Because of its all-weather and all-time characteristics, SAR image is particularly effective to monitor disaster events. When a disaster occurs, the image acquired from a SAR sensor changes dramatically. As a result, damage estimation for natural and human-made disasters from SAR images can be achieved by applying a change detection technique. However, it is very difficult to detect changes in SAR images because of the speckle noise in nature. Theoretically, SAR signal can be characterized as a chaotic phenomenon because that the scattering signals within a resolution cell are summed up coherently. Accordingly, SAR signal can be represented by a spatial chaotic model and characterized by its fractal dimension. In this paper, based on the spatial chaotic model, a simplified SAR image change detection procedure is proposed. The proposed method is then applied to estimate the flood-damage area caused by flooding events. Experimental results reveal that the proposed method is an effective and efficient tool for damage estimation from SAR images.

A k-nearest-neighbor classifier expects the class conditional probabilities to be locally constant. In this paper, we use the local separability based on nonparametric weighted feature extraction criterion to establish an effective metric for computing a new neighborhood. For each test pattern, the modified neighborhood shrinks in the direction with high separability around this pattern and extends further in the other direction. This new neighborhood can often provide improvement in classification performance. Therefore, any neighborhood-based classifier can be employed by using the modified neighborhood. Then the class conditional probabilities tend to be more homogeneous in the modified neighborhood.

For hyperspectral imagery, greedy modular eigenspaces (GME) has been developed by clustering highly correlated hyperspectral bands into a smaller subset of band modular based on greedy algorithm. In this paper, we introduce a simulated annealing mechanism, which is a typical heuristic method in the optimization process, in place of greedy paradigm as adopted in GME approach for hyperspectral imagery. This paper proposes a new method called the two-dimension simulated annealing band selection (2DSABS) for hyperspectral feature extraction. The 2DSABS selects the sets of non-correlated hyperspectral bands for hyperspectral images based on simulated annealing (SA) algorithm while utilizing the inherent separability of different classes in hyperspectral images to reduce dimensionality and further to effectively generate a unique clustered eigenspace (CE) feature. The proposed 2DSABS features can (1) avoid the bias problems of transforming the information into linear combinations of bands as does the traditional principal components analysis (PCA); (2) select each band by a simple logical operation, called CE feature scale uniformity transformation (CE/FSUT), to include different classes into the most common feature clustered subset of bands; (3) provide a fast procedure to simultaneously select the most significant features, and therefore dramatically improve the eigen-decomposition computational complexity. The experimental results show that the proposed 2DSABS approach is effective and can be used as an alternative to the existing feature extraction algorithms.

In Taiwan, approximately 73% of the total water resource is been used for agricultural purposes. Among the 73% of agricultural water usage, more than half of the water is used for the paddy rice irrigation. As a result, it is important to evaluate the evapotranspiration characteristic of the paddy rice fields. Traditionally, evapotranspiration is measured by using basin or empirical function with surface meteorological data. However, these methods only measure points and may not effectively address the evapotranspiration over a large spatial domain. To resolve such issue, using remote sensing technology to estimate the evapotranspiration becomes an interest field, as the remote sensing technology has the advantage of acquiring large surface data instantaneously and the data can be used to retrieve surface parameters, as well as thermal prosperities. The research presented here is to use satellite images in conjunction with the in-situ surface meteorological measurement, such as temperature and wind velocity, to estimate the surface heat fluxes that include soil heat flux, sensible heat flux, and latent heat flux. This method is based on surface energy balance to estimate the surface heat fluxes. Once the heat fluxes has been determined, the evaporation then can be derived from radiation and evaporation controlled lines. The satellite data used for this study are the MODIS level 1D image obtaining from CSRSR (Center of Space and Remote Sensing Research), National Central University. In additional, derivative products of the MODIS sensor from USGS/NASA were also obtained for comparative study. The result shows that using MODIS level 1D data as input parameters produces better result and is closer to the ground measurement. Hence, the MODIS level 1D data were used to study the surface heat fluxes of two rice agricultural cycles in 2006. With the correlation coefficient over 0.60, the result shows a potential in estimating the surface heat fluxes of paddy rice fields in Chiayi area, and can be helpful to control the water used to irrigate the paddy rice.

