elevation information as ancillary data for building and vegetation delineation in urban landscapes in Indiana State, as well as for land-cover mapping of the entire Tippecanoe County in Indiana State. Hierarchical classification schemes were developed to discriminate image objects for specific classes. The advantages of different segmentation algorithms were combined to identify feature similarities, both among image objects and with their neighbors. Image growth took place if neighbors of image objects satisfied a series of criteria given a set of class-defined object features. This study also proposed an object-oriented method to extract road information from the NLCD 2001 and to reclassify urban-related fields from the NLCD 1992 and 2001 data sets. The comparability between NLCD 1992 and NLCD 2001 in urban areas was tremendously increased after the recoding and modification of spectral classes. The results of the urban land cover change analysis from the modified datasets were more logical. Classification results from the combined methods were compared with traditional pixel-based classification methods at each study area. Both statistical and visual comparison demonstrated that the object-oriented decision rules, as well as their combination with the pixel-based methods, were more effective for land-cover mapping, and achieved much higher classification accuracies than traditional classification methods.
Degree
Ph.D.
Advisors
Shao, Purdue University.
Subject Area
Geographic information science|Forestry|Remote sensing
Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server.