Machine learning-based retrieval of benthic reflectance and Posidonia oceanica seagrass extent using a semi-analytical inversion of Sentinel-2 satellite data. Issue 24 (17th December 2018)

• Record Type: Journal Article
• Authors: Traganos, Dimosthenis; Reinartz, Peter
• Is Part Of: International journal of remote sensing. Volume 39:Issue 24(2018)
• Page Start: 9428
• Page End: 9452
• Publication Date: 2018-12-17
• Subjects: Remote sensing -- Periodicals; Télédétection -- Périodiques; 621.3678
• DOI: 10.1080/01431161.2018.1519289
• Languages: English
• ISSNs: 0143-1161

Abstract

ABSTRACT: In the epoch of the human-induced climate change, seagrasses can mitigate the resulting negative impacts due to their carbon sequestration ability. The endemic and dominant in the Mediterranean Posidonia oceanica seagrass contains the largest stocks of organic carbon among all seagrass species, yet it undergoes a significant regression in its extent. Therefore, suitable quantitative assessment of its extent and optically shallow environment are required to allow good conservation and management practices. Here, we parameterise a semi-analytical inversion model which employs above-surface remote sensing reflectance of Sentinel-2A to derive water column and bottom properties in the Thermaikos Gulf, NW Aegean Sea, Greece (eastern Mediterranean). In the model, the diffuse attenuation coefficients are expressed as functions of absorption and backscattering coefficients. We apply a comprehensive pre-processing workflow which includes atmospheric correction using C2RCC (Case 2 Regional CoastColour) neural network, resampling of the lower spatial resolution Sentinel-2A bands to 10m/pixel, as well as empirical derivation of water bathymetry and machine learning-based classification of the resulting bottom properties using the Support Vector Machines. SVM-based classification of benthic reflectance reveals ~300 ha of P. oceanica seagrass between 2 and 16 m of depth, and yields very high producer and user accuracies of 95.3% and 99.5%, respectively. Sources of errors and uncertainties are discussed. All in all, recent advances in Earth Observation in terms of optical satellite technology, cloud computing and machine learning algorithms have created the perfect storm which could aid high spatio-temporal, large-scale seagrass habitat mapping and monitoring, allowing for its integration to the Analysis Ready Data era and ultimately enabling more efficient management and conservation in the epoch of climate change.