Methodology for updating terrain object data from remote sensing data : the application of Landsat TM data with respect to agricultural fields

<p>This thesis describes some methods for updating the thematic and geometrical data of terrain objects that are contained in a Geographic Information System (GIS). The updating is based on the application of digital interpretation techniques on high resolution satellite data. The potential for updating terrain object data from remote sensing (RS) data is largely determined by two factors:<br/>(i) The thematic and geometrical characteristics of terrain objects that can be extracted from RS data depend on their relationship with the spectral and spatial information present in the RS data applied.<br/>(ii) Furthermore, digital interpretation techniques cannot directly yield the information required due to the complexity of real world images.<p>The idea underlying this thesis is that information extraction from RS data (based on digital interpretation techniques) can be improved and optimized by using ancillary data and knowledge about the static and dynamic properties of the terrain objects of interest. Such an approach requires integrated processing of different types of data and knowledge. Important aspects of an integrated approach are the integration level (pixel-based versus object-based data integration), the spatial aspects (co-registration and vector/raster integration) and error propagation.<p>The terrain objects of interest in this thesis are agricultural fields. A data set was established consisting of a Landsat TM image and (multi-temporal) data on the crop type and field geometry of agricultural fields in a polder area in the Netherlands. Three updating methods by means of an integrated approach were developed and tested with the available data.<p>Knowledge about crop rotations was formalized by means of transition matrices which store transition probabilities. The transition probabilities, corresponding to the crop type grown in the preceding growing season, were used as (conditional) a-priori probabilities in a pixel-based maximum likelihood classification. For the test area, overall classification accuracy increased with 2 % to 17 % depending on the spectral separability and the set of a-priori probabilities applied.<p>Object-based classification was used to determine the crop type of agricultural fields for which the geometry was already contained in a GIS. In the same process the field geometry is used to derive a reliable classification result by excluding boundary pixels which are most often mixed pixels. For 92 % of the fields in the test area a correct crop type was determined.<p>An integrated segmentation and classification method was applied to determine both the field geometry and crop type of agricultural fields. The results of an edge detection on the TM image were integrated with the fixed boundaries contained in the GIS by using knowledge about the aggregation structure and shape of the fields. The resulting field geometry corresponded for 87 % with field geometry derived from visual interpretation of the TM image.<p>Several aspects of data integration were identified. Object-based data integration, which means that knowledge is formulated in terms of terrain objects that have geometrical and thematic properties, is required for updating. A large number of representations are possible for formalizing knowledge; different methods for representation were used in this thesis: transition matrices, statistical functions and geometrical functions.<p>For the integration of the vector-structured terrain object data with the rasterstructured RS data two approaches can be adopted: data conversion (vector-to-raster and vice versa) or 'direct integration'. The last approach was used to identify the raster elements that are located within a polygon in object-based classification.<p>The case studies showed that terrain object data can be updated based on digital interpretation of remote sensing data and that the ancillary data and knowledge are effective for improving and optimizing the information extraction. Nevertheless, the information (type and quality) that can be extracted still largely depends on the (spectral and spatial) relationship between the terrain objects of interest and the RS data applied