|
| 低对比度场景下冰水特征与冰封率智能识别研究 |
| Research on intelligent recognition algorithms for ice-water features and ice concentration in low-contrast scenarios |
| 投稿时间：2025-07-28 修订日期：2025-12-26 |
| DOI： |
| 中文关键词: 低对比度 冰封率 智能识别 可变形卷积神经网络 南水北调中线工程 |
| 英文关键词: Low contrast ice concentration intelligent recognition deformable convolutional neural network the Middle Route of South-to-North Water Diversion Project |
| 基金项目:国家重点研发计划项目(2022YFC3202500)；国家自然科学基金(U2443221,U2243221,U2243239)；中国水科院科研专项(HY0145B032021) |
|
| 摘要点击次数: 19 |
| 全文下载次数: 0 |
| 中文摘要: |
| 寒区江河渠库普遍存在冰凌灾害风险，冰封率是影响水体失热成冰和评价冰灾风险的重要参数，其高效监测、准确识别对冰凌洪水灾害预防十分重要。相比于河道，人工输水工程由于边界和水动力条件变化少、水质优等特点，使得冰、水对比度低，区分难度大，造成基于图像的冰封率识别方法误差增大。为此，提出了基于可变形卷积神经网络的冰封率智能识别算法，该算法引入可变形卷积层，能够自适应调整卷积核的采样位置，以实现更精准地捕捉低对比度场景下复杂冰、水的特征。构建了含有330张图像的南水北调中线工程流冰数据集，并采用五折交叉验证方法对算法参数进行了优选，16个典型流冰图像测试结果显示：本文算法冰封率识别准确率(ACC)平均值为0.96，交并比(IoU)平均值为0.91，与OTSU算法、支持向量机等常用冰封率识别算法相比，本文算法的ACC平均值分别提升16%和10%，IoU平均值分别提升19%和9%。本文成果为人工输水渠道的冰封率识别提供了另一种方法手段。 |
| 英文摘要: |
| Ice jam disaster risks are prevalent in rivers, canals, and reservoirs in cold regions. The ice concentration is an important parameter that affects water body heat loss and ice formation, as well as the evaluation of ice disaster risks. Its efficient monitoring and accurate identification are crucial for the prevention of ice jam flood disasters. Compared with natural rivers, water conveyance projects are characterized by less variation in boundaries and hydrodynamic conditions and superior water quality, which results in low contrast between ice and water and increased difficulty in differentiation. This leads to higher errors in image-based ice concentration recognition methods. Based on this challenge, an intelligent ice concentration recognition algorithm based on a deformable convolutional neural network is proposed. This algorithm incorporates a deformable convolutional layer, which can adaptively adjust the sampling positions of convolutional kernels to achieve more accurate capture of complex ice and water features in low-contrast scenarios. A floating ice dataset containing 330 images from the Middle Route of South-to-North Water Diversion Project was constructed, and a five-fold cross-validation method was used to optimize the algorithm parameters. Test results of 16 typical floating ice images show that the average accuracy (ACC) of ice concentration identification reaches 0.96, and the average intersection over union (IoU) reaches 0.91. Compared with commonly used ice concentration recognition algorithms such as the OTSU and SVM, the proposed algorithm improves the average ACC by 16% and 10%, and the average IoU by 19% and 9%, respectively. The findings of this study provide an alternative method for ice concentration recognition in water conveyance channels. |
|
查看/发表评论 下载PDF阅读器 |
| 关闭 |