基于WEU-Net模型的贺兰山东麓滞洪区水体信息提取

针对经典U-Net模型在贺兰山东麓滞洪区水体信息提取中存在的过拟合、泛化能力有限等问题,基于Sentinel-1合成孔径雷达卫星和Sentinel-2多光谱卫星影像提出了一种水体信息提取卷积神经网络模型(WEU-Net)。WEU-Net模型通过减少编码器与解码器的跳跃连接以及卷积核数量使网络结构简化,并引入残差块增强特征提取能力,弥补了因简化模型而损失的图像信息;在数据集方面,采用逐步回归法结合改进的归一化差异水体指数构建了Sentinel-1水体指数,优化了Sentinel-1卫星影像数据集特征丰富度。试验结果表明:WEU-Net模型预测总体精度为98.19%,F1分数为0.946 9,分别较经典U-Net模型提高了0.357 7%和0.948 8%,训练时长缩短了49.30%；融合Sentinel-1水体指数后,模型预测总体精度和F1分数分别提高了0.51%和3.16%。

Water body information extraction for flood detention area in the eastern foot of Helan Mountain based on WEU-Net model

Aiming at the issues of over-fitting and limited generalization ability of classical U-Net model in extracting the flood retention area in the eastern foot of Helan Mountain, a new convolutional neural network model (WEU-Net) for the water body information extraction was proposed based on images of Sentinel-1 synthetic aperture radar(SAR) satellite and Sentinel-2 multispectral instrument(MSI) satellite.This model simplifies the network structure by reducing the number of convolutional kernels and the skip connection levels between encoder and decoder, and introduces residual blocks to enhance the feature extraction ability, which makes up the loss of image features due to the simplified model.In terms of data set, the Sentinel-1 water index was constructed by stepwise regression method combined with modified normalized difference water index (MNDWI), and the feature richness of data set from Sentinel-1 satellite was optimized.The main conclusions are as follows: the overall accuracy of the WEU-Net is 98.19% and the F1 score is 0.946 9, which are 0.357 7% and 0.948 8% higher than the classical U-Net model, respectively, and the training time is shortened by 49.30%; after fusing the sentinel-1 water index, the overall accuracy and F1 score were improved by 0.51% and 3.16%, respectively.