Study On Terraced Field Extraction With A Deep Learning Method Combined With Both Spectral And Topographic Features

Being one of the main ways of arid farming development in the Loess Plateau of China,terrace field plays dual roles in both agricultural production and soil and water conservation,therefore it is very important in agricultural and rural development.As a part of the basic farmland in the Loess Plateau,terrace field have always been the key content of national geographic state monitoring.Using remote sensing technology to monitor the dynamic changes of terrace is of great significance for the construction and protection of terrace field.Unlike normal croplands extraction,the terrace field was widely distributed in the complicated topography of the Loess Plateau,it’s necessary to avoid the mix with the farmland in the basin area.At the same time,terraced fields are greatly affected by different states of crop growth or returning farmland to forest and grassland.Traditional pixel-based analysis methods can only use spectral data,which are not easy to extract from middle and high-resolution remote sensing images because of the influence of " same objects but different spectrum " and " same spectrum foreign matter ".Deep learning semantic segmentation method developed in recent years can comprehensively use the multi-level semantic features contained in the image,which opens up a new way for remote sensing information extraction.In view of this,this paper proposes a deep learning extraction method for terraced fields in the Loess Plateau based on deep learning,combined with the optimal combination of remote sensing image spectrum and terrain feature parameters,and has achieved good results in the study area of Zuli River Basin in Gansu Province.The main content of this article is as follows:（1）A set of sample datasets for terrace semantics segmentation was constructed.Based on Sentinel-2 Spectrum image and medium resolution DEM data,This paper combines the slope,aspect,curvature derived from elevation information and visible and near infrared bands to generate different input data combinations,and then we expand the sample dataset through data enhancement technology.In this process,two different image clipping schemes are applied in the training and testing phases respectively.Finally,5919 sets of image blocks are generated,which lays a foundation for subsequent model training.（2）Terrace identification based on improved U-Net semantics segmentation network.In response to the complex topographic conditions and vegetation cover types at the Loess Plateau terrace,an improved deep learning model for combined wave spectrum optimization and topographic analysis is proposed in this paper.Among these,spectrum spectroscopy is used to retrieve terrace texture and vegetation.In addition,topographic analysis resolves confusion with non-terrace croplands.In order to unite the different features,this paper uses an improved U-Net model that adds a pyramid squeeze attention module with a multi-level feature aggregation upsampling module.The encoder phase enhances the feature extraction capability with multiscale spatial information,while different layers of features are utilized in the decoder to provide auxiliary information for the process of recovering spatial resolution.Experiments have shown that the improved model can increase the integrity of the terrace information extraction results and obtain smoother and consistent with practical boundaries.Moreover,the extraction results of 66% and 91% were obtained in Io U and overall accuracies,which were better than the commonly used models such as U-Net,FCN,Deep Lab V3+.In addition,pure spectral images or topographic features have a large disadvantage in the application of terrace extraction,with more misclassifications in non-terrace areas in valley croplands or hill slopes.However,the data that combined spectral and topographic characteristics were obviously improved,especially the sixchannel data combination consisting of red,green,blue,slope,slope,curvature achieved the best extraction result.（3）The optimal training model is used to extract the terrace information of Zuli River Basin,the extraction effect of the model under different terrain and vegetation environment is quantitatively analyzed,and further model migration experiment is carried out.The results show that the method in this paper has a high spatial generalization ability and shows a good extraction effect for terraces located in Loess ridges,loess hill,stubble fields and cultivated land.However,it is difficult to maintain a good stability in the area of returning farmland to forest and grass,most of the misclassification and omission classification are due to this.