Techniques of GIS data acquisition have been steadily improved over the past years, therefore, how to efficiently manage these volumes of complex data and effectively take advantage of their existence become a bottleneck for the future GIS development. Past research demonstrated that the accuracy of image classification can be improved by introducing ancillary information like landcover or landuse data. This research focuses on the data selection process in the GIS environment. We suggest a knowledge-based approach, which replaces the human users ‘interpretation and judgement with a knowledge base loaded with rules about human spatial and domain knowledge. The system is designed to understand users, requirement, infer their requirement with the understanding about GIS data, and then transform the requirement to a query executable with GIS DBMS. It is expected to solve the semantic gap about reality phenomena due to the different definitions from application domains(e.g., remote sensing) and existing GISs. This research started with a general analysis about building a KB based on the understanding of GIS data, we then use remote sensing image classification as examples to demonstrate how a KB can help choosing appropriate ancillary information to meet certain domain requirement. With careful selection and formalization of spatial and domain knowledge, it is optimistic to expect GISs to be able to meet more domain needs in the future.

In recent years use of multispectral satellite data to analyze plant cover and for devel-oping phenological models has been prosperous. However, a ground-based continued and detailed study of spectral measurements throughout the growing season would be useful by which not only spectral differences within and among seasons may be clarified but also the characterized wavelengths of crop species and variety may be identified. In the present study the reflectance spectra from 350 to 1100 nm were measured and the characterized wavelengths were recognized for the canopy cover of sweet potato (Ipomoea batatas cv. Tainung 57) grown in the field in 1995 and 1996. It was found the reflectance spectra of fresh canopy grown in different years were roughly similar, varied with environment and growth status. The ratio of reflected radiance in the UV region (740 nm) were fluctuated in 70-80%. Patterns of spectral irradiance and reflectance look similar to other green crops. The same 28 wavelengths were identified from the spectral radiance curves while only 6 characterized wavelengths were selected from reflectance spectra of sweet potato in both 1995 and 1996. Three wavelengths from spectral irradiance, 554, 688, and 754 nm, were further chosen for the calculations of vegetation indices. The vegetation index regress ion curves were curvilinear after planting, the increase of the curves stands for the building up of plant canopy and the curves fall off as plants aged.

The topic of this research is about the scheme of extracting roads automatically which at first transform the remote sensing images with special wavelet function, and linear features of original images are detected to attain the gradient magnitude and direction of each pixel. Secondly, according to the some roads properties on the images and getting an application from above data, we make the rules to detect seeds of road automatically. The third, we have a generic road model which consists of some geometric and photometric properties of roads, and a merit function is formulated by some constraints which embodies a notion of the generic road model. A merit function is the conditions of selecting road segments. Last, the seeds of road are used in beginning to search for possible roads which are according to above conditions. Then, we take advantage of dynamic programming to extract the roads from image.

Data integration for multiple images is getting more important as more resource satel­lites become operational. Image registration is an indispensable preprocessing in integrating multi-source images. The automated densitification technique for registration control points is getting mature. However, the establishment of initial correspondence still needs efforts to increase the degree of automation. The objective of this investigation is to correspond the conjugate feature pattern those are extracted from a reference images and its counter­part i.e., a sensed image. The major components of the proposed scheme include: (1)quan­tatively describing the feature patches using shape matrix, (2)similarity assessment between patch pairs, (3)considering the topological and geometrical relationship between candidate patch controids to filter out unreliable matching (4)establishing initial transformation according to the reliable centroid pairs, and (5)applying fine matching to generate registration control points followed by a local transformation procedure.