Deep-Learning-Based Marine Aquaculture Zone Extractions From Dual-Polarimetric SAR Imagery

Abstract

Efficient monitoring of marine aquaculture zones 1 (MAZs) is crucial for facilitating coastal resource management. 2 To achieve this, we developed a specialized deep convolutional 3 neural network tailored for extracting MAZs from Synthetic 4 Aperture Radar (SAR) imagery, integrating prior analytical 5 knowledge of MAZ imaging features. 47 Sentinel-1 dual-polarized 6 (VV and VH) SAR images spanning 2016-2023 in China's Subei 7 Sandbanks along the Yellow Sea coast were collected due to 8 appropriate tidal level and acquired time. We first compre-9 hensively analyzed of Normalized Radar Cross Section (NRCS) 10 values for MAZs under varying tidal levels and aquaculture 11 facility structures. Rising tide-induced submergence resulted in 12 a significant mean NRCS reduction of 7.01 dB (VV) and 4.54 dB 13 (VH), causing MAZ signals to resemble seawater. Additionally, 14 during low tide, volume scattering from the net screen on the 15 aquaculture rafts increased VH-polarized image recognizability, 16 with a smaller NRCS overlap (64%) between MAZs and tidal 17 flats compared to VV-polarized images. Hence, VH-polarized 18 images taken during low tide with intact aquaculture facilities 19 were selected for dataset construction due to their reliability in 20 characterizing MAZs. Building upon the classical U-Net frame-21 work, we introduced four modifications informed by our imaging 22 characteristics analysis to enhance the model's performance. 23 Testing experiments demonstrated an impressive F1-score of 24 94.77%, highlighting the effectiveness of incorporating prior 25 knowledge into refining deep learning models. Applying the 26 model to SAR images from 2016 to 2023 revealed concentrated 27 MAZs in the relatively flat southeastern Subei Sandbanks, with a 28 noticeable scale decline post-2021 resulting in a 67.65% reduction 29 over the years. 30 Index Terms-synthetic aperture radar, Sentinel-1, imaging 31 characteristics, deep convolution neural networks, marine aqua-32 culture zones. 33 I. INTRODUCTION 34 M ARINE aquaculture is crucial and steadily expanding 35 worldwide [1]. In 2020, marine aquaculture production 36 reached 33.6 million tons, serving as a substantial source 37