Identification of offshore mariculture area based on UNet 3+ convolutional neural network and time-series SAR features

Abstract Aiming at the problems of strong randomness, small application range, and great influence of cloud and fog in the identification method of offshore mariculture areas based on optical images, a method for identifying offshore mariculture areas based on the UNet 3+ convolutional neural network and time-series SAR features was proposed. Taking the offshore mariculture area of Guangdong Province as an example, multiple SAR images were preprocessed to synthesize time-series SAR remote sensing images, samples were produced and trained using UNet 3+ convolutional neural network model, and then time-series SAR remote sensing images of different years were used for identification and verification. The results showed that the UNet 3+ model was more suitable for offshore mariculture area identification from time-series SAR remote sensing images than neural network models such as U-Net and MultiResUNet. The accuracy evaluation of time-series SAR data in 2020 showed that the average F1-Score, IoU, Kappa, and final identification accuracies of the method were 0.7544, 0.6092, 0.7383 and 0.9273, with recognition precision and accuracy; the offshore mariculture area in Guangdong Province were mainly distributed in Jiangmen City and Chaozhou City, and the overall trend from 2018 to 2022 was first rising, then falling and then rising again. This method provides a new approach for high-precision identification of offshore seawater mariculture, and provides data and technical support for the construction and management of offshore marine ranches and rapid emergency monitoring of marine fisheries disasters.