| The information extraction of remote sensing data has always been considered as a core and basic work in remote sensing technology,which needs to be improved and improved continuously.The most critical pace for the extraction of ground feature information is remote sensing image classification.With the rapid development of science and technology,high-resolution image is widely used in many fields,but the mesoscale image is rich in more abundant ground feature information,and the mesoscale image is mostly open-source data,which is used more frequently and involves a wide range of areas.In practical application,it can meet the requirements of most information extraction.Therefore,the demand for mesoscale remote sensing image cannot be completely met by high-resolution remote sensing image.It is replaced by sensory image.There are different classification results when choosing different classification methods,which is the key to the classification of mesoscale remote sensing images.In this paper,BP neural network is selected as the original method to improve the accuracy of remote sensing image classification.Back propagation algorithm(BP)is a well-known method for training multilayer feedforward artificial neural network(FFANNs).BP algorithm is widely accepted in remote sensing image classification,but it has some shortcomings in long-term use.Because the core of BP neural network is gradient method,the problem of slow convergence speed and easy convergence to local minimum value cannot be avoided in the calculation process.Therefore,the effective application of BP neural network is limited.By replacing the original gradient method in BP algorithm,the problem of local minimum value can be avoided.In this paper,bat algorithm is used to improve BP neural network and replace the weight threshold algorithm in the original BP neural network to solvethe problem that the network is sensitive to the initial weight threshold.On this basis,it is found that bat algorithm has premature convergence and imprecise solution to high-dimensional problems.The adaptive pace size is used to optimize bat algorithm to update the position of individual bat cooperatively,and Gauss disturbance is added in group iteration.Dixon price function is used to test the optimization ability of the improved algorithm after completing the above optimization work.According to the landform and land cover types,three different areas were selected as experimental areas in China,namely Kangding city located in the western mountainous landform,Guangzhou city located in the southern hilly landform and Bengbu City located in the eastern plain landform.There are differences in land cover in the three experimental areas.As bat algorithm,it has an excellent test value for the optimization effect of BP neural network algorithm.The original BP neural network,bat algorithm improved BP neural network(BABP model)and bat algorithm improved BP neural network(SBABP model)were used to analyze the Landsat of Kangding,Guangzhou and Bengbu 8 remote sensing image classification experiment,and three groups of classification results of the three experimental areas are analyzed,on this basis,the classification results of the three experimental areas are comprehensively analyzed and compared.It is found that SBABP algorithm is the plus in the classification effect of remote sensing image in the three experimental areas.The overall classification accuracy of Bengbu City,Guangzhou City and Kangding city is 93.89%,88.56% and 84.2%,respectively.Kappa coefficient is 0.9183,0.857 and 0.8153,respectively.Compared with the original BP neural network,the overall classification accuracy is improved by 3.54%,3.28% and 4.23% respectively.SBABP model is applied to the classification of remote sensing image.In the single land type,the effect of classification accuracy improvement is the most prominent is bare land,which is 26.2% higher than the original BP neural network,followed by construction land which is 12.8% higher. |
PDF Full Text Request