This paper proposes an algorithm for reconstruction of an object surface covered with 3D points. It utilizes the 2D Daubechies scaling functions of 3rd order, which can describe fractal geometry, to derive the observation equation for each point. The linear system is then solved by the least-squares adjustment (LSA) and the reconstructed surface can then be generated. To overcome the ill-posed problem which often emerges in LSA, we employ a from-coarse-to-fine strategy and use the pseudo observations on dyadic points, called POI (Pseudo Observations by Interpolation) and PHO (Pseudo Height Observations). Moreover, a full-automated weighting model is proposed in order to eliminate the so-called Gibbs effect. It reduces the weights of the points whose absolute residuals are larger than twice the a priori height accuracy of the LIDAR point. Some tests are done by using airborne LIDAR points. They verify that the artifacts caused by the Gibbs phenomenon can be eliminated to the large extent by combining the pseudo observations and the weighting model. While the dyadic points have approximately the point interval of LIDAR points, the a posteriori standard deviations of unit weight in our tests are about ±20~23cm which are all to the extents of the a priori height accuracy, ±25cm.

Usually, a standard least-squares image-matching functional model has radiometric shift and drift parameters, and geometric affinity parameters. This paper is focused to improve on a conventional stochastic modeling. Single-difference gray-levels are no longer dealt with to be independent and identically distributed. Scaling variance and covariance components are associated with some processed image segments. The estimated variance and covariance components are then used to form a new measurement covariance matrix, leading to iteratively adjusted weights until a steady parameter state is achieved. In theory, the proposed Blue-estimator is akin to the best invariant quadratic unbiased estimator. In practice, two Radarsat-1 synthetic aperture radar image scenes were made available to study an image-matching applicability to features such as an angular section of a pond and an intersection of roads. As a result, both the line and sample coordinates can be determined more accurately.

This study investigates the application of the Support Vector Machine (SVM) for image classification. The images used for the experiment include multi-temporal Formosat-2 images of the Chiayi area and multi-temporal SPOT images of the Hsinchu area. There are a number of kernel functions to be selected with SVM. In experiment, Gaussian Maximum Likelihood Classification and Radial Basis Function (RBF) neural network are used for comparison. The Polynomial Kernel Function is the best for Chiayi and RBF is the best for Hsinchu. The overall accuracy is 89.830% for Chiayi and 84.989% for Hsinchu. The kappa index is 0.79303 for Chiayi and 0.68269 for Hsinchu. In terms of the classification accuracy, Support Vector Machine is shown to be better than Gaussian Maximum Likelihood Classification and Radial Basis Function (RBF) neural network.

In this paper, modified Bagging and Boosting voting methods were proposed in the multiple classifiers system for terrain classification of remote sensing images. The improvement is achieved by introducing a confidence index to reduce the ambiguities among the targets being classified. Performance of the proposed multiple classifiers system was tested using fused radar and optical images. Experimental results show that the classifier is able to substantially improve the classification accuracy.

Due to its standardization and simplicity, rational function model has been widely used in the geometric correction for high resolution satellite images. The rational function model includes direct and indirect methods. The direct method uses the ground control points to determine the rational polynomial coefficients directly. The indirect one uses on-board data to establish the rational polynomial coefficients. Then, a small number of ground control points are applied to refine the model. This investigation compares the geometric precision of direct and indirect rational function model for high resolution satellite images. Test data includes an IKONOS satellite image with 1 meter resolution. Experimental results indicate that, direct method needs more control point when compared to the indirect method. Both of the direct and indirect methods may reach sub-pixel accuracy when 21 and 6 control points are applied, respectively.

The percentage of cloud coverage is generally considered the most important factor to affect the availability of optical satellite images for land resource applications. Consequently, an effective cloud assessment process can not only offer general users a guideline for image browsing, but also provide valuable information for orbit planning of ground receiving station. At present, the cloud coverage of Formosat-2 data is visually examined and manually estimated by the operators on the QUICK-LOOK images. The QUICK-LOOK images basically are the reduced images created by satellite data receiving system for browsing purpose. As to the current estimation of cloud coverage, every QUICK-LOOK image is divided into 8 portions and the cloud coverage for each portion is indicated with a cloud index with 5 grades. Apparently, the visual and manual cloud assessment method is labor-intensive, time-consuming, and dependent on the operator’s experience. Furthermore, the existing 5-grade cloud index is inadequate to describe the accurate percentage of cloud coverage. Thus, there is an urgent need for developing an automatic approach for cloud assessment. In this study, a two-stage scheme based on image segmentation and fractal theory is used to automatically extract cloud pixels from QUICK-LOOK images. The experimental results indicate that the cloud areas extracted by proposed scheme are very similar to the manual interpretation. In addition, when the automatically extracted cloud coverage is transferred to the existing cloud index, a good agreement between our automatic approach and current manual estimation is achieved.