Satellite-Derived Bathymetry Using Random Forest Algorithm and Worldview-2 Imagery

Abstract

M, et al. (2016). Satellite-derived bathymetry using random forest algorithm and worldview-2 imagery.

Monitoring Beach Topography and Nearshore Bathymetry Using Spaceborne Remote Sensing: A Review

“…However, to overcome the instrumental limitations (a low spectral resolution and noise-to-signal ratio) inherent in these sensors, recent developments focus on the use of multi-temporal and machine learning approaches to derive bathymetric maps in shallow waters. For example, various machine learning approaches have been employed, such as using artificial neural network (ANN) [123][124][125], support-vector machine (SVM) [126][127][128], and random forest (RF) [79,129] approaches. Good accuracies can be obtained using these methods, with RMSEs that can reach 35 cm by comparing to observations with high accuracies [126].…”

Satellite Derived Bathymetry Using Machine Learning and Multi-Temporal Satellite Images

“…Another method that is currently in use for bathymetric measurement is depth estimation via remote sensing, using multi-spectral or hyper-spectral sensors. Many studies have been conducted using this method since the 1970s [3][4][5][6][7][8][9][10][11][12][13][14]. The use of satellite images helps to obtain information about areas that are difficult to access by boat or airplane.…”

Determination of the Best Methodology for Bathymetry Mapping Using Spot 6 Imagery: A Study of 12 Empirical Algorithms

“…MLR algorithm assumed that water quality and atmospheric condition is uniform, and the number of bottom types is less than a number of used bands are unrealistic for much shallow water environment (Kanno et al 2011). RF algorithm used in this study is run on auto-tuning mode, however, to get the best result of random forest algorithm, it is necessary to do an optimization on the hyper-parameters (Manessa et al 2016a). LR and PC algorithm focused on noise reduction (Stumpt et al 2003 andVan Hengel andSpitzer 1991) but not consider that the linear regression works well with a number of explanatory variables.…”

“…A decade after, for the first time used a commercial satellite data LANDSAT TM to extract the depth of using a linearized regression of single band (Lyzenga 1985), this method was based on previous publication (Lyzenga 1978). Since that time the algorithm has been redeveloped and applied to the newest multispectral image: LANDSAT-TM and ETM (Clark et al 1987;Van Hengel and Spltzer 1991;Bierwirth et al 1993;Daniell 2008), SPOT 4 and SPOT 5 (Melsheimer and Chin 2001;Lafon et al 2002;Liu et al 2010;Sánchez-Carnero et al 2014), IKONOS (Stumpf et al 2003;Hogrefe et al 2008;Su et al 2014), QuickBird (Conger et al 2006;Mishra et al 2006;Lyons et al 2011), LANDSAT-OLI (Pacheco et al 2015;Vinayaraj et al 2016;Kabiri 2017;Pushparaj and Hegde 2017), and Worldview-2 (Lee and Kim 2011;Deidda and Sanna 2012;Doxani et al 2012;Bramante et al 2013;Kanno et al 2013;Yuzugullu and Aksoy 2014;Eugenio et al 2015;Manessa et al 2016b;Guzinski et al 2016;Hernandez and Armstrong 2016;Kibele and Shears 2016;Manessa et al 2016a).…”

“…Our research focuses on the finding the best empirical SDB algorithm for SPOT 6 multispectral data. Twelve empirical SDB algorithm was intensionally chosen: 1) Ratio transform (henceforth named "RT") by Stumpt et al (2003); 2) Multiple linear regression (henceforth named "MLR") by Lyzenga et al (2006); 3) Multiple non-linear regression (henceforth named "RF") by Manessa et al (2016a); 4) Second-order polynomial of ratio transform (henceforth named "SPR") by Mishra et al (2006); 5) Principle component (henceforth named "PC") by Van Hengel and Spitzer (1991); four extension of Lyzenga's SDB algorithm by Kanno et al (2011): 6) Multiple linear regression using relaxing uniformity assumption on water and atmosphere (henceforth named "KNW"); 7) Semiparametric regression using depthindependent variables (henceforth named "SMP"); 8) semiparametric regression using spatial coordinates (henceforth named "STR"); 9) Semiparametric regression using depth-independent variables and spatial coordinates (henceforth named "TNP"), and three statistic new statistical approach of Mohamed et al (2017): 10) Bagging Fitting Ensemble (henceforth named "BAG"); 11) Least Squares Boosting Fitting Ensemble (henceforth named "LSB"); and 12) support vector regression (henceforth named "SVR").…”