Coral Reef Benthic Substances Extraction Based On Convolutional Neural Networks

Coral reef ecosystems are rich in species and are important for socio-economic and coastal protection.However,affected by climate warming and human activities,the coral reef ecosystem is gradually being destroyed.In order to better protect the coral reef ecosystem,it is necessary to map the distribution of benthic substances to understand the distribution of each benthic substance.At present,the mainstream mapping methods for the distribution of coral reef benthic substances are field experiments,machine learning algorithms,and object-oriented algorithms.However,the cost of field experiments is high,the accuracy of machine learning algorithms is not high,and the object-oriented algorithms are accurate but not automatic enough.At present,deep learning methods have been widely used in the field of remote sensing information extraction,but there are few studies on coral reef benthic material information extraction using deep learning technology.There are still many studies on coral reef information extraction based on deep learning algorithms that can be optimized.Where it is necessary to design a refined labeling method;a more effective feature extraction block can be designed to improve the network feature extraction capability;the amount of network parameters needs to be reduced to facilitate deployment on portable devices.In order to solve the above problems,this work designs a method based on multi-scale segmentation to make fine-grained labels,proposes a FASU-Net network for fine-grained extraction,and proposes a DRDAU-Net network to achieve the goal of lightweight.The specific research contents are as follows:(1)A training label making method for coral reef benthic material information extraction based on image segmentation was designed.For different regions,three levels of different scale segmentation and post-segmentation processing are designed,and the KNN algorithm is used for final classification.Compared with the labeling results of the labelme tool and the SVM algorithm,it is proved that the results of this method are more refined and more suitable for labeling of coral reef benthic material information extraction tasks.(2)A coral reef benthic information extraction network based on FASU-Net is proposed.Combining factorized convolution,attention mechanism and channel shuffling operation,a feature extraction block is designed,a multi-input structure is designed to enrich network detail information,attention mechanism is used to improve long-hop connections to suppress irrelevant features,and encoder structure designed as a Res Net-34 structure to extract richer features.The effectiveness of each module is proved by ablation experiments.Compared with some commonly used extraction algorithms,it is proved that this method has better extraction effect.(3)A DRDAU-Net based coral reef benthic material information extraction network is proposed.A lightweight feature extraction block is designed by combining depthwise separable convolution,attention mechanism and adaptive normalization,and a short path with attention mechanism is designed on the network far-hop connection to suppress irrelevant features.Experiments were carried out using GF-2 and Sentinel-2 data.Compared with some commonly used networks,DRDAU-Net has better extraction effect and generalization ability while having fewer parameters.