Conference Agenda

During over the past 10 years, the massive green macro-algal bloom has regularly occurred in the Yellow Sea, the spatial coverage of which is mainly derived by the remote sensing community from satellite images with moderate/low resolution (30-m~1000-m), such as the 250-m-resolution MODIS (Moderate Resolution Imaging Spectroradiometer). In this paper, the MODIS estimates are compared for the first time with the concurrent high resolution (3-m) airborne Synthetic Aperture Radar (SAR) data. We find that the MODIS results are overestimated by more than a factor of 3 when each algae pixel is assumed to be pure (i.e. 100% algae cover), whereas the overestimation is significantly reduced to 1.14 when the pure pixel assumption is abandoned and the genuine (fractional) algae coverage is derived with the linear pixel un-mixing method. These results, together with the re-sampling processing of the high resolution images, indicate that the mixed pixel effect, that is inherent with images with moderate and low resolutions, is the key factor for the satellite extraction of the macro-algae coverage, and these findings are further confirmed by the satellite data with different resolutions. Besides, significant correlations (R²>0.9) are found between the macro-algae coverage from 3-m resolution SAR images and those from concurrent satellite images with various resolutions (30-m~1000-m) under the pure pixel assumption, which provides an alternative statistics-based method (in addition to the linear pixel un-mixing) for the accurate macro-algae coverage extraction from satellite images with coarse resolution (e.g. HJ-1 CCD, AQUA MODIS, COMS GOCI). This new method is independently validated with high resolution optical images, and applied to derive the annual maxima of the massive green macro-algal bloom areas (fractional coverage) in the Yellow Sea from 2007 to 2016, which ranges from 45.6~732.9-km² with an average of 247.9 ± 199.3-km².

High-resolution satellite earth observation data are available in large archives, as data collection increases the ability to inspect and investigate each scene becomes impossible due to the scale and quantity of observations. Computer assisted classification, segmentation and description of satellite data over aquatic bodies can provide invaluable information for focusing analysis to experts and the general public on everyday use of water resources.

Convolutional Neural Networks are capable of classifying and segmenting objects across thousands of images in a fraction of the time a human operator would require inspecting these images visually. These Deep Learning networks have previously been applied to classifying both land usage and land cover, they have been shown to be accurate using multi- or hyper-spectral data such as those collected by the Sentinel-2 MultiSpectral Instrument.

In this work, a training dataset consisting of coastal and in-land waters has been assembled from Sentinel-2 imagery covering multiple sites across North America, South Africa and China and extensively labelled to be compatible with Deep Learning methods. Convolutional Neural Networks developed and trained for natural image classification and segmentation have been extended and retrained through transfer learning to detect and segment the extents of Algal Blooms and River Plumes in the imagery.

Current Convolutional Neural Network architectures are evaluated to establish best approaches to leverage spectral and spatial data in the context of water classification. Several spectral data configurations are used to evaluate competency and suitability for generalisation to other Optical Satellite Sensor configurations. The impact of the atmospheric correction technique applied to data is explored to establish the most reliable data for use during training and requirements for pre-processing data pipelines.