Coral Reef Bottom Classification Method Based On Semi-supervised Learnin

Coral reef substrate classification is an important research content in the fields of marine science,marine resource exploration and marine military.The rapid development of remote sensing technology has provided a fast,effective and accurate means to acquire coral reef substrate information over large areas.Although deep learning convolutional neural networks have shown excellent performance in remote sensing image segmentation tasks,fully supervised deep learning labeling that relies on pixel-by-pixel labeling is costly and not suitable for large-scale and high-frequency substrate classification work,while semi-supervised deep learning methods based on semi-supervised deep learning can effectively use a small amount of labeled data to generate pseudo-labels for unlabeled data,thus effectively reducing labor costs.However,the performance of existing semi-supervised methods is vulnerable to interference from pseudolabel noise,and the feasibility of semi-supervised coral reef substrate classification has not been verified.To address these problems,this paper proposes a semi-supervised learning-based coral reef substrate classification method,and validates the method with shallow substrate habitat datasets from Buck Island Reef and Pearl and Hermes Atolls,specifically:(1)Pseudolabel generation based on joint multi-model decision makingTo address the problem that the classification accuracy of semi-supervised substrate is easily disturbed by pseudolabel noise,a pseudolabel generation method based on multi-model joint decision making is proposed.First,the existing label data are divided into several sub-datasets,and each sub-dataset is used to train several deep learning networks to generate several sets of pseudo-label data;then,the substrate type in the pseudo-label is judged by multi-model joint decision making,and the confidence of the final pseudo-label data and the substrate type of each pixel is generated.The experimental results show that the multi-model joint decision pseudolabel generation method proposed in this paper can effectively solve the serious problem of single-model pseudolabel noise and can provide higher quality substrate classification pseudolabel data for unlabeled data.(2)Soft and hard collaborative substrate classification under 3Closs guidanceIn order to further reduce the impact of pseudolabel noise on the substrate classification results,a soft-hard collaborative substrate classification method that takes into account the loss function of pseudolabel pixel confidence(Collaboration Choice of decision Confidence Loss function,3CLoss)is proposed.First,the pseudo-labeled data and confidence data are mixed with the existing manual labeled data to generate a new dataset;then,multiple models are retrained with the mixed dataset and confidence data under the guidance of the 3CLoss loss function;then,the substrate information is obtained by the joint decision of the prediction results of each model using soft and hard voting;finally,the above steps are repeated to obtain The final coral reef substrate classification results.The experimental results show that the m Io U of coral reef substrate classification under 3CLoss guidance can be improved by 2.38% compared with the cross-entropy loss function.Compared with the fully supervised method with a single model,the m Io U of the semi-supervised substrate classification method proposed in this paper is improved by 2.81%.Compared with mainstream semisupervised semantic segmentation methods,m Io U improves by 3.08%.In conclusion,the semi-supervised learning substrate classification method proposed in this paper demonstrates the effectiveness of semi-supervised classification techniques in coral reef substrate classification,extends the research of deep learning in the field of coral reef remote sensing,and thus provides a new technical means for coral reef substrate classification.