Wetland Information Extraction Based On ICA And Adaptive Minimum Distance Classification

Wetland is a transitional ecosystem between land and water, which enjoys the same important status as forests and marine ecosystems. Sustainable utilization and protection of wetland resources are vital contents of Earth science. However, wetland resources have not received effective protection due to various factors. Therefore, increasing its importance has been attached to the exploratory study of wetlands. Past decades have witnessed the rapid development of remote sensing technology, and the modern technology that based on information extraction has been widely used in monitoring wetland resources. Traditional methods for information extraction mainly includes artificial visual interpretation, supervised classification, unsupervised classification, etc. But traditional methods have disadvantages such as slow speed, poor accuracy and complicated process. That is why traditional ways’ improvement and explore for new methodologies of information extraction has become hot issues in the research of wetland remote sensing.My research takes Western Dongting Lake area as the main research district, and employs Landsat-TM and high resolution remote sensing images in SPOT-5as data sources, in order to find best band combination that applies to Landsat-TM and SPOT-5images wetland classification. Then introduces ICA(Independent Component Analysis) into information extraction of wetlands, combined with adaptive minimum distance classification, which can classify wetlands and evaluating results. This research aimed at improving the system of wetland information extraction, providing a basis for future wetland remote sensing and looking for scientific, accurate and reasonable approach to Dongting Lake wetland research. Main findings are as follows:(1) Best band combination analysisFine for wetland information extraction requirements, quickly and accurately identifying different typical wetlands, often needs in-depth analysis of images band, imaging features and the spectral characteristics of surface features. This research integrate analysis of spectral characteristics with evaluation based on information, so as to finding best band combination that suitable for wetland information extraction.This research has determined the best band combination for Landsat-TM image is RGB=453and the best band combination for SPOT-5image is RGB=453. (2) Independent component analysisUsing PCA(Principal Component Analysis) and ICA to deal with Landsat-TM and SPOT-5images and analyze results, which can find that if the PCA and ICA has caused the loss of images. Additionally, using minimal distance to classify results of PCA and ICA and make precision evaluation. From the experiment, I can conclude that PCA and ICA processing does not reduce the amount of image information, cause the loss of images and affect the visual interpretation of typical wetlands. PCA and ICA processing can increase the separability of typical types of Dongting Lake wetlands. And what’s more, ICA is better than PCA. ICA, based on higher-order statistics, not only can remove the correlation between the bands, but also have independent property between different components, enhace the separability of different land types, remove the negative impact on remote sensing image classification of correlation and improve the accuracy of wetland’s information extraction.(3) Improvement of FastICAUsing FastICA and M-FastICA to deal with Landsat-TM and SPOT-5images,which can help do statistics of two algorithms’calculation amounts, in order to find whether improved algorithms have improved the efficiency of the ICA processing. Minimal classify and precision test on results of two algorithms’images can explore whether improved algorithms have impacts on wetland information extraction. By experiment, there is no obvious difference in accuracy of M-FastICA processing and former algorithm, but the convergence of the algorithm is greatly improved and the speed has significantly increased. It is safe to conclude that M-FastICA can improve the effectiveness of independent component analysis on wetland classificaton.(4) Adaptive minimum distance classificationUsing minimum distance classification, maximun likehood classification and adaptive minimum distance classification to classify Landsat-TM and SPOT-5images, which can help evaluate the accuracy of results. Conclusion:Landsat-TM image maximum likelihood classification reduces the overall classification accuracy than the minimum distance0.87percentage points, kappa coefficient decreases0.0097, compared to the minimum distance classification, adaptive minimum distance classification can improve the overall accuracy by3.48%, kappa coefficient increased by0.0410; for SPOT-5images, maximum likelihood classification is lower than the minimum distance classification overall accuracy1.74%, kappa coefficient decreases0.0200, compared to the minimum distance classification, adaptive minimum distance classification overall accuracy is improved by3.47%points, kappa coefficient increased0.0203. It is safe to conclude that when it comes to wetland classification, minimal distance classification is better than maximun likehood classification, but the superiority is not obvious. Adaptive minimum distance classification is superior to the other two methods, and the classifuication accuracy is improved significantly.