Research On The Method Of UAV Remote Sensing Terrace Recognition Based On Semantic Segmentation

Terraces are the fastest and most effective soil and water conservation projects on hillside farmland.According to the planning of the "Thirteenth Five-Year Plan" for the comprehensive management of soil and water loss on slope farmlands nationwide,large-scale slope-to-ladder projects are about to be completed.Timely and accurate information on terrace distribution is very important for soil and water conservation monitoring and evaluation.At present,the research on the method of terrace recognition based on high-resolution UAV remote sensing images is still at the object-oriented level.The feature learning and recognition classification of terraces rely on manual assistance.In order to solve the problem that the characteristics of terraces cannot be automatically and deeply learned in the study of remote sensing terrace recognition,this paper using deep learning technology,the semantic segmentation method based on Fully Convolutional Networks(FCN)and Conditional Random Field(CRF)is studied to complete the identification of terraces.This paper mainly completes the work as follows:(1)The scheme of generating semantic segmentation data set of terraces.Since the existing semantic segmentation data sets lacks the terrace block label samples,this paper first uses drones to collect the original data,combined with geometric correction,image mosaic and other data preprocessing methods to obtain orthophotos of terraces,after Open CV cropping,using Labelme to label the sample set based on visual interpretation,finally,the data enhancement technology is used to generate the pixel-level terrace block semantic segmentation data set,which lays the foundation for the follow-up terrace recognition research.(2)Research on FCN-based terrace recognition method.In order to extract the deep-level features of the terrace blocks and then identify the blocks,the VGG-19 network model is used as the basis to design and implement the FCN-32 s,FCN-16 s,FCN-8s three FCN terrace recognition methods through multi-scale feature fusion.Three different types of terrace test samples were used to test the recognition ability of the three FCN methods,and the accuracy and comparison of the recognition results of the test samples were tested.The results show that the FCN-8s terrace recognition method is better than the other two FCN methods.(3)Research on CRF-based terrace recognition results optimization method.In order to reduce the recognition error of FCN-8s method recognition results for unclear terrace blocks,and further improve the recognition accuracy,this paper researches the optimization scheme of terrace block recognition results based on CRF,uses FCN-8s terrace block recognition results as the first-order potential function,and uses Gaussian kernel linear combination to define pairwise potential functions to construct 4-CRF,8-CRF and full connection CRF(Dense CRF)optimization method,the method results are calculated by the average approximate field solution method.The accuracy evaluation and comparative analysis of the recognition results of the three CRF optimization methods on the test area show that the Dense CRF method has achieved the best optimization effect,and the overall accuracy,F1 score and Kappa coefficient of the recognition on the ridge terraces,dense horizontal terraces and irregular terrace images were improved by an average of 5.41%,7.3% and 10.08% compared with before optimization,and finally reached 86.85%,87.28%,80.41%,achieving fine semantic segmentation results.The research in this paper promotes the research and application of deep learning technology in the field of remote sensing terrace recognition.