Research On Extraction Method Of Cotton Planting Area Based On Satellite Remote Sensing

Cotton is an important economic crop in my country.Xinjiang’s cotton planting area accounts for more than 80% of the country’s total,and its output accounts for more than 85%.Quickly and accurately extracting cotton planting areas has very important strategic significance for maintaining social stability and sustainable economic development.However,the large cotton planting area and variety in Xinjiang region of my country,coupled with the complicated farmland planting conditions,intercropping,desertion and natural disasters,make it difficult to obtain specific cotton planting areas,and traditional methods have been unable to meet our actual needs.Deep learning semantic segmentation technology is a relatively cutting-edge research direction in the field of machine learning,and in recent years,with the development of remote sensing image detection technology,large-scale,multi-temporal and high-resolution remote sensing data are used for our deep learning research.Provides massive data sources.This thesis takes this as a research entry point and applies deep learning technology to the extraction of cotton planting areas.It is hoped that this technology can improve the extraction accuracy of cotton planting areas in cotton fields and meet the needs of national agricultural development.Based on this,this thesis carried out the following research work.(1)Extraction of cotton planting area based on U-Net neural network model under complex background.Based on the remote sensing data of Sentinel-2 satellite,the traditional agricultural remote sensing detection technology has low efficiency,complex types of farmland objects,data imbalance between classes,and the model is prone to misclassification and misclassification,resulting in poor extraction accuracy in cotton planting areas.high question.The U-Net model was used to extract the cotton planting area under the complex background of the 31 st Regiment and 33 rd Regiment in Yuli County,Xinjiang.In view of the problem of the unbalanced phenomenon of remote sensing satellite image categories,the results of cotton planting area extraction were low.Resampling and changing the loss function to adjust the sample imbalance of the dataset and improve the accuracy of cotton planting area extraction,it is found that the method in this thesis can improve the extraction accuracy of cotton planting area by more than 6 percentage points compared with other methods,which can effectively change the model.Misclassification and misclassification.(2)Extraction of cotton planting area based on optimal time-phase combination of satellite remote sensing images.Based on the deep learning method,the remote sensing satellite images from April to October in the cotton growth cycle were used to explore the impact of the data of a single phase and its combination of remote sensing images on the extraction of cotton planting areas.The deeplabv3+ model is used to extract the cotton planting areas of the remote sensing images in each phase of the study area,and the piecewise function evaluation model is used to quickly compare the extraction results.According to the extraction results,single-phase remote sensing images suitable for cotton planting and extraction are arranged.,added to the Deeplabv3+ model in turn,and the optimal time-phase combination of the extraction results of cotton planting areas from remote sensing images under this model is June,July,August and October,which can significantly reduce the difference with non-cotton planting areas.In the case of misclassification,the edge of the cotton planting area can be recovered better,and the result can provide a reference for the extraction of cotton planting area when the remote sensing data is insufficient.(3)Extraction of cotton planting area based on ShuffleNet neural network from satellite remote sensing images.For the large-scale semantic segmentation network,the network depth is deep,the structure is complex,the network training time is long,and the number of parameters is large.It takes a lot of time to extract the cotton planting area,resulting in insufficient immediacy of extraction.In this section,based on the optimal time-phase combination of cotton extraction,through several deep learning networks,the results are obtained by comparing the training time and parameters of the lightweight network.Compared with U-Net,Deeplabv3+ has greater advantages in training time and network parameters,but its network extraction accuracy differs by about 10 percentage points from Deeplabv3+ network.For the problem that the extraction result of this network is low,the original band of Sentinel 2 is used and the red edge band,which has special advantages in plant detection,is used.The results show that the ShuffleNet network after adding the red edge band can effectively improve the extraction accuracy of cotton planting areas.,the network is small in model parameters and training time,which provides a reference for cotton extraction using Sentinel-2 remote sensing images.