| Dramatic urbanization processes have happened in many regions and thus created a number of metropolises in the world. The Pearl River Delta (PRD) is one of such typical areas, where the urban land use/land cover has been significantly changing in the recent past. As one of the most important implications, a large increment of impervious surface (IS) turned out to be one of the features of fast urbanization process and has been influencing the urban environment significantly, including urban flooding, urban climate, water pollutions, and air pollutions. Therefore, the estimation of IS would be very helpful to monitor and manage the urbanization process and its impacts on the environment. However, accurate estimation of urban IS remains challenging due to the diversity of land covers. This dissertation attempts to fuse optical and SAR remote sensing data to improve the accuracy of urban impervious surface estimation (ISE) in humid subtropical regions (HSR). The seasonal characteristics of land covers and its impacts on ISE in HSR are all investigated. Some interesting findings are summarized as follows.;Firstly, the study demonstrates quite a special pattern of the seasonal effects of ISE in humid subtropical areas that is different from that in mid-latitude areas. According to the results, in subtropical monsoon regions, winter is the best season to estimate IS from satellite images. There are little clouds, and most of the Variable Source Areas (VSA) is not filled with water. On the other hand, autumn images obtained the lowest accuracy of IS due to the clouds coverage and the water in VSA. Autumn is a rainy season in a subtropical monsoon region, for which clouds occur very often and VSA areas are always filled with water. Consequently, clouds are confused with bright IS due to their similarly high reflectance, and more water in VSA is confused with dark IS due to their similarly low reflectance.;Secondly, a novel feature extraction technique, based on the shape-adaptive neighborhood (SAN), is proposed to incorporate the advantages of human vision into the process of remote sensing images. Quantitative results showed that improvement of SAN features is particularly significant improvement for the unsupervised classifier, for which the overall accuracy increased from 0.58 to 0.86, and the Kappa coefficient increased from 0.45 to 0.80, indicating promising applications of SAN features in the unsupervised processing of remote sensing images.;Thirdly, a comparison study of ISE between optical and SAR image demonstrates that single optical image provides better results than using single SAR image. In addition, results indicate that support vector machine (SVM) is a better choice for ISE using Landsat ETM+ (optical) images, while artificial neural network (ANN) turns out to be more sensitive to the confusion between dry soils and bright IS, and between shades and dark IS. However, ANN gets a better result using ASAR (SAR) image with higher accuracy, while the SVM classifier produces more noises and has some edge effects. Considering both the merits and demerits of optical and SAR images, synergistically fusing the two data sources should be a promising solution. Comparison of three different levels of fusion shows that pixel level fusion seems not appropriate for optical-SAR fusion, as it reduces the accuracy compared to the single use of optical data. Meanwhile, feature level fusion and decision level fusion obtained better accuracy, since they improves the identification of IS from shaded areas and bare soils.;Fourthly, a methodological framework of fusing the optical and SAR images is proposed. Three different data sets are used to assess the effectiveness of this methodological framework, including the Landsat TM and ASAR images, the SPOT-5 and ASAR images, and the SPOT-5 and TerraSAR-X images. In addition, different methods (e.g. ANN, SVM and Random Forest) are employed and compared to fusion the two data sources at a mixed level fusion of pixel and feature levels. Experimental results showed that the combined use of optical and SAR image is able to effectively improve the accuracy of ISE by reducing the spectral confusions that happen easily in optical images. Moreover, Random Forest (RF) demonstrated a promising performance for fusing optical and SAR images as it treats the two data sources differently through a random selection procedure of variables from different data sources.;The major outcome of this research provides evidence of the seasonal effects on IS assessment due to phenological and climatic characteristics, as well as provides an applicable framework of methodology for the synergistic use of optical and SAR images to improve the ISE. Since the PRD region is highly typical of many fast growing areas, the methodology and conclusions of this research would serve as a useful reference for other subtropical, humid regions of the world. |
