基于DeeplabV3+的无人机遥感影像识别  

UAV Remote Sensing Image Recognition Based on DeeplabV3+

作　　者：赖丽琦 LAI Liqi(School of Information Engineering,Zhejiang Agriculture and Forestry University,Hangzhou 311300,China)

机构地区：[1]浙江农林大学信息工程学院,浙江杭州311300

出　　处：《林业调查规划》2021年第3期11-16,62,共7页Forest Inventory and Planning

摘　　要：针对无人机植被覆盖航拍影像分类精度不高、工程量大的问题,开发适用于无人机植被覆盖影像识别、具有较高准确性的语义分割模型,以提高植被分类精度。通过无人机航拍获取研究区影像,对获取到的图像进行标注,建立图像数据集,构建随机森林、SegNet、U-Net及DeeplabV3+模型,对DeeplabV3+模型进行改进,将主干网络替换为MobileNetV2,在本实验数据集上训练和测试模型,以验证集指标作为模型评估指标。结果表明,语义分割模型表现优于传统图像分割模型;原始模型中,DeeplabV3+模型表现最好,像素准确率达93.64%,平均交并比达66.43%;改进的DeeplabV3+模型在植被覆盖影像上取得较好的识别效果,像素准确率和平均交并比分别为94.71%和70.89%,相比原始模型分别提升了1.07%和4.46%,参数量仅为原始模型的6.74%,训练时长为原始模型的78.76%,具有一定的适用性和实用价值。Aiming at the problems of low classification accuracy and large amount of work of UAV vegetation cover aerial images,the semantic segmentation model with high accuracy for UAV vegetation cover image recognition was developed to improve the accuracy of vegetation classification.The images of the study area were acquired by UAV aerial photography,and the acquired images were labeled.The image data set was established,and the random forest,SegNet、U-Net and DeeplabV3+models were constructed.The DeeplabV3+model was improved,and the backbone network was replaced by MobileNetV2.The models were trained and tested on the experimental data set,and evaluated by validation test.The results showed that the semantic segmentation model performed better than the traditional image segmentation model;among the original models,the DeeplabV3+model performed best,with a pixel accuracy rate of 93.64%and a mean intersection-over-union of 66.43%;the improved DeeplabV3+ model performed well with pixel accuracy of 94.71%and mean intersection-over-union of 70.89%on the validation set,which was 1.07%and 4.46%higher than that of the original DeeplabV3+model,while the parameter quantity and the training duration was 6.74%and 78.76%of the original model respectively,which had certain applicability and practical value.

关 键 词：无人机 语义分割模型 DeeplabV3+模型 植被覆盖影像 参数量 训练时长 

分 类 号：S718.521.2[农业科学—林学] S771.51