Shallow Water Depth Inversion Based on Data Mining Models

This thesis focuses on applying machine-learning algorithms on water depth inversion from remote sensing images, with a case study in Michigan lake area. The goal is to assess the use of the public available Landsat images on shallow water depth inversion. Firstly, ICESAT elevation data were used to determine the absolute water surface elevation. Airborne bathymetry Lidar data provide systematic measure of water bottom elevation. Subtracting water bottom elevation from water surface elevation will result in water depth. Water depth is associated with reflectance recorded as DN value in Landsat images. Water depth inversion was tested on ANN models, SVM models with four different kernel functions and regression tree model that exploit the correlation between water depth and image band ratios. The result showed that the RMSE (root-mean-square error) of all models are smaller than 1.5 meters and the R2 of them are greater than 0.81. The conclusion is Landsat images can be used to measure water depth in shallow area of the lakes. Potentially, water volume change of the Great Lakes can be monitored by using the procedure explored in this research