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

Abstract

Shallow water bathymetry is important for nautical navigation to avoid stranding, as well as for the scientific simulation of high tide and high waves in coastal areas. Although many studies have been conducted on satellite derived bathymetry (SDB), previously used methods basically require supervised data for analysis, and cannot be used to analyze areas that are unreachable by boat or airplane. In this study, a mapping method for shallow water bathymetry was developed, using random forest machine learning and multi-temporal satellite images to create a generalized depth estimation model. A total of 135 Landsat-8 images, and a large amount of training bathymetry data for five areas were analyzed with the Google Earth Engine. The accuracy of SDB was evaluated by comparison with reference bathymetry data. The root mean square error in the final estimated water depth in the five test areas was 1.41 m for depths of 0 to 20 m. The SDB creation system developed in this study is expected to be applicable in various shallow water regions under highly transparent conditions.

Towards Routine Mapping of Shallow Bathymetry in Environments with Variable Turbidity: Contribution of Sentinel-2A/B Satellites Mission

“…The multi-scene compositing approach addressed limitations inherent in conventional methods and reduced the impact of turbidity, performing better than the standard "pick the best scene" method that relies on a single image (Figures 3 and 4c vs. Figure 4a in Cape Lookout; and Figure 10e vs. Figure 10a,c in Saint Joseph Bay). The final corrected SDB produced robust depths up to the limit of the lidar surveys, with typical errors ≤0.4 m. These excellent results from Sentinel-2 compared favorably with those produced in relatively low turbidity water in south Florida [22,44], and in regions with transparent waters [10,18,20,30]. Whereas some researchers suggested there is still work to be performed regarding the identification of the optimal period throughout the year where bathymetric errors are minimized [18,29], others asked for novel strategies to allow seabed mapping without the laborious analysis per image and the visual inspection of the "clearest scene" [19].…”

“…Whereas some researchers suggested there is still work to be performed regarding the identification of the optimal period throughout the year where bathymetric errors are minimized [18,29], others asked for novel strategies to allow seabed mapping without the laborious analysis per image and the visual inspection of the "clearest scene" [19]. Recent studies have already indicated the potential of multi-scene approaches in order to select the optimal scene or eliminate noise over clear waters [16,17,19,20]. However, our temporal compositing strategy successfully reduced the turbidity impact without requirement of visual inspection, thereby enhancing SDB performance in an easy way.…”

“…The multi-scene strategy applied here did not require any screening or manual adjustment of the imagery prior to compositing. It automatically picked the pixels least impacted by turbidity (e.g., Figure 8) from the set of scenes provided (Table 1), substantially simplifying the effort compared with other studies where the selection of optimal images with variation in the water quality conditions were essential in the extraction of SDB [20,22,29,55]. Manual selection of the optimal scene is not only highly subjective, but requires considerable time and effort, and may still include regions having patches of turbidity.…”

“…These This paper proposes the first multiple-image technique with Sentinel-2A and Sentinel-2B that allows semi-automated SDB estimation in areas of variable turbidity. Although some recent works have already obtained accurate results with Sentinel-2 and similar satellites, they were implemented in regions with transparent waters [18][19][20]29,30]. We focus on application of the log ratio model [31] because it was designed to support routine mapping in areas with extremely limited calibration data.…”

“…There have been previous studies inspecting water quality issues on SDB [11][12][13][14][15]. Recent studies have already indicated the potential of multi-scene approaches in order to eliminate noise over clear waters [16][17][18][19][20]. The impact of turbidity on SDB is not random and often appears as a false shoal [21,22].…”