基于深度语义分割的无人机影像烟草种植面积提取

Extraction of Tobacco Planting Area from UAV Images Based on Deep Semantic Segmentation

目前,深度学习广泛用于作物信息提取,但是存在模型复杂以及训练时间较长的问题。为此,提出一种基于DeeplabV3+深度语义分割模型的无人机遥感影像烟草种植面积精确提取方法。首先通过采用四种经典的轻量化骨干网络(Xception、MobilenetV1、MobilenetV2和Resnet50)替换DeeplabV3+原始的空洞卷积结构(Atrous Conv)结构来训练数据集;其次比较分析不同骨干网络对烟草特征的提取能力;最后,通过训练的模型对预测图像进行语义分割,并使用平均交并比(mean Intersection over Union,mIoU)来评价准确率。实验结果表明,使用Xception、MobelnetV1、MobelnetV2和Resnet50对71张预测图像进行语义分割,得到的mIoU分别为95.58、93.95、94.86和90.45,语义分割准确率较高。

At present,deep learning is widely used for crop information extraction,but there are problems of complex model and long training time.For this reason,this paper proposes an accurate extraction method of tobacco planting area from UAV remote sensing images based on DeeplabV3+deep semantic segmentation model.First,it trains the dataset by replacing original DeeplabV3+ASPP(atrous spatial pyramid pooling)structure with four classic lightweight backbone networks(Xception,MobilenetV1,MobilenetV2,and Resnet50);Then,it compares and analyzes the extraction capabilities of different backbone networks for tobacco features;Finally,the predicted images are segmented semantically through the trained model,and the mIoU(mean Intersection over Union)is used to evaluate the accuracy.The experimental results indicate that by using Xception,MobelnetV1,MobelnetV2 and Resnet50 to perform semantic segmentation on 71 predicted images,the mIoU obtained are 95.58,93.95,94.86 and 90.45,respectively,and the accuracy of semantic segmentation is high.