Study On Crop Rapid Classification Based On High Resolution Remote Sensing Images

Accurate crop classification is an important data source for agricultural monitoring,food security assessment and national macro-control.It is an efficient method to realize crop classification by high-resolution remote sensing image classification.A large number of algorithms have been proposed for remote sensing image classification,such as maximum likelihood method based on pixel,support vector machine based on object-oriented and decision tree classification based on expert knowledge.However,due to the arable land,their own conditions caused by different crop growth status difference,causes on the image the same crops have different spectral differences,existing algorithms for some minor crops recognition effect is not good,fine crop classification at present stage still need to use artificial visual interpretation,it is need to improve the existing algorithms learn the characteristics of crops at a deeper level,In order to improve the accuracy of automatic classification and recognition.Existing deep learning mainly focuses on bulk crops such as corn and wheat,and pays little attention to small-scale crops.However,small-scale crops in some regions are the main economic source of farmers.Accurate planting area,region and yield prediction of crops can help farmers accurately and benefit them.Therefore,improving the classification accuracy of remote sensing image on small population crops also plays a crucial role in the agricultural field.In this thesis,some towns in Luozhuang District of Linyi City and Lanling County of Shandong Province were selected as research areas.After field investigation,the cross test areas of Mopan Town,Shenshan Town and Zhudun Town were determined as sample refinement areas,and sample sets of 256*256*3 were made,and the number of samples was enhanced.In this thesis,samples are selected in a local concentration with diversified sample forms to make large area prediction,and then the method of sample set selection is verified in a large area.The sample model is based on the U-Net model,which is a mini U-Net model that reduces the number of channels and optimizes the hyperparameters.The optimal combination is selected by formulating experimental parameter groups.By comparing and analyzing the classification results of the three object-oriented methods and mini U-Net model classification methods,the model generalization ability of the three test areas selected by Bienzhuang subdistrict,Luzuo Town and Changcheng Town is verified.The results show that the mini U-Net model has the advantages of smaller misscore and missing score in the cross test area,but there is no great difference with the three object-oriented methods.In the classification results of the three verification test areas,the mini U-Net method has the advantage of equal accuracy with the cross test area,and is obviously superior to the three object-oriented methods.