Abstract: To investigate the application effectiveness of the mixture density network deep learning algorithm in nearshore water depth inversion,a portion of Haizhou Bay in northern Jiangsu province was selected as the study area.Landsat 8 satellite imagery and measured water depth data were utilized,and the accuracy of the results was compared and evaluated against those of the support vector machine model,random forest model,and fully connected neural network model.The results indicate that,overall,all four models exhibit commendable accuracy.The mixed density network model exhibits higher inversion accuracy in the experimental area compared to the other three models,with a coefficient of determination for water depth inversion of 0.86,a mean absolute error of 0.85m,and a root mean square error of 1.36m.The inversion results align more closely with the actual water depth.

Key words: water depth inversion, mixture density network, deep learning, multispectral remote sensing, Haizhou Bay