This study applied satellite SAR images from ERS-1 and ERS-2 for water line extraction and change detection along the west costal area, covering from Kuan-in, Taoyuan to Tung-Kang, Pintung a total of about 400 km long. The images were acquired in the years of 1996 and 2005 with 10 years time span with 23 image tracks (total 92 image sets of 100 km by 100 km) In particular, the environmental sensitive areas of Fang-Yuan, Hai-Feng-Dao Chou, and Wai-San-Ting Chou, were chosen for detailed analysis. In addition to SAR images, we also incorporated optical satellite images, water depth map from Water Resource Bureau, and tidal records from tidal stations operated by Central Weather Bureau. This is important to ensure all the water line can be calibrated to the same tidal level. From the results, it is found that Wai-San-Ting Chou moved southward about 700 meters during 10 years period and it was shrunk to just one-third of its size in 1996. It is also found that the surroundings costal waters of Wai-San-Ting Chou and estuary of Pei-Kang River moved inward to west costal. The water channel was becoming more deposited as a result. Finally, Hai-Feng-Dao Chou itself moved southward about 1.5 km although its size remained quite the same during 1996 and 2005, and it also presented a clear tendency of inward moving. From the results we reported here, it is concluded that satellite remote sensing by SAR, aided by ground tidal data, bathematric map, and optical images, provides an effective and efficient tool for costal water changes in terms of large area coverage and long time span.

This study investigates the application of the Rough Set Theory 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. Gaussian Maximum Likelihood Classification and Back-Propagation neural network are used for comparison. The overall accuracy for Rough Set Theory is 86.947% for Chiayi and 81.44% for Hsinchu. The kappa index is 0.73826 for Chiayi and 0.61448 for Hsinchu. In terms of the classification accuracy, Rough Set Theory is shown to be better than Gaussian Maximum Likelihood Classification but inferior to Back-Propagation neural network for Chiayi and Hsinchu area.

Synthetic aperture radar (SAR) which adopts a focusing process to improve the spatial resolution in azimuth direction of the real aperture radar (RAR) brings the use of radar system in remote sensing applications into practical. Traditionally, SAR platform moving with a constant speed along azimuth direction results in a Doppler frequency shift relative to a target on the ground. A time domain Fourier Transform is then applied so that the spatial resolution in azimuth direction is improved. The Remote Sensing Image Processing Lab at the Department of Electronic Engineering in the National United University cooperated with the Microwave Remote Sensing Lab at the Center for Space and Remote Sensing Research in the National Central University to design a ground based SAR remote sensing system. In the previous studies, which are performed by the research group in the National Central University, the step frequency continuous wave technique was adopted to modulate the transmitted electromagnetic waves so that Inverse Discrete Fourier Transform could be applied to the received radar echoes and image focusing in the range direction were achieved. However, without Doppler frequency, image focusing in the azimuth direction can not be realized by making use of the conventional approach because that this system has no moving mechanism. Therefore, this paper aims to propose an effective approach to conduct SAR image focusing process in azimuth direction for the ground based SAR remote sensing system. Because the architecture of this system is similar to that of synthetic array, this paper propose to apply synthetic array technique to transfer time domain Fourier Transform into spatial domain Fourier Transform with a virtual speed of the system so as to attain SAR image focusing in azimuth direction. At first, in this paper, a computer simulation according to the system parameters designed in the precious studies is done to verify the feasibility of this approach. Simulation result shows that this approach fails to perform SAR image focusing in azimuth direction. Then, this approach is applied to the real radar image acquired by the ground based SAR system. Experimental result indicates that this approach is still unable to achieve SAR image focusing in azimuth direction. Therefore, this paper attempts to re-design the system parameters. It is concluded that the failure of the proposed approach results from the sampling rates of the system which are over-sampled in the range direction and under-sampled in the azimuth direction. Finally, in this paper, a computer simulation according to the modified system parameters is done again to verify the feasibility of this approach. Simulation result shows that this approach does perform SAR image focusing in azimuth direction effectively. In conclusion, this paper has successfully overcome the difficulty of the SAR image focusing process in azimuth direction for a ground based SAR remote sensing system with no moving mechanism. The synthetic array technique is applied to transfer time domain Fourier Transform into spatial domain Fourier Transform with a virtual speed of the system so that the conventional SAR image focusing process in azimuth direction is still applicable without any extra efforts to be paid.

It is of considerable interest to find an optimal classifier that has been discussed in the field of spatial information. In essence, there are many image classification methods, e.g. Maximum likelihood (MLH), K-nearest…. However, most of the linear classifiers are not capable of handling the complexity and the huge amount of the very high resolution image data. Thus, Support Vector Machine (SVM) is one of the powerful non-linear data mining classifier which is adopting to resolve the classification problems in this study. The high resolution QuickBird satellite images with additional texture information are the study material. The MLH method is used as a parallel study for the comparison on overall accuracy. The contribution of this study found that the overall accuracy of SVM is stable than that of MLH. More specifically, the overall accuracy of SVM is 87.3% (Kappa=0.8416) which is apparently higher than that of MLH (overall accuracy of SVM is 83.73% with Kappa= 0.7994). On the other hand, SVM can display better classification outcomes in the image pattern of “paddy rice” than that of MLH. In fact, the additional texture information can deal with noise effectively. The study find out that SVM can potentially perform higher image classification ability than the conventional MLH method.

Each element in the production chain of airborne lidar operation will inevitably influence the final quality of the product. By selecting different parameter values in each check item, and adopting different procedures, one can define the quality level of the product required for the application. This study investigates the production flow of airborne lidar for the digital elevation model and digital surface model. The procedures and different phases of the operation are analyzed. A series of check items for quality assurance and their schedule are proposed. Three phases are identified, mission planning, data collection, and data processing. In the data processing, it is further divided into point cloud generation, classification, and editing.

The 3D road modeling becomes an important task in the geoinformatic realm. The proposed scheme integrates LIDAR data and road networks to automatically reconstruct 3D road models. The whole process includes three parts: initial modeling, precise modeling, and validation. First, the elevation of the model surface is initially derived from LIDAR data. Then, the vertical alignment of each road segment is modified by line fitting. In the second part, the continuity in height, slope, and slope difference of 3D road models are fine-tuned with least squares adjustment. Finally, the validation is done to examine the consistency between point clouds and the generated model, the difference between the reference model and the generated model, and the rationality of the generated road profiles.