Research On Automatic Extraction Of Typical Water Conservancy Elements

With the rapid growth and wide application of remote sensing data,the remote sensing monitoring capability of water facilities and water bodies has been continuously improved.The contents of water information collection have been greatly increased,and the integrated monitoring system of water facilities and water resources space-ground network have initially been formed.Remote sensing technology has become the foundation of water resources information construction,supplies new tools for the construction of the water ecological civilization,guarantees the rational allocation and efficient utilization of water infrastructures,and promotes the process of water ecological civilization and water-saving society construction.However,vast lands,variety climates,different geographical environments and unbalanced regional economy take great challenges to water conservancy management and monitoring.This paper aims to address problems about the extraction of lakes,reservoirs,canals and dams by researching their spectral characteristics,morphological characteristics and spatial topological relationships among different types of water conservancy elements in remote sensing images.Based on lakes and reservoirs,we explore the relationship and the extraction methods of different water conservancy elements.Although there have been a lot of extraction methods of water,it is difficult for a single method to accurately extract large-scale water conservancy elements due to peculiar geographic conditions and the climate characteristic of our country.At present,there is little research on correlations among the geometry property of point,line and polygon water conservancy elements.And the extraction methods of these different types of elements have not been integrated together as a complete system.We try to solve these issues by proposing a preliminary system with several extraction methods on different water conservancy elements.At first,we propose a modified two-mode method to extract water by researching the spectral characteristics of lakes.To start with,the cubic spline interpolation method is used to smooth the histogram and eliminate the interference of random noise points.We introduce the concept of neighborhood to redefine the extremum.The number and location of peaks and troughs in the histogram are confirmed by comparing the image frequency values with its neighborhoods.According to the requirements of the number and relative position of peaks and troughs,the width of the neighborhood is determined by the traversal method.Then the segmentation threshold is chosen automatically from the median of the two peaks and the trough by the characteristics of water in the spectrum.To address the issue of extracting lakes and reservoirs,we propose a new lakes and reservoirs automatic extraction process.We apply the modified two-mode method to extract water from partial images by calculating thresholds in the NDWI and the NIR band;then we estimate the segmentation thresholds from NDWI thresholds of neighborhood images for the remaining images.Subsequently,a series of feature factors are used to remove rivers,shadows and ponds.And the complexity is introduced to describe the edge characteristics of water bodies,distinguishing lakes and reservoirs from other water bodies with the approximate area.The results indicate that the new method performs well,and can discover new built reservoirs and have high application value.With long spans and narrow widths,it is difficult to extract canals from river networks along with them.A new automatic extraction method based on heuristic knowledge to detect canals is proposed in this paper.The buffers are generated by the historical vector data of the canals,and are made to cut out canals with coastal areas in a certain range from the images.we utilize the modified two-mode method,OTSU and the iteration method to obtain different thresholds,respectively.The ratio of canals and clipped images is within a certain range.The appropriate threshold is automatically selected by comparing the proportion of extracted water bodies with the interval.Then,multiple feature factors are selected to remove noises.After that,the Zhang parallel thinning algorithm and the mathematical morphology method are applied to extract canal skeleton lines and connect breakpoints.At last,a length threshold is set to remove branches in skeleton lines according to the characteristics of buffer and the morphological characteristics of rivers.The scale of dams and their other features keep changing due to the influence of the surrounding environment,which makes it difficult to extract the dams.To address this problem,we propose an automatic method combining the lakes and reservoirs extraction process and the deep-learning algorithms.The Faster R-CNN algorithm is used to train the neural network feature model of the dams.Then we extract water bodies which may contain dams.The trained model is utilized to identify the dam near or in the water body in the images.The results indicate that the new method has high accuracy,and the accuracy is higher than that using the deep-learning algorithms alone.