基于卷积神经网络的城市水体提取方法研究

针对传统方法在城市水体提取中容易受到建筑物阴影影响和难以精确提取细小水体等问题,提出了一种基于逐像元分类和多尺度分割技术的卷积神经网络遥感水体提取方法。该方法利用像元的光谱特征向量构建光谱特征矩阵,作为卷积神经网络输入特征训练水体提取模型,以多尺度分割结果抑制分类离散点与水体边缘误分现象,进一步提高提取精度。试验结果表明,该方法在细小水体的提取精度和细节上比改进的归一化水体指数算法表现更好,不仅能有效抑制建筑物阴影的影响,还能够有效区分一些相对细小的建筑对象如桥梁等,提取结果边缘也更光滑。

Research on Convolutional Neural Networks for Urban Water Body Extraction

Aiming at the problems that traditional water body extraction methods are easy to be affected by building shadows and it is difficult to accurately extract small water bodies,this paper proposes a method based on convolutional neural network with pixel-by-pixel classification and multi-scale segmentation technology for remote sensing water body extraction in city area.The spectral feature matrix is constructed by spectral feature vector of each pixel,which is used as the input feature for convolutional neural network training to construct the water feature extraction model.The multi-scale segmentation technique-is used to reduce discrete points and water body edge misclassification in classification results.The experimental results show that this method is better than the traditional method-MNDWI in both rate of accuracy and details,which can not only suppress the influence of building shadows,but also distinguish some small constructors as bridge,and the extraction results are smoother.