【摘要】 利用高分辨率卫星遥感影像提取建成区边界对于城市扩张监测和城市发展规划具有重要意义。为获取高精度高空间分辨率的建成区数据,本研究通过归一化建筑指数(normalized difference built-up index,NDBI)加人工目视解译方法构建城市建成区遥感影像数据集,分别采用传统机器学习方法和包括PSPNet在内的4种深度学习语义分割网络对Sentinel-2影像进行建成区提取,训练结果表明PSPNet网络对于建成区的提取具有最高的精度(训练集交并集比(intersection over umion,IOU)为79.5%)。提出Overlapsize方法对PSPNet的提取结果进行优化,进一步提高了建成区提取准确率,该方法在训练集上的IOU达到80.5%,在测试集上的IOU达到了83.1%,利用PSPNet+Overlapsize提取建成区的方法相较于传统机器学习方法具有实际应用意义。更多还原

【Abstract】 Using high-resolution satellite remote sensing images to extract the boundary of the built-up area is of great significance for urban expansion monitoring and urban development planning. In order to obtain high-precision and high-resolution built-up area data, this study uses the NDBI index and artificial visual interpretation methods to construct remote sensing image datasets of urban built-up areas and uses traditional machine learning methods and four deep learning methods including PSPNet semantic segmentation network to extract the built-up area of Sentinel-2 images. The training results show that the PSPNet network has the highest accuracy for the built-up area extraction( IOU of the training set is 79. 5%). This paper employs Overlapsize method to optimize the extraction results of PSPNet,which further improves the accuracy of the built-up area extraction. The IOU on the training set reaches 80. 5%,and the IOU on the test set reaches 83. 1%. Compared with the traditional machine learning method,the method of PSPNet + Overlapsize has practical application significance in built-up area extracting.更多还原