Cloud-Native Coastal Turbid Zone Detection Using Multi-Temporal Sentinel-2 Data on Google Earth Engine

Abstract

The lack of clarity in turbid coastal waters interferes with light attenuation and hinders remotely sensed studies in aquatic ecology such as benthic habitat mapping and bathymetry estimation. Although turbid water column corrections can be applied on regions with seasonal turbidity by performing multi-temporal analysis, different approaches are needed in regions where the water is constantly turbid or only exhibits subtle turbidity variations through time. This study aims to detect these turbid zones (TZs) in optically shallow coastal waters using multi-temporal Sentinel-2 surface reflectance datasets to improve the aforementioned studies. The herein framework can be paired with other aquatic ecology remote sensing studies to establish the clear water focus area and can also be used by decision makers to identify rehabilitation areas. We selected the coastlines of Guinea-Bissau, Tunisia, and west Madagascar as our case studies which feature wide-ranging turbidity intensities across tropical, subtropical, and Mediterranean waters and applied three different methods for the TZ detection: Otsu’s method for bimodal thresholding, linear spectral unmixing, and Random Forest (RF) machine learning method on Google Earth Engine as an end-to-end process. Based on our experiments, the RF method yields good results in all study regions with overall accuracies ranging between 88 and 96% and F1-scores between 0.87 and 0.96. TZ detection is highly site-specific due to the inter-class variability that is mainly affected by the nature of the suspended materials and the environmental characteristics of the site.

Earth observation for ecosystem accounting: spatially explicit national seagrass extent and carbon stock in Kenya, Tanzania, Mozambique and Madagascar

“…First, this technological component integrates and combines recent developments in multi‐temporal composition and scalability, optically deep water (areas with no emanating signal in the water surface from the seafloor) and turbidity masking (Pertiwi et al., 2021; Poursanidis et al., 2021; Thomas et al., 2021; Traganos, Aggarwal, et al., 2018; Traganos, Poursanidis, et al., 2018). Here, it used the 33 095 raw Sentinel‐2 L2A tiles ingested by GEE within the area of interest and filtered them for cloud coverage using the GEE filterMetadata function, retaining only tiles with less than 25% CLOUDY_PIXEL_PERCENTAGE, which halved the raw archive to 16 453 S2 tiles.…”

“…Second, the percentile composite is masked for land and optically deep water pixels utilizing a non‐parametric, highly adaptive algorithm (Donchyts et al., 2016) based on a combined Otsu‐based thresholding (Otsu, 1979) and Canny edge filter applied to the modified normalized difference water index, MNDWI (Pertiwi et al., 2021; Thomas et al., 2021; Xu, 2006). The MNDWI index integrated the green (B3; 560 nm) and shortwave infrared (B11; 1610 nm) bands of the Sentinel‐2 L2A percentile composite.…”

“…An in‐depth description of the turbidity processor is given in Pertiwi et al. (2021). The utilization of this pre‐processing step was deemed necessary due to high reflectances (bright targets) caused by yearlong, persistent turbidity across the entire East Africa, remnants of the MT composition.…”

“…Finally, we employed a recently developed turbidity processor to detect and mask medium to high turbidity pixels by applying spectral unmixing on the percentile composite aided by the annotated data points described in section 2.2.2 and Table 1. An in-depth description of the turbidity processor is given in Pertiwi et al (2021). The utilization of this pre-processing step was deemed necessary due to high reflectances (bright targets) caused by yearlong, persistent turbidity across the entire East Africa, remnants of the MT composition.…”

Deteksi Turbidity Front Menggunakan Citra Satelit Sentinel-2 Hubungannya Dengan Oseanografi Di Estuari Bengawan Solo

“…Pengindraan jauh memiliki kemampuan yang baik, meliput area yang luas dan sulit dijangkau sehingga pengambilan data lebih efektif dan efisien (Munthe et al, 2007;Katlane et al, 2020). Beberapa algoritma yang dikembangkan untuk mengestimasi TSS dari citra satelit umumnya bersifat lokal sehingga perlu dikembangkan untuk memperoleh algoritma yang sesuai pada area penelitian ini (Budiman, 2005;Parwati & Purwanto, 2014;Heltria et al, 2021;Pertiwi et al, 2021;Selamat & Ukkas, 2020).…”