A Multi-Scale Spatial Attention Network for Aquaculture Extraction Based on Time Series SAR Image

Synthetic aperture radar (SAR) is a kind of radar system with high resolution remote sensing imaging capability, which can be used to obtain remote sensing images at different times and under complex weather conditions. In this study, a multi-scale spatial attention network based on time series aquaculture extraction (MSATUnet) is proposed. The model uses multiple time series SAR images to extract sequence features, and uses a encoder decoder network design to extract spatial features of different scales. In the encoder part, the network calculates the spatial attention weight of the multi-scale feature map to improve the feature expression ability of the convolutional layer for the target. Through the design of convolutional long short-term memory (ConvLSTM) and self-attention mechanism in the deep layer of the network, the network can capture the temporal features of sequence images and improve the ability of global semantic information of the model. In addition, in order to solve the problem of target adhesion, a edge loss function is introduced to adjust the edge details. The experiment uses the time-series images taken by GF-3 satellite as the data input, and compares with other networks. The results show that MSATUnet outperforms traditional single-input networks and multiple time series models, which verifies the effectiveness of multiple time series in marine aquaculture extraction.