The Classification Of High Resolution Remotely Sensed Image Based On Gravitational Search Algorithm Optimized Neural Network

With the rapid development of remote sensing technology,obtaining geographic information by means of remote sensing techniques has become one of the most common ways in geographic study.Nowadays,since the satellite sensors have been improved a lot,the resolution remote sensing image has been able to reach to the decimeter level.Therefore,how to deal with the spectral and texture information in high-resolution images and apply them to solve some pracical problems have become a key issue in the field of high-resolution remote sensing image classification.In this paper,we will discuss the utilization of artificial neural network(ANN)and error back propagation algorithm(BP)in the domain of high-resolution remote sensing image classification.Compared with some traditional classification algorithm,the BP neural network algorithm is more flexible and can deal with the nonlinear dataset,which is an effective method to solve the high resolution remote sensing image with different objects which have the same spectrum and the same objects have different spectrum.However,BP neural network has the disadvantages of slow learning speed and local minimum trapping problems.Thus,in order to tackle these issues,we will introduce a gravitational optimization algorithm(GSA)to obtain the optimal parameters in neural network,and further enhance its classification accuracy.In this paper,the Gaofen-2 high resolution remote sensing images are used as the basic experimental data and the main content of this paper is as follows:(1)After analyzing the influence of parameter ? on the convergence mechanism of the swarm in gravitational search algorithm,introduce a dynamic parameter adjusting strategy based on the agents search behaviour during the different evolutionary iterations.In addition,make use of the stability conditions to constraint ? changes in order to guarantee the stable convergence.This paper employed four modified GSA algorithms and then tested the searching performance of SCAA and comparing algorithms on 13 conventional problems and 15 CEC2015 rotated and shifted functions.The results indicated that SCAA can avoid being trapped into local optima,have more stable converge trajectory and show more superior convergence rate and accuracy.(2)According to the different characteristics of objects in high resolution images,extract the spectral and spatial features of the typical objects,including four spectral bands,normalized differential vegetation index,normalized differential water index,grayscale symbiotic matrix and morphological attribute profile.After that,combine all the features and select the optimal feature set via intelligent optimization algorithm.This paper summarize and analyze the optimal eigenspaces of two different remote sensed images.The experimental results demonstrated that the intelligent optimization algorithm-based feature selection method can realize the feature dimension reduction effectively and improve the classification accuracy.(3)On basis of the structure and parameter settings of ANN,construct the swarm in SCAA and design the objective function.Then using iterative loop of SCAA to optimize the weight and biases settings of ANN.This paper employed the iris classification dataset to test the performance of SCAA optimized ANN clasifier.It was verified that SCAA can avoid the slow learning speed and premature convergence problems in BP.Besides,the different classifiers were utilized on the Gaofen-2 high resolution remote sensed data in urban area and wetland in Shanghai city and the results showed that the SCAA optimized ANN had the best performance according to the overall accuracy and kappa value.