Using genetic algorithms and artificial neural networks for multisource geospatial data modeling and classification
Date of Completion
January 1998
Keywords
Geography|Remote Sensing|Artificial Intelligence
Degree
Ph.D.
Abstract
In the remote sensing and geographical information system (GIS) community, huge amounts of digital geospatial data are produced daily. On the other hand, we lack efficient tools to extract useful information from these raw data. The bottleneck of information processing, e.g., modeling and classification, becomes most challenging. The applicability of the traditional statistical approaches or expert systems are hindered when dealing with geospatial data of inaccuracy, and multiple measurement scales. In addition, prior knowledge, e.g., distribution assumption for Maximum Likelihood Classifier (MLC) or weight assignments for Multiple Criteria Evaluation (MCE), is required for the traditional methods. ^ The purpose of this dissertation research is to address this information extraction problem in remote sensing and GIS industry. The major methods used are genetic algorithms, artificial neural networks and wavelets. A genetic learning artificial neural network (GLANN) model was developed whose learning algorithm is a genetic one rather than traditional backpropagation. It is concluded that GLANN can provide an alternative of and improvement over conventional overlay, multicritiera evaluation or maximum likelihood methods. ^ To improve the classification accuracy of the high resolution remote sensing images, e.g., 10 meter SPOT panchromatic image, where pixel-by-pixel classification approaches have proven to be inadequate, a wavelet transform method was presented to extract texture feature in a multiple scale fashion. The classification results indicated that texture feature of multiple scale represented by wavelet energy achieves much better classification accuracy than those derived by the conventional fixed window size texture extraction algorithms. ^ To take advantage of the high spectral resolution of Landsat TM images and the high spatial resolution of SPOT panchromatic images (SPOT PAN), a wavelet transform method to merge the two data types was presented. A quantitative method for the quality evaluation of the multiresolution fused image was proposed. Experiments showed that the wavelet transform methods performed better than other data fusion approaches examined. ^
Recommended Citation
Zhou, Jiang, "Using genetic algorithms and artificial neural networks for multisource geospatial data modeling and classification" (1998). Doctoral Dissertations. AAI9918105.
https://digitalcommons.lib.uconn.edu/dissertations/AAI9918105