Reservoir Capacity Calculation Method Based On Machine Learning And Satellite Remote Sensing

Reservoir projects play an important role in regulating regional surface runoff and terrestrial water circulation,and are essential for hydropower generation,river navigation,and flood storage and regulation.The correct and rapid calculation of the reservoir capacity is particularly important for the reservoir to play a normal flood regulation role and ensure the safety of the reservoir area.In recent years,the development and application of satellite remote sensing technology provides a new method for reservoir capacity calculation.This study proposes a new method of reservoir storage calculation based on machine learning and satellite remote sensing,which can quickly determine the reservoir storage capacity,and has achieved good results through experimental verification in Meizhou Reservoir.This study uses the unmanned ship sounding system to obtain the original water depth data as the data source of the water depth inversion model constructed by the machine learning algorithm.The fitting effects of different algorithms on water depth data are discussed in detail.The inversion accuracy of different algorithms in different water depth intervals is compared and analyzed,and the influencing factors of the errors are analyzed.Finally,select the optimal water depth inversion algorithm to calculate the reservoir capacity.The main tasks of the paper are as follows:1.Aiming at the problem that the traditional water depth inversion model data source is calculated with low accuracy based on charts and tide survey data.In this paper,the method of collecting regional water depth data of unmanned ships equipped with single wave speed sounder and global positioning system is used to improve.The depth data source of the water depth inversion model is obtained by eliminating pre-processing measures such as abnormal points in the original water depth data,transforming the coordinate system and random sampling.Download the mid-scale and high-resolution Sentinel?2 satellite image and preprocess it by atmospheric correction and resampling as the image data source of the model.2.Select a supervised learning regression algorithm to construct a water depth inversion model based on the correspondence between water depth data and remote sensing images in the study area and the required target values.Finally,a random forest algorithm,an extreme gradient boosting algorithm,and a support vector machine algorithm are selected to construct a water depth inversion model.Extract 10% of the data from all the collected water depth data for training and testing of the model,and compare and analyze the test results to select the best model.Research shows that in the process of machine learning algorithm tuning,the inversion accuracy obtained by using the default parameters of the model is low.By changing different parameter values,the model accuracy is improved,but the random forest model is not very sensitive to the adjustment of parameters.The support vector machine model has greater fluctuations and higher accuracy.The coefficient of determination has been increased from 0.02 to 0.78.A comparative analysis of the three machine learning models shows that the model with the highest degree of fit is a random forest model.Its determination coefficient R2 is 0.80,and the average absolute error is 1.51 m.Second is the extreme gradient Booting model,and last is the support vector machine model.Through segment analysis of different water depth intervals(5m intervals),The best fitting accuracy of the random forest model in the 20-25 m interval is 1.21 m,while the extreme gradient lifting model performs best in the 0-5m water depth interval,and the error is only 0.63 m.Support vector machine performance in general,Through the input of reducing the sample size,it is found that the random forest model shows higher stability and fitting accuracy to the increase and decrease of the sample size,while the support vector machine model has relatively large fluctuations.3.Comprehensively compare the advantages and disadvantages of previous people in the extraction of water area,combine the features of the study area and the band characteristics of remote sensing images,and select the appropriate water extraction method(normalized difference water index method),After further analysis of the water body image obtained by automatic extraction,the area of the water body image extracted is calculated.By comparing and analyzing various storage capacity calculation methods such as grid method,contour line method,triangle network method,cross-section method,analysis method,etc.,select a method suitable for storage capacity calculation in this area.The water depth data obtained from the machine learning water depth inversion model is combined with the storage capacity calculation formula set in the program to automatically calculate the reservoir storage capacity.The calculated of Meizhou Reservoir storage capacity is 48714763.57 cubic meters,while the actual reservoir storage capacity is 48639372.38 cubic meters and the cause of the error is analyzed in detail.