基于深度学习的新增建设用地信息提取试验研究: 全国土地利用遥感监测工程创新探索

目前基于遥感图像的城市功能区分类方法通常采用光谱特征解译、兴趣点数据辅助、评价策略优化等方式，依赖大量人工操作，并借助遥感图像外的其他信息源。为了解决以上问题，提出利用卷积神经网络进行滑窗识别，提取图像语义标签，结合语义推理机制实现城市功能区分类的滑窗-推理方法。首先，建立两级城市功能区分类，以二级城市功能区为标识标注训练样本，并以此训练卷积神经网络作为识别器；然后，设计有重叠的滑窗识别模式，使用识别器辨识滑窗区域内图像块的二级城市功能区类型；最后，提出一个带权重的打分机制，作为语义推理方式，语义推理对象为全部识别结果，确定各图像块的一级城市功能区类型，实现遥感图像城市功能区分类。实验使用模拟图像和高分辨率遥感图像，两种图像的总分类精度分别可达94.50%、92.04%。滑窗-推理方法旨在通过语义推理处理滑窗识别产生的多语义标签，根据多语义标签确定对象的真实城市功能区。实验结果表明，所提方法无需辅助信息，直接利用遥感图像进行城市功能区分类是可行和有效的。

Research on New Construction Land Information Extraction Based on Deep Learning: Innovation Exploration of the National Project of Land Use Monitoring via Remote Sensing

  Objectives  Although remote sensing image is one of the main data sources for the classification of urban functional areas, it is rare to use remote sensing images to classify urban functional areas and extract their attribute information. At present, the classification methods of urban functional areas based on remote sensing images usually need manual interpretation, point of interest data assistance, information questionnaire survey and so on. This kind of method not only needs a lot of manual operation, but also needs other external information sources except remote sensing images.  Methods  In order to solve the above problems, a sliding window-reasoning method for urban functional area classification is proposed by using convolutional neural network (CNN) to identify sliding window, extract image semantic tags, and combine semantic reasoning mechanism, which can quickly realize urban functional area classification only by remote sensing images. First, a two-level classification table of urban functional areas is established, and the training samples are marked with the second-level urban functional areas, and the training CNN is used as the recognizer. Second, using the designed overlapping sliding window recognition pattern, the trained re-cognizer is used to identify the types of features in the sliding window area, and to determine the type of urban functional area. Finally, a weighted scoring mechanism is designed as the implementation of semantic reasoning, and the semantic reasoning objects are all the recognition results. And, the type of urban functional areas of each region is determined, and the urban functional areas of large-scale remote sensing images are classified.  Results  Using simulated images and high-resolution remote sensing image experiments, the total classification accuracy of simulated image experiments based on confusion matrix can reach 94.50%, and the total classification accuracy of real remote sensing image experiments based on confusion matrix can reach 92.04%.  Conclusions  The purpose of the sliding window-reasoning method is to deal with the multi-semantic label results produced by the sliding window recognition through the semantic reasoning method, and to determine the real urban functional area results of the identified objects through multiple semantic tags. The results show that the method of CNN sliding window recognition combined with semantic reasoning is feasible and effective to classify urban functional areas directly using large-scale remote sensing images without the assistance of other information.