Abstract: Aiming at the problem that the traditional high-cost and low-efficiency sea area supervision methods such as field investigation and evidence collection and manual comparison of remote sensing images can not meet the current regulatory requirements, we use satellite remote sensing images and deep learning algorithm to propose a comprehensive control method for large-scale, fast and dynamic sea use. Based on massive images and multi-source marine basic data, multi planning fusion analysis, fast R-CNN(Regions with Convolutional Neural Networks features) algorithm, an artificial intelligence recognition model is established to realize automatic recognition and early warning of maritime targets, illegal sea occupation and ecological environment damage. We analyze the principle of fast R-CNN algorithm, uses satellite remote sensing data of different years, different satellites and different resolutions, establishes more than 10 000 image sample database for five kinds of common marine targets, and conducts training and testing with VGG16 and RestNet101 network models. The experimental results show that the computational complexity of RestNet101 model is slightly larger than that of VGG16 model, but it has stronger ability of complex feature extraction, and is more suitable for the detection and recognition of complex sea targets; the overall recognition accuracy of the five types of targets is more than 80% . Combined with the marine planning data, the model realizes the rapid automatic supervision and early warning of illegal user behavior in large-scale sea area, which provides a new idea for the intelligent marine supervision.

Key words: faster regions with convolutional neural networks features( Faster R-CNN), convolutional neural network, multi planning fusion, remote sensing image, sea area supervision