Remote Sensing Image Recognition Of Construction Land For Rural Settlements Based On Convolution Neural Network

Construction land is an important place for human production and life.How to effectively manage and supervise the construction land is very important.At present,urban construction land has been fully approved and recorded,and construction land information can be obtained directly from the relevant management departments.However,due to the lack of unified planning and decentralization,the size and location of settlements in rural areas can not be grasped in detail.Traditional recognition methods need to extract the characteristics of construction land manually.This method is not only inaccurate in recognition,but also depends on experience in feature design.With the development of deep learning,convolution neural network can automatically extract complex features of construction sites and effectively improve the accuracy of identification.Under this background,this paper combines the advantages of the convolution neural network with the object-oriented method and the deep learning architecture,and takes the rural area of Dan Ling County,Meishan,Sichuan Province as the research area,and uses the 1:2000 unmanned aerial vehicle(UAV)remote sensing image data to study the best region of interest extraction method in the construction land recognition process,and quickly transport the neural network to the neural network.Training samples.By comparing the traditional machine learning algorithm and different types of convolution neural network,the convolution neural network with the highest construction land recognition rate is constructed.The main results of this paper are as follows:(1)this paper studies the method of extracting interesting area of remote sensing image.Through the experiment,we know the best scale k=300 of the selective search method,the best threshold n=0.05 of the edge location object acquisition method and put forward the method of region of interest extraction with Multi Strategy fusion.The initial image is extracted with the edge location object acquisition method,and then the data is filtered by the maximum value suppression method.Then the selected search method is used to extract the region of interest image.The best result is that the overlapping rate of Io U can reach 0.85 in all experimental methods.Although the extraction speed is slightly weaker than the edge location object acquisition method,it is the best choice to integrate the Multi Strategy fusion method of region of interest extraction.(2)the influence of training parameters on training in remote sensing image recognition is analyzed,and the optimal parameters are selected from it.When the learning rate is at 0.07,the loss value is the lowest and the maximum iteration number becomes stable after 8000 times,the training time and resource can be stopped,and the optimization algorithm type selection is SDG optimal.This paper also uses 4 kinds of scales to standardize the image extracted from the region of interest.Through the experiment comparison,we find that the selection of 64*64 scale remote sensing images can get better network convergence and recognition effect.(3)this paper realizes the Ada Boost algorithm based on Haar-like and LBP features,several common convolution neural networks and PCANet for construction land recognition.Through the comparison of accuracy,it is found that Caffe Net network has the highest score of 98.52% in all algorithms,and it is most suitable for the identification of construction land.In view of the shortcomings of Caffe Net,which has less speed than simple convolution neural network Le Net-5 and PCANet,the speed of training and testing is accelerated,and the accuracy is increased to 99.66% by simplifying the network structure and adjusting the parameters.