The study planed to combine SPOT satellite remote sensing (RS) image and traditional fix point water quality samplings by using a two-stage unsupervised model to classify remote sensing image, then adding a few laboratory analytical water quality samplings with co-kriging (CK) interpolation method to construct an overall distributed surface river pollution index map (RPIM). This research developed an unsupervised co-kriging interpolation model (UCK) which adopting the automatic optimization method to classify the pollution of a river in a satellite image, then secondly adding digital elevation model, and on-site water quality sampling data which has the same time with the unknown existed classified satellite image in the first stage for spatially co-kriging interpolation (UCKD). The procedure took into account the merits of using spot image, digital elevation model, and on-site sampling data. Moreover its results showed the advantage when comparing to supervised co-kriging (SCK and SCKD) and traditional interpolation methods (IDW).

Paddy rice is one of the major food crops in Taiwan. The government investigated the paddy rice area through aerial photography every year. However, paddy rice thematic maps require using much manpower which is very time-consuming and funding. How to quickly obtain this geo-information through very high-resolution aerial photographs is an important research issue. In this study, we want to develop a novel decision model (Region Object-oriented Entropy Based Classifier, ROEBC) which integrates the patch-scale information and Entropy Base Classifier. It also compares these results of traditional concept of pixel-based classifier by Maximum Likelihood Classifier (MLC). The ROEBC categories decision model can find the ideal cutting point from each spectral band through the value of attributes on Information Gain (IG). Based on these IG values, we can obtain rules from image information successfully. This method can clearly show the differences on image knowledge rule content by the traditional statistical parameter classifier (such as MLC). After that, in this study, we check the point accuracy and area accuracy at paddy rice thematic maps with MLC and ROEBC methods, respectively. It shows that the regional based classifier of ROEBC methods is better than those of the pixel based classifier of MLC. Finally, this study discussed the feasibility of paddy rice object results combine with other GIS data on the agriculture information management issues of future projects.

Cloud cover is an inevitable interference when mapping land use/cover with optical satellite imagery. In this study, we apply region growing processing to delineate unrecoverable thick cloud and use Fourier analysis to recover ground information from hazy areas with single temporal imagery. Several methodologies across literature successfully solve cloud problems, but most methods require additional cloud-free reference areas or imagery, which may be unavailable in the real world. Moreover, visual methods rather than quantitative methods are used for assessing results, which can be subjective and arbitrary. Most importantly, the feasibility of applying haze-off imagery to image classification is seldom discussed. To overcome the existing limits, expert method is applied to assess the thick cloud delineation and image classification and normalized difference vegetation index (NDVI) is used to evaluate the recovery degree of ground information after the haze-off processing for quantitative verification of the results. This study revises the image enhancement and region growing algorithm to delineate unrecoverable thick cloud. Accuracy assessment shows the overall accuracy of delineation could be 90% above in each study area. For hazy areas, Fourier analysis is used to reduce haze interference and recover ground information. The proposed haze filter increases the overall accuracy of the whole scenes by about 4%. The overall accuracy of hazy areas in the imagery increases the most (by 6%), while that of shadow areas decreased slightly. The influence of haze on NDVI is also reduced with statistical significance (p < 0.01). Both thick cloud and hazy areas processing can be achieved with no cloud-free area or reference imagery required. Future applications include preprocessing of satellite imagery in land use/cover mapping, which can decrease the manpower to interpret and remove cloud areas and increase the usability of the satellite imagery.

Tropical sea surface temperature (SST) data derived from the Geostationary Operational Environmental Satellite (GOES) is analyzed by using data mining to explore the error sources of data and to further improve its accuracy. The SST data has been pre-processed into two kinds of data set, the daily mean and hourly. The root mean square error (RMSE) of daily SST estimate is reduced from 0.58 K to 0.37 K and the mean absolute percentage error (MAPE) is 1.03% by using the Back Propagation Network (BPN) algorithm. For the hourly SST estimate, its RMSE is also reduced from 0.66 K to 0.44 K and the MAPE is 1.1%. This indicates that the BPN algorithms improve the accuracy of the SST. While the proportion of cloud contamination is in different circumstances, the RMSE of retrieval satellite SST by using the BPN algorithm can be maintained below 0.38 K. This demonstrated the efficiency ability of anti-noise analysis of the neural algorithm. The factor analysis also shows that the errors are mainly caused by air temperature and then followed by wind speed and relative humidity.

This study utilized decision tree and Bayesian network algorithms to extract the knowledge of shallow landslides based on susceptibility analysis in the Shimen reservoir watershed. Furthermore, the uncertainty filter, significant analysis and feature reduction methods for landslide factors were also proposed. The objective is to develop a post- and data-driven analysis system for landslide detection and risk assessment in a regional scale. This study did not distinguish different types of landslides, and all landslides were induced by heavy rainfall. Experimental results demonstrate that the developed landslide factor model is effective for landslide detection in the study site. After filtering uncertain data, the reliability of landslides verification and susceptibility assessment has been improved significantly. Based on the experiments, Bayesian network can provide more reliable prediction and susceptibility results than decision tree in the study case. In addition, feature reduction can improve the computation efficiency while maintaining acceptable check and prediction accuracy.