基于软硬协作决策的半监督珊瑚礁底质分类方法

A semi-supervised coral reef substrate classification method based on soft and hard collaborative decision making

珊瑚礁底质分类对海洋资源开发和海洋生态环境保护起到至关重要的作用。目前,深度学习语义分割方法在遥感图像分类领域应用广泛,但在底质分类方面的研究较少。由于基于全监督深度学习的方法中逐像素标注标签的成本较高,不适用于大规模、高频次的底质分类工作,基于半监督的深度学习方法能够有效利用已标注标签为无标签数据产生伪标签,从而有效降低人工成本,然而现有半监督方法的性能易受伪标签噪声的干扰。针对以上问题,本文提出了一种基于软硬协作决策的半监督底质分类方法。首先,利用多模型联合决策生成高质量的伪标签;然后,提出了一种能够顾及伪标签像素置信度的损失函数来指导模型进行训练;最后,采用软硬协作的决策方式得到精确的底质分类结果。在美属维尔京群岛圣克罗伊岛北部的巴克岛礁和夏威夷群岛的中途岛东南约400km处的珍珠与爱马仕环礁的浅层底栖生物栖息地地图数据集上评估了本文方法的精度,实验结果表明,本文提出的方法与全监督学习方法精度相当,比主流的语义分割方法精度平均高3.08%,能够有效服务于珊瑚礁底质调查工作。

Coral reef substrate classification plays a crucial role in marine resource development and marine ecological protection.At present,deep learning semantic segmentation methods are widely used in the field of remote sensing image classification,but less research has been conducted in substrate classification.Due to the high cost of pixel-by-pixel labeling in the fully supervised deep learning-based method,it is not suitable for large-scale and high-frequency substrate classification work.The semi-supervised deep learning-based method can effectively use the labeled labels to generate pseudo-labels for unlabeled data,thus effectively reducing the labor cost,however,the performance of the existing semi-supervised method is vulnerable to the interference of pseudo-label noise.To address the above problems,this paper proposes a semi-supervised substrate classification method based on soft and hard collaborative decision making.First,a high quality Pseudo tag is generated using joint decision making of multiple models;then,a loss function(Collaboration Choice of decision Confidence Loss function,3CLoss)is proposed to take into account the confidence of Pseudo tag pixels and guide the model for training;finally,a soft and hard collaborative decision making approach is used to obtain accurate substrate classification results.The accuracy of this paper was evaluated on the shallow benthic habitat atlas datasets of Buck Island Reef in the northern part of St.Croix,U.S.Virgin Islands,and Pearl and Hermes Atolls,about 400 km southeast of Midway Island,Hawaiian Islands,and the experimental results show that the accuracy of the proposed method is comparable to that of the fully supervised learning method,and 3.08%higher than that of the mainstream semantic segmentation methods on average,which can effectively serve the coral reef substrate survey.