A Segmentation Model For Farmland Extraction From High Resolution Remote Sensing Image

Farmland is an important type of surface cover.The extraction of farmland information is of great significance to the management of agricultural production.The timely and accurate extraction of farmland information by using high spatial resolution remote sensing images can provide important reference for land grade evaluation,basic farmland protection,dynamic monitoring of cultivated land and decision making.High score two is the first optical remote sensing satellite with the first spatial resolution superior to 1 meters in our country.It is an important data source for high precision extraction of farmland.However,the area of single plant covering the farmland is very small on the high score two remote sensing image.It is necessary to take full consideration to the design of the covolume neural network structure of farmland from remote sensing images.The ratio between the spatial resolution of the image and the coverage area of the plant is proportional.In order to better extract farmland information,this paper takes the January 11,2018 Shandong Zhangqiu high score two remote sensing image as the experimental research object.A convolution coder network model(Convolutional Encode Neural Networks,CENET)is proposed,and the segmentation of farmland image is realized with supervised training method.The main work and conclusions of this article are as follows:1.The design of the convolution encoder network model.The single plant coverage area of farmland crops is small,contains few pixels,contains little details and continuous plants,and the traditional image segmentation algorithms can not solve the above problems effectively.In this paper,a convolution coder model(CENET)is designed for the farmland to be image language.It is divided by meaning.Through the "width" convolution method,we can make full use of the characteristics of farmland images to achieve the segmentation of farmland image pixel level.2.Based on convolution coder model(CENET),the image segmentation algorithm of farmland is realized.The model is divided into two sub models.The first model is the training model,which is divided into three parts.The first part is the convolution kernel,which is used to study the characteristics of farmland.The second part is the full connection layer,which is used to transform the two-dimensional feature map of convolution output into one vector of one dimension;the third part is the encoder and the encoder uses the conversion.The function encodes the learned features and maps the coding results to the corresponding categorynumbers.The second model is the recognition model.On the basis of the encoder,the Bayesian principle is used to analyze the classification values of the pixels in the training process,to obtain the prior knowledge and to judge the output of the all connected layer.In the model training phase,we use the labeled farmland and other categories of samples to train the models,so that the model can get enough discriminating power.The training model is used to segment the image by pixel segmentation and get the final result.3.Model optimization and comparison experiment design.In this paper,Batch Nomalization(BN)is used to optimize the model.Before activating the function,it can speed up the learning speed of the model,make the model training process more stable,and play the role of regularization.Using Bias as an auxiliary means of classification and recognition,the accuracy of segmentation is further improved.In this paper,SegNet model and total convolution neural network(FCN)method are used to compare and analyze the CENET model,and the accuracy evaluation is carried out.The results show that: compared with the SegNet model and the full convolution neural network(FCN)method,the CENET network model of this paper can more accurately excavate the characteristics of farmland,and achieve the target of high precision extraction of farmland from high score two image.