| Water resources are the most indispensable condition for human production and life.Due to our country's geographical location and uneven distribution of water systems,the density of river networks in the south is greater than that in the north.In addition to natural disasters such as lack of water resources,uneven distribution of precipitation,and floods in some areas,the water environment is very fragile.From the perspective of national development and the protection of people's livelihood,protecting water resources is of great significance.To grasp the domestic water system situation from a macro perspective,to use remote sensing to extract water bodies and monitor them in real time is a major measure to protect water resources.After development,the methods for extracting image water bodies by remote sensing have been quite mature,mainly including single-band threshold method,supervised classification method,spectrum relationship method and water body index method.These four methods basically meet the accuracy requirements of waer body extraction,but the accuracy had also reached saturation.It is difficult to continue to improve the accuracy by remote sensing alone.It is necessary to find other methods to achieve higher accuracy.Neural network is a research hotspot in recent years.The neural network simulates the working mode of human neuron cells.A large number of nodes are connected to each other.Each node can perform an operation,and the place where two nodes are connected will have a signal weight value.The neural network continuously learns and adjusts its own parameters through node calculation layer by layer,so as to fulfill various needs in the research process well.This paper selects the BP neural network of error back propagation for research.The network can reach a higher level of accuracy after a certain number of learning.In addition,the number of hidden layers and the number of nodes can be set arbitrarily,which can be regarded as a universal model.Set the input nodes of BP neural network to nine,including first six bands of image data in huainan area,NDVI,NDWI index and texture image of gray level co-occurrence matrix calculated from the image.At the same time,in view of the multi-directionality of the gray-level co-occurrence matrix,the local binary pattern was introduced to improve it,so as to solve the problem and reduce the amount of calculation.The random forest model was introduced and established to learn sample data and identify water bodies in the imagery of huainan area.The accuracy of BP neural network,random forest and maximum likelihood method was compared.After adjusting the format of the sample data,training the BP neural network,and continuously modifying the parameters,the accuracy of the trained BP neural network was tested to be 87.13%.Finally,the BP neural network was used to classify the images of the study area in Huainan City.The overall accuracy of the classification results was 0.9609,the kappa coefficient was 0.8940,and the accuracy was increased by 4.7%and 15.38%respectively compared with the supervised classification,compared with the random forest.Increased by 0.56%and 2.01%.The study area changed into taihu,and identified water bodies with the same methods.accuracy has the same performance in both huainan and taihu.Finially,flooded area in chaohu in the summer of 2020 was extracted.The following conclusions were reached:1)The classification accuracy of the BP neural network that combines pixel and texture features is higher than classification with a single feature.2)The recognition accuracy of BP neural network is higher than that of random forest.3)Compared with a single image band,the BP neural network using multiple image bands has higher accuracy in water recognition.4)The texture image,NDVI and NDWI are added as auxiliary data to the BP neural network model,which can reduce the impact of buildings,and the recognition effect at the water boundary is also improved.Figure30 table 6 reference 15... |
PDF Full Text Request