CN:36-1239/TH

ISSN:1672-3872

半月刊

JST日本科学技术振新机构数据库(日)(2022)

中文核心期刊(遴选)数据库收录期刊

中文科技期刊数据库收录期刊

中国学术期刊(光盘版)全文收录期刊

中国期刊网收录期刊

中国学术期刊综合评价数据库 统计源期刊

搜索
搜索
这是描述信息

基于对抗域适应网络的土壤水分智能识别算法

访问量:

DOI:10.3969/j.issn.1672-3872.2023.20.014

作者:孙德鑫(长安大学工程机械学院,陕西 西安 710064)

 

摘 要:【目的】使用传统卷积神经网络测量土壤含水率的识别准确率低、泛化性能差。为了满足重大工程扰动下的含水率测量需要,需提出一种适用于不同土壤类型且不受深度制约的探测方法。【方法】笔者搭建了室内实验平台,采集了延安、兰州 和蓝田三个地区间隔等级为2%的8类不同土壤水分下的图像,构造了用于神经网络训练和测试的数据集。然后,基于对抗域适应算法(CDAN),比较了ResNet、MobileNetV2、Xception和ViT四种网络,选取识别准确率最高的模型作为其特征提取器,构建基于对抗域适应的土壤含水率识别模型,对比分析含水率识别模型在不同迁移任务上的识别效果。【结果】域适应模型CDAN的测试准确率明显高于MobileNetV2、Xception和ViT等模型,对不同地区土壤含水率的测试准确率均达到68%以 上,最高识别准确率为84.3%。【结论】使用域适应方法具有良好的泛化能力和识别精度,能够为开发土壤地质信息探测机器人的视觉系统提供算法支持。然而,目前的土壤含水率识别算法仅适用于实验室条件下所搭建的土壤含水率数据集,若想实现实际的工程应用,则需要进一步完善土壤含水率图像数据集,使室内搭建的算法实现室外的原状土壤含水率识别。

关键词:土壤水分;智能识别;Vision Transformer;迁移学习

 

Intelligent Identification Algorithm of Soil Moisture Based on Adversarial Domain Adaptive Network

Sun Dexin (School of Construction Machinery, Chang’an University, Shaanxi Xi’an 710064)

 

Abstract: [Objective] Using traditional convolutional neural network to measure soil moisture content has low recognition accuracy and poor generalization performance. In order to meet the needs of water content measurement under the disturbance of megaproject, it is necessary to propose a detection method that is suitable for different soil type and is not restricted by depth. [Method] The author established an indoor experimental platform and collected images of 8 different soil moisture levels with an interval of 2% in Yan’an, Lanzhou, and Lantian regions. A dataset for neural network training and testing was constructed. Then, based on the conditional adversarial domain adaptation algorithm (CDAN), four networks, ResNet, MobileNetV2, Xception and ViT were compared. The model with the highest recognition accuracy was selected as its feature extractor, and a soil moisture content recognition model based on conditional adversarial domain adaptation was constructed. The recognition effects of the moisture content recognition model on different migration tasks were compared and analyzed. [Result] The testing accuracy of the domain adaptation model CDAN is significantly higher than that of models such as MobileNetV2, Xception and ViT. The testing accuracy of soil moisture content in different regions is above 68%, with the highest recognition accuracy of 84.3%. [Conclusion] The use of domain adaptation methods has good generalization ability and recognition accuracy, and can provide algorithm support for the development of visual systems for soil geological information detection robots. However, current soil moisture content recognition algorithms are only applicable to soil moisture content datasets built under laboratory conditions. If practical engineering applications are to be achieved, it is necessary to further improve the soil moisture content image dataset, so that indoor algorithms can achieve outdoor undisturbed soil moisture content recognition.

Keywords: soil moisture; intelligent recognition; Vision Transformer; transfer learning

 

引文信息:[1]孙德鑫.基于对抗域适应网络的土壤水分智能识别算法[J].南方农机,2023,54(20):53-57.

查看全文请下载PDF文件↓

相关下载

分类:
2023年
文件大小:
1.6M
2023-11-02 10:40:23
所属人群:
所有人
上一页
1
底部logo

公众号

地       址:江西省南昌市红谷滩红谷中大道1326号江报传媒大厦908室

联系电话:0791-86202556

投稿邮箱:nfnj@vip.163.com

版权所有:江西南方农机杂志社有限责任公司.  All rights reserved.   SEO     赣ICP备2023003226号-1       技术支持:中企动力-南昌