Integration of multisource remote sensing data for improvement of land cover classification

摘要

The use of multisource remote sensing data for land cover classification has attracted the attention of researchers because the complementary characteristics of different kinds of data can potentially improve classification results. Such a multisource approach becomes increasingly important with the ready availability of a variety of satellite imagery. However, using more input data does not necessarily increase the classification performance. On the contrary, using too many remotely sensed input datasets increases data volumes, including noise, redundant information and uncertainty within the dataset. Therefore it is essential to select relevant input features and combined datasets from the multisource data to achieve the best classification accuracy. Other challenging tasks are the development of appropriate data processing and classification techniques to efficiently exploit the advantages of multisource data. The goal of this thesis is to improve the land cover classification process by using multisource remote sensing data with recent advanced input feature selection, data processing and classification techniques. The capabilities of non-parametric classifiers, such as the Artificial Neural Network (ANN) and Support Vector Machine (SVM), were investigated using various multisource datasets over different study areas in Vietnam and Australia. Results showed that the multisource datasets always gave higher classification accuracy than the single-type datasets. The non-parametric classifiers clearly outperformed the commonly used Maximum Likelihood algorithm.The feature selection (FS) technique, specifically the wrapper approach, based on Genetic Algorithm (GA) was proposed to search for the appropriate combined datasets and classifier’s parameters. The integration of GA and SVM classifier was employed for classifying multisource data in Western Australia, including multi-date, multi-polarised SAR and optical images. It was revealed that the SVM-GA model gave significantly higher classification accuracy with less input data than the traditional method. The GA algorithm also performed better than the conventional Sequential Forward Floating Search algorithm.The Multiple Classifier Systems (MCS) or classifier ensemble technique, which can potentially improve classification performance by exploiting the strengths and alleviating the weaknesses of different classifiers, was also evaluated using different algorithms and combination rules. An experiment was carried out with the MCS technique using the ANN, SVM and Self-Organising Map (SOM) classifiers over a study area in New South Wales, Australia. The investigation shows that the MCS technique, in general, provided higher classification accuracy than individual classifiers.Finally, a synergistic model using the FS based on GA and MCS techniques was developed to further increase the performance of multisource data classification. Results confirmed that the hybrid model of FS-GA and MCS outperformed other methods and significantly improved on the performance of both the FS-GA and MCS algorithm. Results and analyses presented in this thesis emphasise that using multisource remote sensing data is an appropriate approach for improvement of land cover classification. The proposed methodologies are very efficient for handling high dimensional, complex datasets such as combined multisource data.