Digital Terrain Model with a 40m grid was adopted to analyze the landscape characteristics of volcanic morphology in Tatun Volcano Group of northern Taiwan. By scrutinizing the slope map derived from the DTM, 10 volcanic craters, 29 volcanic cones, some 40 lava lobes and more than 200 ogives are readily identifiable. The slope gradients of volcanic cones are in the range of 26°-48°. The gradients of the table surface of the lava lobes are mostly less than 10°, which implies high viscosity of lava flows in this area. It is also shown that the morphologic features of volcanoes erupted in earlier stage are not remained so well as those 10 craters formed in a relative late stage when visualized in 3D perspectives, derived from DTM.

The main purpose of this study is to evaluate approaches of automatic classification for landslide on a SPOT image. Additionally, 8 SPOT images that taken from 1996 to 2001 were used to understand the impact of the Chi-Chi Earthquake on the occurrences of landslide in the Choushuichi watershed. Compared to the interpretation of aerial photos, theoretical calculation shows that the average of producer's and user's accuracy in automatic classification of landslide can reach 70%. However, in comparison with theoretical calculation, 20-40% errors can be produced when we apply these techniques. In this study, the producer's accuracy can be increased to 50% when the ratio of band 3lband 2 is applied in vector image transformation. Besides, the produce's accuracy can further significantly be increased to 76% if the hill slope is used as a filter to delete the areas with similar characteristics as landslides when its slope is less than 10°• Although the accuracy of automatic classification of landslides in a SPOT image is still less than that identified from aerial photos, it is a fast, economic, and powerful tool to continuously monitor occurrences of landslides.

The bathymetry of coastal waters is of vital importance to coastal ocean engineering projects as well as the shipping safety. The conventional method for measuring water depths uses the shipboard echo sounder, which is accurate for point-measurement, but is a time-consuming and labor-intensive task. For periodic survey of bathymetry at synoptic scale, the remote sensing method may be a viable alternative. The goal of this research is to develop an indirect and more expedient method to map the bathymetry of coastal waters. There are several indirect methods for remotely sensed water depths. One is the optical bathymetry method, which is based on the principle that the total reflected energy of electromagnetic waves from a water column (including the water surface, the water body and the sea bottom) varies with the water depth. Actually, the reflectivity of visible light depends on the quality of the water column and the type of sea floor; the depth of highly turbid water is difficult to be determined by the optical bathymetry method. Presently, satellite image of QuickBird with 2.4m/pixel is one of high spatial resolution images. Avoiding the affection of variant quality of water column, One QuickBird image, covering the shallow water of Kending and Nanwan with light pollution, was chosen and analyzed to derive the bathymetry. This research finds that the optical bathymetry method is more feasible to bathymetric mapping in shallow water when the water depth is shallower than 12m.

Landuse classification is one of the major applications of satellite remote sensing. Many classification methods exist including nearest-to-means, maximum likelihood classification method, Bayes classification method, and neural network classification. Although these classification methods differ in their statistical and mathematical basis, the work of image classification is essentially a feature space partition problem, i.e., we look for algorithms that divide the n-dimensional feature space into k sub-spaces. Scattering of sample points of individual classes in the feature space may be anisotropically or randomly distributed; therefore, the spatial modeling algorithms such as kriging are ideal for estimation in feature space. In this study, we consider the scattering of each individual class as an indicator random field and developed a classification algorithm based on anisotropic indicator random field estimation. For each point X in the feature space, the proposed algorithm calculates the probabilities that X belongs to each individual classes using indicator kriging, and then assign X to the class with highest probability. The proposed indicator kriging classification algorithm has the following advantages: (1) it can deal with anisotropic problem in feature space, (2) it is a nonparametric method, and need not to know the type of probability distribution, and (3) it yields 100% classification accuracy for the training data. Preliminary results of landuse classification using SPOT images for a test area in Taipei show the indicator kriging classification algorithm outperformed the maximum likelihood.

This paper analyzes the error property of the so-called TOPSAR_DEMs, which are the digital elevation models (DEMs) determined by NASA/JPL airborne interferometric synthetic aperture radar (INSAR) technique in the mid Taiwan area using C-and/or L-bands during the PacRim 2000 Mission. After horizontal and vertical datum transformation operations, we estimate and analyze the systematic bias, random error, and blunder as well involved in the TOPSAR_DEMs. Also, the areas where airborne INSAR cannot determine the height are examined. Analysis results indicate that both TOPSAR_DEM and reference DEM have very similar terrain surface trend. However, height differences between both still have significant systematic bias. In general, the systematic bias is a non-linear function of point position. The RMS value of the random error components in the flat areas (about l~5m) is smaller than in the hilly areas (about 10~15m). In c.a.4746 km(superscript 2) of study areas with about 2.5 million DEM points, there is about 1~2% of DEM points with blunder. The percentage of points on which TOPSAR cannot determine the heights is about 6.37% in average and 18.22% in mountainous area, respectively.

The elements of concrete structures suffering from chemical reactions tend to be an expansion phenomenon, which leads to cracks. Non-contact measurement is employed to analysis the relationship between concrete expansion and crack width. This paper deals with semi-automatic crack feature extraction from digital multi-temporal close-range images with non-metric camera and processes the image adjustment with lens distortion parameters. Furthermore, the relationship between image and object space is obtained by 2-D projective transformation to do tilt rectification. Only few roughly manual seed points are required on crotches of crack features. Then, the shape and position may be illustrated with automatic feature extraction in two-dimensional image space. Because manual seed points are not exactly located on the crack features, two methods are proposed to search the geometric crack features, such as crotches. Wavelet transformation, clumpping, thinning and crossing algorithm are processed sequentialy to reach the semi-automatic feature extraciton purpose. Depending on these seed points, crack tracks and edges are determined by the method in this paper. Local image registration is utilized to raise relative accuracy and acquire the expansion of concrete cracks.