湖北农业科学 ›› 2023, Vol. 62 ›› Issue (11): 176-182.doi: 10.14088/j.cnki.issn0439-8114.2023.11.031

• 农业工程 • 上一篇    下一篇

基于深度学习和遥感数据的水稻种植面积提取

邱儒琼1,2, 彭少坤2, 李孟璠2   

  1. 1.中国地质大学(武汉)国家地理信息系统工程技术研究中心,武汉 430074;
    2.湖北省发展规划研究院有限公司,武汉 430071
  • 收稿日期:2022-03-29 出版日期:2023-11-25 发布日期:2023-12-25
  • 作者简介:邱儒琼(1977-),女,湖北公安人,正高级工程师,在读博士研究生,研究方向为遥感数据智能处理的理论和应用,(电话)18602740369(电子信箱)404251392@qq.com。
  • 基金资助:
    湖北省自然资源厅科研项目(ZRZY2021KJ03)

Extracting rice planting area based on deep learning and remote sensing data

QIU Ru-qiong1,2, PENG Shao-kun2, LI Meng-fan2   

  1. 1. National Engineering Research Center of Geographic Information System,China University of Geosciences(Wuhan), Wuhan 430074,China;
    2. Hubei Development Planning Research Institute Co., Ltd., Wuhan 430071,China
  • Received:2022-03-29 Online:2023-11-25 Published:2023-12-25

摘要: 针对现有基于深度卷积神经网络进行水稻(Oryza sativa L.)遥感识别中样本采集工作量大、样本标注要求高及水稻感受野尺度选择难等问题,构建了一种基于像元和多尺度的深度卷积神经网络(DCNN)水稻遥感识别模型。首先,针对水稻种植分布特点,综合深度卷积神经网络方法的特点,设计了基于像元的DCNN提取模型;其次,将多尺度与DCNN相结合,构建多尺度DCNN模型,增加感受野的多尺度特性;最后,为了验证多尺度DCNN模型提取水稻的效果,以高分一号和高分二号卫星影像为数据源,选取传统机器学习SVM模型、语义分割D-Linknet模型、单一尺度DCNN模型进行分类精度对比分析。结果表明,本研究提出的多尺度DCNN模型的准确率、精确率、召回率、平衡F分数分别为97.75%、96.68%、99.08%、97.85%;与其他模型相比,多尺度DCNN模型结构简单、样本制作简便、识别精度高,具有较好的应用价值。

关键词: 水稻(Oryza sativa L.), 高分辨率遥感影像, 深度学习, 种植面积提取, 像元分类, 深度卷积神经网络(DCNN), 多尺度DCNN模型

Abstract: A pixel and multi-scale Deep Convolutional Neural Networks (DCNN) rice(Oryza sativa L.) remote sensing recognition model was constructed to address the issues of large sample collection workload, high sample labeling requirements, and difficulty in selecting the scale of rice receptive fields in existing rice (Oryza sativa L.) remote sensing recognition based on Deep Convolutional Neural Networks. Firstly, based on the distribution characteristics of rice planting, a pixel based DCNN extraction model was designed by integrating comprehensively the characteristics of Deep Convolutional Neural Networks methods;secondly, by combining multi-scale and DCNN, a multi-scale DCNN model was constructed to enhance the multi-scale characteristics of the receptive field; finally, in order to verify the effectiveness of the multi-scale DCNN model in extracting rice, the traditional machine learning SVM model, semantic segmentation D-Linknet model, and single-scale DCNN model were selected for classification accuracy comparison and analysis using Gaofen-1 and Gaofen-2 satellite images as data sources. The results showed that the accuracy, precision, recall, and equilibrium F-scores of the multi-scale DCNN model proposed in this study were 97.75%, 96.68%, 99.08%, and 97.85%, respectively;compared with other models, the multi-scale DCNN model had a simple structure, simple sample production, and high recognition accuracy, which had good application value.

Key words: rice (Oryza sativa L.), high resolution remote sensing images, deep learning, extraction of planting area, pixel classification, deep convolutional neural networks (DCNN), multi-scale DCNN model

中图分类号: