Investigation On Classification Of High-Resolution Satellite Image By Back-Propagation Neural Network

From the 1970s, the remotely sensed imagery has shown itself a trend of meter-scale resolution and the high-resolution satellite imagery has been the main product with the development of satellite technology. The classification techniques of remotely sensed image can extract effectively rich information from this high-resolution imagery. By data processing and analyzing, the scientific results are obtained to meet the actual production and application, and play a major role in commercial and civilian fields.ANN (Artificial Neural Network) is one nonparametric classification method. It has good adaptability and complicated mapping capability, which can detect and exploit nonlinear data patterns. Compared with those classification methods based on the conventional statistical theory, this method, which needn't suppose or estimate probability distribution of sample space in advance, can attain good classification results and high classification accuracy when executing the classification using the remotely sensed image with highly ambiguous information. In addition, this method can efficiently integrate spectrum features with texture features to execute high-resolution image classification, and is helpful improving the classification accuracy by utilizing fully spectral and spatial information.In this thesis, the multi-spectral (resolution 2.44 m) and panchromatic (resolution 0.61 m) images acquired by the satellite QuickBird in November of 2006 around Xipu plain areas in Chengdu, China, are selected as data source. Back-propagation (BP) neural network is selected as the classification approach to investigate and analyze the image classification and accuracy of the study area. The investigation performed and relevant conclusions are outlined as follows:First, based on the characteristics of QuickBird image, this thesis presents data processing and image analyzing of study area. The Optimum Index Factor (OIF) analysis and image fusion are performed to obtain the optimum spectrum features, while texture analysis based on Gray Level Co_occurrence Matrix (GLCM) is used to extract texture features. And then, the feature data to be classified are prepared by combining the 432-band image with the contrast (CON) texture image obtained from panchromatic image.Second, BP algorithm is processed for image classification. The training samples are selected from the feature data of the study area and the BP network structure is constructed. To ensure the training stability and improve the convergence rate, some approaches are explored to improve the BP algorithm, such as normalized pretreatment of training samples, importing momentum factor, setting different training rates in each layers and adjusting network structure. The computer program of BP algorithm is designed and developed in the environment of ENVI/IDL (Interactive Data Language) to implement the QuickBird image classification and accuracy evaluation of study area. The experiments show that this method can attain high-accuracy classification results, whose overall accuracy is higher than 93% and Kappa coefficient is greater than 0.9.Third, the classification results derived by the improved BP algorithm are compared with those derived by a conventional method, i.e., maximum likelihood classification (MLC) approach. The testing results show that the BP classification approach can identify some small-area objects that are erroneously classified by the MLC method and the BP classification has a higher overall quality. The classification results derived by the BP neural network have a higher overall accuracy and Kappa coefficient than the MLC.Based on the above experimental results, the primary conclusion drawn from this research is that the BP classification approach is more robust than the MLC and it is more suitable for performing classification for this study area.This thesis implements the automatic classification of high-resolution satellite image using the BP neural network with the joint use of spectral and textural features and accomplishes information extraction of ground objects. These experimental results and conclusions are of significance for the resource investigation and mapping of land use and land cover with high-resolution satellite imagery in the southwest part of China.