Linear spectral mixture analysis (LSMA) has been widely used in remote sensing applications, and the Least Squares (LS) approach is one of the most effective methods for solving LSMA problem. Since the noise in LSMA from each band may not be independent and identical distributed (i.i.d.), it has been proven mathematically that the Noise Whitened Least Squares (NWLS) will outperform the original LS by making the noise i.i.d. with the noise whitening process. But how to estimate the noise covariance matrix is remain a challenge problem. Many methods have been proposed in the past including spatial high-pass filter, frequency domain high-pass filter, orthogonal subspace projection, principal component analysis and Fisher’s Linear Discriminant Analysis (Fisher’s LDA) based approach. They all perform well for Gaussian noise but encounter problems when the noise is ill distributed. In this study, we adopt Nonparametric Weighted Feature Extraction (NWFE) to estimate the noise distribution and compare the results. Furthermore, we also apply the weighted factor for Constrained Energy Minimization (CEM) to overcome its sensitivity problem to the target spectrum.

In the study, two adaptive classifiers based on image eigen-features are proposed for multispectral image classificationm. One is based on a liner filter with weights adaptively updated by the principal eigencompoments, and the other is an artificial neural network (ANN) with weights trained by the image eigen-features. We first propose an adaptive signal subspace projection (ASSP) approach to detect and extract target signatures in unknown background. The weights of ASSP are adjusted adaptively by using the eigen-features which are updated recursively by the adaptive eigen-decomposition algorithm. Then, we proposed an ANN classifier based on back propagation multilayer perception (BPMLP) with weights trained by the image eigen-features. Simulation results validate the image eigen-features can alleviate the noise effect in classification and the proposed ASSP and ANN classifiers have lower detection error and fast convergence rate than conventional Wiener filter and per-pixel ANN methods.

Transformation of coordinates usually invokes level-wised processes wherein several sets of complicated equations are calculated. Unfortunately, the accuracy may be corrupted due to the accumulation of inevitable errors between the transformation processes. In this thesis, we present a new symbolic regression engine for generating regressive models for direct transformation from GPS signals to 2-D coordinates. The functional models to be found by the regression engine are represented in tree-based chromosome. Genetic operators like mutation, crossover, and reproduction are applied on such trees. In order to improve the efficiency of the regression process, we design some adaptive mechanisms which adjust the probabilities of applying genetic operations. Methods of trimming and pruning the regressive models for reducing their complexity are also design and implemented. Since target coordinates for a GPS application can be obtained by using simpler transformation formulas, the computational costs and inaccuracy can be reduced. The proposed method, though does not exclude systematic errors due to the imperfection on defining the reference ellipsoid and the reliability of GPS receivers, effectively reduces the statistical errors when the accurate Cartesian coordinates are known from the independent sources. From the experimental results where the target datum TWD67 is investigated, it seems that the proposed method can serve as a direct and feasible solution to the transformation of GPS coordinates.

An approach for building detection and feature lines extraction from airborne LiDAR data is proposed in this paper. The building detection is based on Wavelet Transform and geometric properties of buildings, and the extraction of feature lines is based on Hough Transform and image processing. Although LiDAR data contains rich surface information, the shortcoming is it cannot capture building features such as corners, edges, faces of roofs directly. For data-driven building reconstruction, the feature lines are essential to reconstruct the roof models. The basic idea of the proposed approach is to detect the location of each single building in the raw LiDAR data firstly. Then, the initial feature lines which are divided into external contour lines and internal structure lines are extracted respectively. The external contour lines are extracted using Hough Transform, and the internal structure lines are extracted using collinear analysis. Finally, the roof models are reconstructed by external contour lines as well as the internal structure lines of buildings. The experiment results showed that the regular and simple roofs such as rectangle roofs, gabled roofs and L-type roofs could be reconstructed successfully.

Due to its all-weather capability and stable repeat observations, synthetic aperture radar is able to provide timely information after a natural disaster. In this study, we developed a multi-scale matching technique to efficiently and effectively integrate multi-temporal for fast and accurate change detection. Experimental test results from RADARSAT and ALOS satellite radar images confirm that the proposed automatic registration scheme based on a multi-scale matching greatly reduces the artificial error and required computation, and thus is highly feasible for operational purpose.

