高级检索

    基于卷积神经网络及遥感反演的生产建设项目识别方法及应用

    A methodology of identifying production and construction projects by integrating convolutional neural networks with remote sensing inversion model and its application

    • 摘要: 由于遥感影像中的地物普遍存在异物同谱、同物异谱等现象, 导致面向对象的多尺度和多阈值分割分类等解译方法无法在完全脱离人工干预的情况下使用, 因而给解译带来成本高、效率较低等问题。本研究目的是建立一个实用、高效的生产建设项目扰动图斑自动识别模型, 提高扰动图斑识别的效率和准确性。河南省作为水利部第一批遥感监管示范省, 至今已积累大量生产建设项目水土保持空间数据。为此, 笔者选择河南省作为研究区。提出一种基于神经网络及遥感反演模型的生产建设项目自动识别方法。该方法使用由5个卷积层和3个池化Pooling层的AlexNet对扰动图斑进行初步自动解译, 并使用归一化植被指数、温度植被干旱指数等遥感反演模型进一步判断图斑是否为生产建设项目。笔者以成像时间为2020年、覆盖河南全省的GF-1、GF-6遥感影像为基础, 使用生产建设项目扰动图斑的外包矩形裁剪上述影像获得训练样本, 对样本进行标记, 并将标记好的训练样本输入到AlexNet模型对其进行训练, 完成训练后分别使用单一神经网络模型及神经网络模型+遥感反演模型进行两组测试。结果表明: 1) 本方法在降低误识别率和漏识别率的效果方面均比单纯卷积神经网络识别模型优异, 误识别率平均降低23.88%, 尤其在降低误识别率方面效果明显; 2) 本方法的计算时间, 与卷积神经网络识别模型没有明显差异。综上, 本研究成果可有效降低生产建设项目自动识别的误判率, 有助于进一步提升生产建设项目水土保持遥感监管工作的自动化程度, 在生产建设项目自动识别应用领域有一定的借鉴意义。

       

      Abstract:
      Background Due to the phenomenon of spectral similarities of foreign objects and spectral differences of the same objects in remote sensing images, object-oriented multi-scale and multi-threshold segmentation and classification methods cannot be used without significant manual intervention. This manual intervention results in high costs and low efficiency in the interpretation process. The purpose of this study is to establish a practical and efficient model for automatic identification of disturbed areas in production and construction projects, aiming to improve the efficiency and accuracy of disturbance identification. As one of the first batch of remote sensing monitoring demonstration provinces designated by the Ministry of Water Resources, Henan province has accumulated a significant amount of spatial data on soil and water conservation for production and construction projects. Therefore, this study selects Henan province as the research area.
      Methods The authors proposed an automatic identification method for production and construction projects based on neural networks and remote sensing inversion models. This method utilized AlexNet, which consists of 5 convolutional layers and 3 pooling layers, for initial automatic interpretation of disturbed spots. Additionally, remote sensing inversion models such as normalized difference vegetation index (NDVI) and temperature-vegetation dryness index (TVDI) were employed to further determine whether the identified spots were production and construction projects. Based on GF-1 and GF-6 remote sensing images taken in 2020, covering the entire Henan province, the author used the external rectangle of suspected production and construction project spot to crop the aforementioned images and obtain training samples. The samples were labeled and then input into the AlexNet model for training. After completing the training, two sets of tests were conducted using a single neural network model and a neural network model combined with remote sensing inversion.
      Results 1) This method performs better than a simple convolutional neural network recognition model in reducing both false recognition rate and missed recognition rate, with an average reduction of 23.88% in false recognition rate. It is particularly effective in reducing false recognition rate. 2) This method does not show any significant difference in terms of runtime cost compared to a simple convolutional neural network recognition model.
      Conclusions This achievement can effectively reduce the misclassification rate in the automatic identification of production and construction projects, contributing to the further improvement of the automation level in remote sensing monitoring of soil and water conservation in production and construction projects. It has certain reference value in the application field of automatic identification of production and construction projects.

       

    /

    返回文章
    返回