This paper presents a two-stage approach to classify remotely sensed imagery. At the first stage, an unsupervised adaptive resonance theory neural network is invoked to perform a multi-spectral classification of satellite images. The outcome of this stage is a set of fine spectral classes. Since our goal is to obtain meaningful classes of the images the fine spec­tral classes are, at the second stage, reorganized into a small set of meaningful classes using a statistical clustering technique. The efficiency and accuracy of the proposed method is examined using a synthetic image. Afterwards, the proposed classifier is applied to a real SPOT image. Compared to another unsupervised classifier of ISOCLS, both classifiers have compatible processing times, while ours has achieved a better accuracy (96% for ours and 94% for ISOCLS).

Forest Remote Sensing consists of forest investigation、disaster estimation, and forest management etc.. One of the main purpose of remote sensing is to recognize the distribution of forest cover type. In this research, we propose a method to estimate the distribution on the remote sensing images with an eye to obtaining important information of forest resources for management purpose. In our work, French Satellite image (SPOT) is adopted as the material, and are integrated forest map and geographic information system as the overlay resampling. We devise the quadtree segmentation method segment the image of National Taiwan University Forest into uniform areas that satisfy the property of gray level homogeneity. T0 maximum likelihood classifier in conjunction with associating correlation value and weight loss are utilized to classify the distribution of forest. Experi­mental results reveal the feasibility of the proposed approach in forest cover classification.

In this study, a sequential sampling was employed to statistically select and evaluate cluster training blocks. Training blocks, each 3 by 3 kilometers in size, were selected from the entire Rocky Mountain National Park. Spectral, topographic, and ecologic criteria were used to assess the adequacy of the sampling. The criterion to determine whether to stop or to continue the sampling was based on the maximum difference between the sample cumulative distribution function (CDF) and the population CDF. The results showed that for this particular data set, twenty-three blocks would be adequate to represent the entire Rocky Mountain National Park.

In combining satellite images with other landuse data in geographic information systems, the images must be coincided with the map coordinate system. Thus, the orthographic correction for satellite images is a necessity. Traditionally, the photogrammetric technique reconstructs the orbit parameters using bundle adjustment for SPOT satellite images followed by an orthographic correction procedure. We, from another point of view, proposed a scheme which corrects the on-board data by using small amount of ground control points (GCPs). Then, a ray tracing is applied to perform the orthographic correction. The proposed scheme includes four major components: (1) using a few GCPs to first fit the orbit parameter with low order polynomials, (2) performing least squares filtering to fine tune the orbit parameter according to the fitting residuals in the first step, (3) determing the corresponding ground location for each image pixel by ray tracing method using the filtered orbit parameter and a digital terrain model, and (4) grid gray value resampling for output orthoimage. Experimental results indicate that the generated orthoimages attain an accuracy better than 0.6 pixels when 5 GCPs are used.

Mapping of surface reflectance has been considered as a primary task for optical remote sensing, since the remotely sensed data is dependent of sensor characteristics, solar-viewing geometry, the atmosphere and topography for a given canopy. In order to retrieve its intrinsic reflective characteristics that is reasonablly indepent of the above external factors, determination of reflectance by using radiometric correction of satellite image is considered to be indispensable. The purpose of the dissertation is radiometric correction of SPOT satellite imagery. Firstly, the variability of surface reflective characteristics due to solar-viewing geometry and the atmospheric haziness is studied. We propose the optimal view angle should be defined as the direction that could be used to estimate the albedo and is insensitive to cover types, solar zenith angle and the optical properties of the atmosphere. Secondly, the atmospheric correction model of satellite image in horizontal surface is proposed. The aerosol characteristics are iteratively retrieved from the image itself. Thirdly, urban bidi­rectional reflectance distribution function (BRDF) is retrieved by using multi-temporal multiangle SPOT imageries. Fourth, we suggested the normalized reflectance determined by its intrinsic BRDF is much suitable than bidirectional reflectance factor (BRF) as the spectral knowledge. Finally, a robust algorithm for retrieving aerosol optical depth from satellite image over a mountainous area is proposed. The results show that the optimal view angle determined by Kimes and Sellers (1985) would be insensitive to the aerosol optical depth. The RMSE of derived surface reflectance in the horizontal surface is about 0.01-0.02 over reflectance range 0.0-0.6. Spectral knowledge determined from normalized reflectance is much suitable than that proposed by Wharton (1987). Determination of reflectance image can reduce the topo­graphic effect of image mountainous area. Without atmospheric correction, terrain related classes of forest class would be classified from unsupervised classification, whereas they automatically merge to one class after atmospheric correction. Radiometric corrected image is more suitable for change detection analysis than raw image.

In recent years, automatic Photogrammetry becomes a major goal for everyone in the Photogrammetric industry. The technique of image processing has been used in aerial triangulation measurement for years, but it has not yet been put on production lines. Although digital photogrammetric workstations have been brought up to the market for some years, they still can not replace analytical plotters because of the limitations in there functions. Therefore mapping will still be done manually on the analytic plotters in the near future. Although fully automatic aerial triangulation measurement has not yet been realized, some kinds of semi-automatic solutions are still possible. In this paper we developed a method for semi-automatic aerial triangulation measurement. The major concern of this development is production, not sophisticated matching theories, i.e. to put the image matching technique actually on production lines to replace the expensive point transfer instru­ments and experienced operators so that it is not jmerely a research topic in the labora­tory﹒We hope that the results of the semi-automatic aerial triangulation obtained here can contribute to the automation of Photogrammetry in the practice.

In developing an automatic process to recover the relative orientation, tie points in each pair of stereo images should be identified. The least-square matching (LSM) has been widely used in the past few years but the fact that it tends to fail to match locally different images. Difficulties arise due to the influence of the image rotation and unequal expo­sures. The purpose of this study is to remedy the drawbacks of LSM. For identifying conju­gate image points, we detect the line segment of building roof using the Hough transform and compute the intersections of the lines to obtain the building corners .The result will be invariant with respect to the size and rotation of the image. In order to solve the occlusion problem, a simple criterion is proposed to match conjugate building corners. In order to test the practicability of this method, some experiments on a pair of real stereo image are processed with this method. The results are compared with the numbers obtained by using LSM. The comparisions show that this method is more robust to the effect of local changes in the images. We therefore conclude that this method can be a backup of LSM in image matching.

The booming applications of geographical information systems(GIS) demonstrates that we are in the age of information science. Traditional maps of many kind become spatial information systems in which one can dynamically query, display, analyze and manipulate spatial data. This paper tries to indicate the direction of photogrammetric field should be heading according to the developments of GIS and photogrammetric technologies, when the age of information science is coming. In addition, the conceptual and structural changes of photogrammetry will also be discussed. Along the track of automation in the field of photogrammetry, the goals and current condition of automation of (1) the generation of DTM and orthoimage, (2) automatic recovering image orientation, (3) automatic stereo mapping, and (4) image recognition and interpretation are examined. Photogrammetry, remote sensing and GIS have to be integrated in the future, and the key factors of integra­tion are the new concept of generating geographical data and the automation of photogrammetry.