Intelligent Extraction Of Terraced Fields From Remote Sensing Images Based On Deep Learning

In the construction planning,acceptance statistics and benefit evaluation of terraced fields,timely,accurate and objective extraction of the area,type and geographical spatial distribution of terraced fields is the core link to achieve dynamic monitoring of soil erosion and accurate agricultural management.Aiming at the problems of low accuracy and slow efficiency in the intelligent extraction of geographical elements of terraced fields from remote sensing images,this paper takes typical terraced fields in Tonglu County,Hangzhou City,Zhejiang Province,Qiaojia County,And Ludian County,Zhaotong City,Yunnan Province as examples.In order to make full use of the semantic information of visible light and digital elevation model of remote sensing image,improve the segmentation accuracy of complex overlapping ground objects and perfect the precise segmentation of boundary,the convolution neural network and Transformer attention mechanism network were used to extract the geographical elements of terraced fields.The main research content of this paper is summarized as follows:(1)A network model for the extraction and segmentation of terracing geographical elements is adopted,a deep neural network model combining image and elevation information.The model mainly includes convolutional neural networks DeeplabV3+and W-NET based on ResNet50 and ResNet101 as skeletons,and SegFormer designed based on Transformer.In the case of large scale and high resolution,the solutions of over-fitting and under-fitting problems formed due to different network layers and parameter Settings are deeply discussed and studied.The accuracy difference between Transformer based network and convolutional neural network in the extraction task of terraced field geographical elements is systematically compared.(2)The basic data sets are acquired by artificial visual interpretation of remote sensing satellite images,so as to solve the problems of small area and small quantity of terraced fields and difficult data set acquisition in the pre-processing stage.The GIS software is used to vectorize and transform the sample labels,data enhancement is used to increase the number of samples,and automatic data set generation method is used to balance the proportion of positive and negative samples and the area size of the data set.In the postprocessing stage,conditional random fields were used for post-processing to smooth the edge of the terrace,fill in the omission of small map spots in large terraced fields,and form a good segmentation effect.(3)Realize vectorization processing of agricultural remote sensing image terraced fields extraction and segmentation.After segmentation of different attributes of geographical elements,geographical attributes can be quickly and automatically extracted,and ultimately achieve the purpose of mapping,visualization,can also further achieve the calculation of area,spatial distribution change,etc.The final experimental results show that the deep learning technology proposed in this paper has achieved a good effect on terrace geographical elements extraction.Among them,the accuracy evaluation index of W-NET convolutional neural network is the best,the overall classification accuracy is 91.83%,Kappa coefficient is 0.7834,MIOU value is 80.04%.However,the segmentation effect of combining digital elevation model and visible remote sensing image is not ideal due to the large difference in resolution between remote sensing image and digital elevation model,low resolution of elevation digital model and registration error between them.