Orienting the photogrammetric imagery by means of existing control entities is one of practical issues which aim to automatically acquire the control information from the database and perform the photo measurements in an attempt to reduce the labor involvement, as well as to promote the working efficiency towards orientation solutions. The author investigates in this study the strategic algorithms where the exterior orientations of the aerial photographic imagery are solved by matching with control patches. The proposed working scheme features itself with the designed platform of the human-machine interface where the measurements of control patches on the imagery can be achieved by through manual, semi-automatic, or fully automatic mode. Furthermore, the disagreements of control patches against the imagery, usually the phenomena of the scene change, are able to be identified and fed back to the system. The experiments show that the positioning quality of oriented imagery performed under the proposed scheme can reach up to sub-pixels and one to two pixels in horizontal and vertical components, respectively, a satisfaction towards orienting the imagery.

Multi-temporal satellite imageries and airborne LiDAR data are used in this study to understand the trend of change of Wai-Shan-Din Sandbar, and, thus to estimate the moving rate of the sandbar toward the inner land in recent 35 years. Results are as follows: (1) Images between 1973 to 1990 shows that the sandbar is retreating from the sea toward the land to the east side in a counter-clockwise movement with a larger amount in the southern part of 1.82 kilometers and smaller amount in the northern part of 0.84 kilometers. (2) From 1992 to 1996, the northern part of the sandbar was stable but the southern part was moving counter-clockwise with a rate of 162m per year. (3) From 1997 to 1999, the southern part of the sandbar shifting 150 meters northward. From 2000 to 2003, the annual rate is 50 meters toward the land to the east. (4) Three times of DEMs from 2004 to 2006 were obtained by airborne LiDAR. The northern part of the sandbar was subject very subtle change according to the overlay of thematic maps of zero-meter lines abstracted from Lidar DEMs. However, the southern part shows a change of around 110 to 130 meters toward the land to the east. (5) The zero-meter thematic maps and the volumes within zero-meter lines also show that the area of the sandbar decreased from 1813 to 1721 hectares, and the volume changed from 17.56 to 16.24 million m ³. It is concluded that satellite imageries give a long time of records for understanding the trend of the 2D coastal lines. However, LiDAR enhances a possibility of understanding the change patterns in 3D, including the volume of change.

Image segmentation is an important topic in many applications of image processing. It is a procedure that divides different regions or objects of the image, and thus we can retrieve the regions of interest. The images of the different tree species usually have the similar color distribution, but the differences between them are only texture information. In this paper, an efficient segmentation method based on seeded region growing is proposed for the color remote sensing images of forest, it performs the unsupervised segmentation. We use the coefficients of discrete cosine transform as the texture features, the seeds are then produced automatically by considering both of the color features and texture features. The seeds grow to their boundaries until all pixels are classified. Finally, a region-merging process is adopted to improve the segmentation result. Experimental results show that the proposed method can separate successfully the different color texture in the remote sensing images of forest, and the performance is better than the JSEG algorithm.

Heavy rainfall often triggers widespread soil erosion and landslide hazard in mountainous terrain, resulting in sediment deposition and loss of reservoir capacity. This research aims to identify major sediment sources and predict watershed yield by using satellite images and GIS techniques. Optical satellite images, specifically those of SPOT series, were used to identify landslide areas from 2001 to 2006. To distinguish landslides from high vegetated areas, field studies were conducted for investigation of normalized difference vegetation index (NDVI) covering several major land covers to determine the threshold value for bare land. Slope features, based on spatial analysis of Digital Terrain Model, were also introduced to differentiate landslide from other low vegetated area. Study showed that landslide covered about 634 hectares at Shihmen Reservoir, 236 hectares at Touchien catchment, 1,189 hectares at Dajia catchment , and 4,308 hectares at Kaoping catchment in post-typhoon season images of 2006. To estimate soil erosion and the mechanism of sediment transport, the Universal Soil Loss Equation (USLE) and soil delivery ratio were used to estimate non-point sources of soil erosion, and those of point sources were determined by the identified landslide areas, which were then multiplied by the predicted depth of weathered layer to obtain the total volume. Study indicated that the sediment yield of Shihmen Reservoir, Touchien Creek, Dajia creek, and Kaoping Creek were estimated as 1,070,097 m3, 323,540 m3, 1,836,431 m3, 4,824,430 m3, respectively. Finally, under the assistance of Northern Region Water Resources Office, our predicted results were validated by the measured sediment yield of Shihmen Reservoir.

