Extraction Method Of GF-2 Remote Sensing Image Based On Convolutional Neural Network

Accurately obtaining the distribution information of cultivated land,residential land,forest land,roads and waters is the main means of obtaining land use information.The high-resolution remote sensing image is used as the data source,and the high-resolution remote sensing image is segmented and extracted by the deep learning model,which can obtain high-precision land use classification related information,which plays an important role in rational development of land resources and promotion of precision agriculture.In the high-resolution remote sensing image,the features of various types of features are complex and diverse.Applying the neural network model to the extraction of remote sensing images can increase the segmentation accuracy and improve the processing speed compared with the traditional extraction method.Deeplab and RefineNet networks are effective in natural image recognition segmentation,but they are not suitable for processing high-resolution 2 remote sensing images.The main reason is that deep convolution ignores the details and there is a multi-layer pooling layer.The pooling discards part of the location information while considering the background space information.Another reason is that the high-resolution 2 remote sensing images have different characteristics compared with the natural images.The convolutional neural network is not suitable for directly extracting high-resolution 2 remote sensing images.This paper analyzes the performance characteristics of cultivated land,woodland,waters,roads and residential areas on high-resolution 2 remote sensing images.Based on the Deeplab and RefineNet models,the design analysis is based on the specific characteristics of each type of features,combined with the training results.Finally,the network structure for the five types of ground objects is obtained separately,and the images are separately classified by the two network models.main tasks as follows:1.Data set production.The integrated high-resolution 2nd remote sensing image is very large.If the network training is directly input,the memory card will have memory overflow problem.This article uses ENVI 5.3 to directly integrate the high-resolution 2 remote sensing image,and then use the annotation software to mark the remote sensing image.The code written in Python crops the PNG format image with a resolution of 960*960,and marks its attribute as a value from 0 to N,that is,the pixel attribute values of different features are marked as 1,2,3,...,N,the pixel points of other objects are marked as 0.2.The image features of high-resolution remote sensing images of various objects are mainly characterized by spectrum,space,texture and shape.This paper summarizes the characteristics of cultivated land,woodland,waters,roads and residential areas in the high-resolution 2 remote sensing image,and theoretically summarizes the various features to lay the foundation for the subsequent model design.For the image of bare land and wheat distribution,this paper proposes an extraction method based on ECLDeeplab(Extraction Cultivated Land Deeplab).Since the cultivated land in the central part of Shandong Province in December was mainly wheat and bare land,the model mainly extracted winter wheat and bare land,and the overall extraction accuracy rate was 88.3%.4.According to the characteristics of each type of features,the network structure is constructed based on the RefineNet model,and the IM-RefineNet(Improved Model RefineNet)model structure is designed.This method can retain the spatial position information of high score 2 remote sensing images.The final experimental results show that the proposed method improves the efficiency of extracting features from high-resolution 2 remote sensing images,and the overall extraction accuracy reaches 93%.