| Lakes are important information carriers to reveal global climate change,and are extremely sensitive to regional and global climate fluctuations.There are many lakes on the Tibetan Plateau(TP).Obtaining high-precision lake water depth and accurately estimating lake water level,water storage and its changes are of great significance to understand the dynamic changes of lakes and regional climate change in the TP.Based on Sentinel-2,GF-1 and Landsat-8 multispectral images and measured water depth data of QiXiang Co,this paper constructed a water depth inversion model,evaluated the accuracy and made inversion mapping.The water depth inversion model includes three statistical regression models(two band log ratio model-Stumpf model,multi band log linear regression model-Lyzenga model,and multi factor combination linear regression model-MLR model)and three machine learning models(BP neural network model-BPNN model,support vector machine model-SVM model,and random forest model-RF model).Using the QiXiang Co water depth retrieved by the optimal model,combined with the boundary data obtained by Google Earth Engine(GEE)platform from 1998-2018,the change of the water level and water storage were estimated and compared with them obtained by traditional spatial interpolation and Shuttle Radar Topography Mission(SRTM DEM)method.The results show that:(1)Logarithmic transformation and combination of spectral band reflectance of the three multispectral images can improve the correlation with water depth.The three images have good application prospects in lake water depth inversion.In this study,Sentinel-2 image has the highest accuracy.(2)In water depth prediction,multi factor as model input can improve the accuracy of the model.Compared with Stumpf model and Lyzenga model,the MLR model proposed by us combined with ratio factor and logarithm factor can effectively improve the accuracy of water depth inversion in three multispectral images.In the range of 028 m water depth,the mean relative percentage error(MAPE)of the verification set is reduced by 2.4%-15.2%,and the mean absolute error(MAE)is reduced by 0.15-1.2 m.The accuracy of machine learning model is higher than that of statistical regression model.In the three multispectral remote sensing images,the determination coefficient(R2)of machine learning model verification set is more than 0.63,MAE is as low as 1.64 m,and MAPE is 14.3%-27.8%.Compared with statistical regression model,MAE is reduced by 0.08-2.25 m.In addition,in the machine learning model,the random forest model has the best stability and prediction accuracy,and the Mae and MAPE of the verification set are the smallest,only 1.64 m and 14.3%,which can meet the needs of high-precision lake bathymetric mapping.(3)In the water depth inversion mapping,the inversion mapping of different models can reflect the overall water depth distribution trend of QiXiang Co.The inversion mapping effect of machine learning model is more accurate and real than the traditional empirical model.Among them,the detailed changes of the mapping effect of random forest model are the richest and closest to the real water depth distribution,In the area without measured value distribution,the water depth obtained by spatial interpolation method is more accurate than that obtained by spatial interpolation method,Remote sensing water depth inversion can be used to quickly obtain the water depth distribution of lakes.(4)Based on the optimal model inversion mapping results of Sentinel-2 image,combined with boundary data of QiXiang Co from 1998 to 2018,it is estimated that the change of water storage is 1.770 km3,water level is about 9.32 m,and the average annual change is about 0.47 m,which is similar to the result estimated by SRTM method,while the error of water storage change estimated by spatial interpolation method is relatively large.In addition,it is analyzed that the main reason for the change of QiXiang Co is the increase of regional temperature,which leads to the accelerated melting of frozen soil and glaciers in the water source area in the upper reaches of the basin,and the increase of water source supplied by lake runoff. |
