Research On Irrigation Area Canal Contour Extraction Based On Machine Learning

Accurate and timely access to canal contours of irrigation districts is of great significance in ensuring adequate irrigation for farmlands,improving the utilization of water resources,and alleviating agricultural water scarcity.At the meantime,it is helpful for farmers to increase crop yield,income,and water efficiency.For the currently widely used remote sensing imagery in irrigation areas,it has drawbacks for its low resolution and difficult extraction.In this study,the image of Haifeng County irrigation area in Linhe District of Inner Mongolia is taken as the object of study,and high-resolution images are acquired using drone aerial photogrammetry.High-precision orthophotos,digital elevation data,and slope data are used as data sources.The K-Means method and Support Vector Machines(SVM)are used to extract the contours of the canal system in the irrigation area,and the object-oriented classification method can be improved.Then,Hough transform methods are compared and evaluated.However,the predecessor's research based on the traditional object-oriented classification method has poor extraction accuracy of the canal system,and the automated degree of canal extraction based on the improved Hough transform method is not high.Therefore,it is a solution to adopt the machine learning method to realize the automatic extraction of canal system in the irrigation area.In this study,the extraction of canal contours in irrigation districts was studied by K-Means method and SVM method.The main research contents and results of this paper are as follows:(1)The canal contour extraction based on K-Means.The K-Means clustering method is designed and implemented to classify the canal features.The spectral features and texture features are used as an important mean to distinguish canal systems from non-canal systems.Geometric feature filters are used to optimize the extraction results.Experiments show that only the second experimental area with simple environment among three experimental areas can achieve the outline extraction effect,but the completeness is only 78.90%,and the canal extraction accuracy rate is less than 70%.(2)Extraction of canal contour based on support vector machine.The trusted training sample points were obtained by manual selection.The mixed kernel consisting of the joint unified kernel and radial basis kernel functions generated by the multi-core multi-core classification algorithm was used as the SVM classification criterion.The canal results were trained and generalized.Convex method and central axis extraction method are used to optimize the canal outline results to ensure the integrity of the extraction results and high-precision cartographic results.Experiments show that the extraction efficiency of the canal system is pretty good for three experimental regions,and the extraction accuracy rate is more than 90%,and some of the agricultural channels in each region are extracted.The extraction accuracy of second experimental area is 19.37% higher than that of the canal contour extraction method based on K-Means.The results show that the accuracy of the K-Means method in this research method can reach the canal,and the extraction accuracy is higher than the object-oriented classification method,but the completeness of the canal extraction is low in a complex environment.At the same time,the SVM method is superior to the improved Hough transform method in terms of completeness and accuracy,and the average extraction accuracy of the canal system can reach 86.34%,which shows a good performance in the extraction accuracy.