基于DenseNet模型的无人机高分辨率影像树种分类研究  

Tree Species Classification Based on DenseNet Model in UAV High Resolution Images

作　　者：张泽宇 王妮[1,2] 朱锦富 孟祥端 董思萌 

机构地区：[1]滁州学院地理信息与旅游学院,安徽 滁州 [2]滁州学院实景地理环境安徽省重点实验室,安徽 滁州

出　　处：《测绘科学技术》2021年第4期139-145,共7页Geomatics Science and Technology

摘　　要：如今深度学习广泛应用于医学、工业、人工智能以及地理学等领域。本文基于DenseNet模型,在其残差块之间加入1 ×1的小型卷积核作为瓶颈层得到了一种改进的DenseNet_BL模型,以琅琊山林场为研究区,使用DenseNet121_BL和DenseNet169_BL模型对研究区的无人机高分辨率光学影像进行分类研究实验。得到的实验结果表明DenseNet121_BL模型在进行树种分类时正确率最高,达到了88.29%。说明改进后的DenseNet_BL模型是一种有效的树种分类算法。Deep learning is widely used in medicine, industry, artificial intelligence, geography and other fields. This paper proposes an improved DenseNet_BL model based on DenseNet model. An improved DenseNet_BL model is obtained by adding a 1 ×1 small convolution kernel between the Residual Blocks as the Bottleneck Layer. Taking Langya Mountain Forest as the research area, DenseNet 121_BL and DenseNet169_BL models were used to classify UAV high-resolution optical images in the research area. The experimental results showed that DenseNet_BL121 model had the highest accuracy in tree species classification, reaching 88.29%. The improved DenseNet_BL model is an effective tree species classification algorithm.

关 键 词：DenseNet 残差网络 无人机 深度学习 树种分类 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]