OPTIMIZATION OF ARTIFICIAL NEURAL NETWORKS IN SATELLITE REMOTE SENSING DA TA ANALYSIS

Satellite remote sensing data have been broadly applied in extraction of thematic land cover information. The multi-spectral classification is the mostly used method. Traditional supervised classification algorithms are in the category of statistical pattern recognition. Due to the limitations of spatial and spectral resolutions of remotely sensed data, traditional classification has always been challenged in thematic accuracy of land cover mapping. To improve the performance, new technical approaches are in demand. Artificial Neural Network is one of the emerging approaches in multi-spectral remote sensing and multi-source spatial data handling. However, none-optimized method always met difficulties to get an ideal result. In this study, Backpropagation ANN (BPANN) was applied for the classification of 1999 Landsat- 7 Enhanced Thematic Mapper Plus (ETM+) data. Several optimization algorithms including Conjugate Gradient, Resilient back-propagation (RPROP), and Gray-code coding method were also applied to improve the performance of the classification. The result shows that the optimized BP ANN is a more robust and advanced approach in remote sensing image analysis than the traditional statistical based or none-optimized BP ANN based classification method. Therefore, application of satellite derived remote sensing data in natural resource and environmental mapping can be improved