Characterizing oil spills using deep learning and spectral-spatial-geometrical features of HY-1C/D CZI images

Marine oil spills cause pollution to the environment, where timely information on the oil characteristics (i.e., oil types, concentration, thickness) is essential for spill response and post-spill assessment. Existing remote sensing models (including deep-learning or DL) have been applied to both synthetic aperture radar (SAR) and optical remote sensing images, which are primarily for presence/absence detection. Here, we propose a framework to take advantage of data richness (i.e., multi-band observations under different viewing conditions) of optical sensors to characterize oil spills beyond presence/absence detection. The framework includes a 2-step DL model (F1-score ∼ 0.88) to account for the spectral-spatial-geometrical features of optical images to achieve the presence/absence detection, and a spectral model to characterize oil type and estimate oil quantity. The framework led to a relatively comprehensive “ground truth” dataset (N = 105 oil spill cases with 1616 oil features of each being at least 3 × 3 pixels) including all possible scenarios to account for various oil types and observing conditions, which can be used in the future for model development. Application of the framework to >25,000 HY-1C/D CZI images between December 2018 and April 2023 revealed substantial oil spills in the marginal seas off China, whose spatial distributions and oil characteristics (e.g., normalized concentration, thickness, and volume) were all quantified. The DL model also appears to be applicable to other Sentinel-1 SAR as well as to other optical sensors such as Landsat Operational Land Imager (OLI) and Sentinel-2 Multispectral Instrument (MSI). The use of multi-sensor data with the proposed approach over an oil spill event in the Gulf of Mexico shows improved temporal coverage as well as better knowledge on where most of the spilled oil can be found.