Search

The weakness of global metrics for assessing the accuracy of satellite-derived bathymetry (SDB) is demonstrated and the magnitude of local accuracy variability quantified. SDB was developed for three small areas (500 m-by-1500 m) near Key West Florida (United States) that varied in depth and geomorphometric complexity. Data employed were 2016 and 2019 airborne lidar and three 2016 and four 2019 Landsat images. The modelling methodology was a widely used quasi-empirical linear regression-based approach. Local accuracy of SDB depth change – which was expected to be pronounced between 2016 and 2019 due to the occurrence of Category 5 Hurricane Irma in 2017 – was also calculated and examined. The global correlation (R2) between lidar-based depth and SDB for 2016 and 2019 varied from about 0.1 to 0.9 which produced relative (to depth) root mean square errors (RMSEs) between 5% and 25%. No image was consistently best for all three areas nor for either 2016 and 2019. SDB-based depth change was poorly correlated to the reference lidar-based depth change for all three areas. Spatial depiction of SDB accuracy showed considerable variability even within the three relatively small areas employed for 2016 and 2019 SDB depth estimates and 2016–2019 SDB depth change estimates. [ABSTRACT FROM AUTHOR]

Color is an important cue in object recognition and classification problems. In underwater imagery, colors undergo strong distortion due to light propagation through an absorbing and scattering medium. Distortions depend on a number of complex phenomena, the most important being wavelength-dependent absorption and the sensitivity of sensors in trichromatic cameras. It has been shown previously that unique reconstruction in this case is not possible—at least for a simplified image formation model. In this paper, the authors use numerical simulations to demonstrate that this statement also holds for the underwater image formation model that is currently the most sophisticated. [ABSTRACT FROM AUTHOR]

Most work to date on satellite-derived bathymetry (SDB) depth change estimates water depth at individual times t1 and t2 using two separate models and then differences the model estimates. An alternative approach is explored in this study: a multi-temporal Sentinel-2 image is created by "stacking" the bands of the times t1 and t2 images, geographically coincident reference data for times t1 and t2 allow for "true" depth change to be calculated for the pixels of the multi-temporal image, and this information is used to fit a single model that estimates depth change directly rather than indirectly as in the model-differencing approach. The multi-temporal image approach reduced the depth change RMSE by about 30%. The machine learning modelling method (categorical boosting) outperformed linear regression. Overfitting of models was limited even for the CatBoost models having the maximum number of variables examined. The visible Sentinel-2 spectral bands contributed most to the model predictions. Though the multi-temporal stacked image approach produced clearly superior depth change estimates compared to the conventional approach, it is limited only to those areas for which geographically coincident multi-temporal reference/"true" depth data exist. [ABSTRACT FROM AUTHOR]

The goal of this work was to evaluate if routinely collected but seldom used airborne lidar metadata – 'point attribute data' (PAD) – analyzed using machine learning/artificial intelligence can improve extraction of shallow-water (less than 20 m) bathymetry from lidar point clouds. Extreme gradient boosting (XGB) models relating PAD to an existing bathymetry/not bathymetry classification were fitted and evaluated for four areas near the Florida Keys. The PAD examined include 'pulse specific' information such as the return intensity and PAD describing flight path consistency. The R2 values for the XGB models were between 0.34 and 0.74. Global classification accuracies were above 80% although this reflected a sometimes extreme Bathy/NotBathy imbalance that inflated global accuracy. This imbalance was mitigated by employing a probability decision threshold (PDT) that equalizes the true positive (Bathy) and true negative (NotBathy) rates. It was concluded that 1) the strength of the bathymetric signal in the PAD should be sufficient to increase accuracy of density-based lidar point cloud bathymetry extraction methods and 2) ML can successfully model the relationship between the PAD and the Bathy/NotBathy classification. A method is also presented to examine the spatial and feature-space distribution of errors that will facilitate quality assurance and continuous improvement. [ABSTRACT FROM AUTHOR]