Calculating Alaska North Slope Lake Depth with Sonar, Multi-Resolution Satellite Imagery and Random Forest Algorithm

The North Slope of Alaska has an abundance of lakes, many of which are < 3 m deep, however some of which exceed 25 m. Due to the large number and striking depth ranges of lakes here, it has been extremely difficult and costly to gain accurate estimates of bathymetry, either from field visits or from satellite/aerial imagery. Machine learning provides a promising, but somewhat untested method to estimate water depths automatically from multispectral imagery. Several studies have previously used machine learning to analyze depth, but these studies were concentrated in tropical and subtropical coastal regions, where water tends to be clear and shallow. Our study represents a first look at the potential for the Random Forest Algorithm (RFA), an ensemble learning method that applies a number of decision trees to various subsets of input data, to provide accurate high-resolution predictions of lake depth. We used multispectral Top-Of-Atmosphere and Surface Reflectance data derived from Landsat-8 and WorldView imagery in combination with in-situ sonar depth points to train and validate Random Forests models. Our results suggest RFA is able to derive depths more accurately at a regional scale than linear models, particularly at moderate (4 - 8 m) depths. Furthermore, RFA requires less tuning and user input than models previously used. Although the accuracy of depth predictions declines as depths surpass 9 m, RFA can classify shallow and moderate depths, thus providing valuable information for most of the North Slope. Implementing models to estimate lake-by-lake bathymetry across northern Alaska is now possible, representing an important advance towards understanding and managing water storage and water supply, e.g. winter ice roads; hydrologic connectivity; aquatic habitat; regional energy balance and methane budget.