Extraction Of Irrigation Networks In Irrigation Area Of UAV Orthophotos Based On Fully Convolutional Networks

With the continuous deepening of agricultural water-saving theory,agricultural watersaving technology is tending to be controllable and accurate.As an important water delivery and water distribution facility of farmland in irrigation area,irrigation networks is the key to achieving refined management of irrigation area.Therefore,the rapid and accurate access to the distribution information of irrigation networks in irrigation area is of great significance for the scientific allocation of regional agricultural water resources and the improvement of water resources utilization rate.Aiming at the issues of low accuracy and lower degree of automation in the study of irrigation networks extraction in irrigation area at present,this paper takes the deep learning algorithm into the study of irrigation networks extraction in irrigation area by using the high-resolution orthophotos of irrigation networks contour in Inner Mongolia Autonomous region collected by UAV as the research object.The extraction of the irrigation networks contour in the irrigation area is accomplished by using the semantic segmentation algorithm based on the fully convolutional networks(FCN).The main research contents and conclusions are as follows:(1)Datasets generation scheme of irrigation networks contour semantic segmentationA complete generation scheme of the semantic segmentation data set of the irrigation networks contour is designed.By using the UAV to complete acquisition of the image information,the high-resolution orthophotos of the irrigation networks in the irrigation area collected by UAV are used as the data sets,and the datasets are labeled by visual interpretation,and in the Open CV environment,the automatic segmentation and data enhancement of the datasets are realized,and the semantic segmentation datasets of the irrigation networks contour is formed.(2)Design and implementation scheme based on FCN irrigation networks extraction methodA complete scheme of contour extraction of FCN irrigation networks is designed,including the design and realization scheme of the contour extraction method of FCN irrigation networks and the automatic splicing scheme of the extraction results of the irrigation networks.Firstly,based on the VGG-19 network,four kinds of irrigation networks contour extraction models of FCN-32 s,FCN-16 s,FCN-8s and FCN-4s are constructed by means of multi-scale feature fusion in Tensorflow depth learning platform,and the GPU is used to accelerate the experiment.Through four kinds of FCN methods,the training sets and verification sets after data enhancement are trained 50,000 times,which effectively avoids the phenomenon of over-fitting.The test sets was tested by four FCN methods respectively,and the test results were automatically stitched and evaluated by precision.The FCN-8s method with the best performance of the model is selected and compared with the Support Vector Machines(SVM)method and the Hough transform method for further comparison and analysis.The results show that the FNS-8s method has the best extraction effect of the irrigation networks contour.The average extraction accuracy of the test area for different complexity is 89.45%,the average extraction accuracy is 4.57% higher than the SVM method,and 8.51% higher than the Hough transform method.The method has good generalization and robustness.