3D structure lines are geometric evidences of building roofs and can be resulted from photogrammetric mapping process, feature extraction of LiDAR data sets or other sources. Hence, 3D structure lines by nature are advantageous for inferring to building roof reconstruction. The objective of this investigation is to present a novel algorithm of building roof reconstruction, termed CSR (Construct-Shape-Refine) in this study. The proposed algorithm aims at reconstructing building roof model purely by employing 3D structure lines or integrating 3D structure lines and imagery line features. Geometric inferences are imposed at the stage of Construct where topological relationship of the 3D structure lines is established while 3D coordinates of roof corners are adjusted and estimated through Shape processes, apart from compensating the hidden boundaries, if any. The efficiencies of two streams, fusion of 3D structure lines that result from different sources or fusing 3D structure lines with photogrammetric line features, are gained as promoting the building roof quality via Refine workflow. The experiments of the Construct-Shape steps demonstrate that the proposed method is independent of building roof type as well as whether 3D structure lines are complete for both successful and efficient performance of building roof reconstruction on 3D line basis. Furthermore, the Refine step offered more robust and satisfactory building reconstruction result as compared to the situation when only single data set was attempted.

Spectral mixture analysis (SMA) approach is one of the most popular methods for reducing the mixed pixel problem on satellite image classification. The objective of this study was to examine the applicability of SMA to the estimation of forest abundance using SPOT image in the Kenting National Park. We acquired three fraction images which were green vegetation (GV) and soil derived from linear spectral mixture model. In general, both the GV and normalized difference vegetation index (NDVI) were to examine the forest abundance, and for the purpose to compare the examine accuracy, the forest abundance from high resolution (1×1 m) IKONOS image was derived. Accuracy assessment of the GV and NDVI was to calculate three indices, which were mean absolute percentage, root mean square and Theil’s U inequality coefficient. The results showed that the GV was higher accuracy than NDVI on the three indices, meaning that the SMA could be an efficient tool for monitoring forest.

Photogrammetry has been considered as the most efficient technique for extracting 3D spatial information in the last few decades. However, the point-by-point measurement of using floating mark has become the bottleneck while reconstructing the 3D city model. In this paper, we expanded the floating mark to the floating model based on the concept of model-based building extraction. The measuring tool is no longer an abstract point but many kinds of 3D model, which can be scaled, rotated, or moved in the space. The floating model is defined with a datum point indicates the model’s position. Each kind of models is associated with a set of pose parameters to describe its rotation about the three orthogonal axes and shape parameters to describe its scales along predefined directions. In other words, the floating model is a flexible entity floating in the space, and can be adjusted to fit the object by these parameters. If the model parameters are good enough to represent the 3D spatial information of the object, the projection of the floating model on every overlapped image will all be coincident to the object’s outlines. In order to meet the needs of building reconstruction, we designed several primitive models that can be categorized into four kinds: point, line segment, surface, and solid. By means of human-computer interface, the operator is able to choose the most suitable model and measure the object on multiple images. The complicated building can be split into several parts, modeled part-by-part, and finally joined into one complex model. Based on the floating model theory and model-based building extraction, we proposed a semi-automated strategy. A friendly human-machine interface is designed for the operator to choose and adjust the floating model to fit the images manually. Then, the computer calculate the optimal fit by an ad hoc Least-Squares Model-Image Fitting algorithm. Thus the 3D spatial information can be extracted object-by-object, which increases the efficiency and accuracy.

One of the key prospects of data fusion is focused on exploiting the complementary information among different sensors. In this paper, the multiple classifiers approach is utilized for the multisource classification/data fusion to fully utilize as much of the available information among different data sources as possible. Three different weighting policies (variance reduction technique, rms distance weighting, and average distance weighting) applied to the multiple classifiers approach are introduced. The performance of each combination method was demonstrated and compared with the fusion of the SAR and optical images for the terrain cover classification. Experimental results show that the classification accuracy is dramatically improved by making use of the proposed method. In addition, both the multiple classifiers using the rms (root mean squared) distance weighting and the average distance weighting outperform that of using the variance reduction technique.