Method Research On Extracting Spatial Distribution Information Of Winter Wheat Using Convolutional Neural Network

Winter wheat is one of the main grain crops in China.Accurate estimation of spatial distribution information of winter wheat is important in yield estimation,production management and food policy adjustment.Remote sensing image interpretation is the main technical means to obtain the spatial distribution information of winter wheat at present with the advantages of wide coverage,short detection cycle and abundant information.GF 2 image provides reliable data guarantee for timely and accurate acquisition of spatial distribution information of winter wheat with the spatial resolution of 1 meter.It is a hot issue of current researchers.to obtain the spatial distribution information of winter wheat by applying convolutional neural network in remote sensing image interpretation.However,the segmentation accuracy is not too high by the existing convolutional neural network model is directly used to conduct semantic segmentation of winter wheat on GF 2 image.The main reason is that there are some defects in the encoder and classifier of convolutional neural network:(1)The regression simulation used by encoder is relatively simple,which is not conducive to the formation of abstract high-level features;(2)The classifier only utilizes the information of the maximum probability value in the category classification judgment and ignores the influence of the difference value between the probabilities on the pixel category classification judgment.In view of the above problem,selected the chapter grave area of jinan of shandong province as the research area,using the high score 2 Remote Sensing images and ground survey data as the data source,the high accuracy of Winter Wheat Remote Sensing Segmentation Model(WWRSSM),obtained the finer spatial distribution information of Winter Wheat in this paper.The main research contents of this paper are as follows:1.The classic SegNet,DeepLab and RefineNet models were selected to extract the spatial distribution information of winter wheat from the GF-2 remote sensing image,and the extraction results of the three models were statistically analyzed in this paper.2.Design and implement of the high-precision remote sensing image segmentation model for winter wheat(WWRSSM).Aiming at the problems of SegNet,DeepLab and RefineNet models in extracting spatial distribution information of winter wheat,a network structure composed of feature extractor,encoder and classifier was designed.RefineNet convolution structure was used as a feature extractor to extract the features of pixels in winterwheat planting area;The depth confidence network(DBN)with strong fitting ability was selected to construct the encoder,and the feature vectors obtained by the feature extractor were encoded to increase the nonlinear capability of the model;The classifier is composed of softmax-ex and the second-level sub-classifier of the maximum posterior probability(MAP)layer.Based on the classification of softmax-ex,confidence coefficient is calculated by probability difference.The original classification result is reserved for pixels with high confidence coefficient,and for pixels with low confidence coefficient the category probability vector is input into the MAP sub-classifier.The MAP layer further identifies the categories of low confidence coefficient pixels based on the bayesian principle.A large number of data samples were used to train the model,and finally the winter wheat segmentation results were obtained.3.Analyze the result of the experiment.In order to verify the WWRSSM model,SegNet,DeepLab,RefineNet and the models coupled with MAP respectively were selected as the comparison model,and the same data set was used for training and testing.According to the experimental results,the extraction accuracy of WWRSSM model is 94.8%,higher than that of the six comparison models.Finally,the advantages and theoretical basis of WWRSSM model are summarized.WWRSSM model constructed in this paper makes up for the deficiency of traditional convolutional neural network in extracting spatial distribution information of winter wheat.Moreover,it is of certain significance to improve the precision and automation level of spatial distribution mapping of large-scale winter wheat,and can also provide certain technical reference for extraction of spatial distribution information of crops and area statistics.