基于改进U-Net网络模型的高分影像养殖池塘提取

Extraction of high resolution image aquaculture pond based on improved U-Net network model

针对养殖池塘提取难的问题,使用深度学习方法对高分辨率遥感影像中养殖池塘进行精细提取。研究基于0.5 m高分辨率遥感卫星影像,采用DenseNet网络结构作为U-Net网络模型的编码器,使用手工矢量化的养殖池塘样本对改进U-Net网络模型进行训练,并用训练后的网络模型对验证集影像中养殖池塘进行提取。结果表明,改进U-Net网络模型提取精确率、召回率、交并比分别达92.77%、92.21%、85.60%。与面向对象方法和D-LinkNet模型方法对比,改进U-Net网络模型效果最佳。该模型为养殖池塘精细提取提供了新的思路与方法,有利于推进渔业养殖资源确权调查及精细化管理。

Aiming at the problem of difficult extraction of aquaculture ponds,the deep learning method was used to extract aquaculture ponds from high spatial resolution remote sensing images. Based on the 0.5 m high-resolution remote sensing satellite images,this study adopted DenseNet network structure as the encoder of the U-Net network model,used the hand-marked aquaculture pond training samples to train the improved U-Net network model,and used the network model after training to extract the aquaculture ponds in the validation set images. Resulte showed that the precision rate,recall rate,and intersection over union of the improved U-Net network model were 92.77%,92.21% and 85.60% respectively. Compared with the object-oriented method and the D-LinkNet model method,the improved U-Net network model possessed the best result. This model provided a new idea and method for fine extraction of aquaculture ponds,which was beneficial to promote the investigation and fine management of aquaculture resources.