基于改进DeepLabv3+的涉县石堰梯田遥感提取方法研究  

A method for extraction of Shiyan terraced fields in Shexian County from remote sensing images based on improved DeepLabv3+network

作　　者：刘诗雨 郑文武[1] LIU Shiyu;ZHENG Wenwu(College of Geography and Tourism,Hengyang Normal University,Hengyang 421000,China)

机构地区：[1]衡阳师范学院地理与旅游学院,衡阳421000

出　　处：《时空信息学报》2024年第6期722-731,共10页JOURNAL OF SPATIO-TEMPORAL INFORMATION

基　　金：湖南省研究生科研创新项目(CX20231240)。

摘　　要：梯田是山区可持续发展的重要基础,是实现土地合理利用及粮食安全的重要部分。针对北方旱作梯田尺度变化大且分布较为密集,导致阶梯状梯田难以迅速、准确的获取,本文基于DeepLabv3+模型进行改进研究。首先,在模型中引入轻量级MobileNetV2,提取多种形状和多尺度的梯田特征;其次,通过调整空洞空间金字塔池化的膨胀率,获取更大感受野和上下文信息;引入注意力机制高效通道注意力模块,关注细小目标边界关键信息;最后,以涉县石堰梯田为例,与常用的三种经典图像语义分割模型PSPNet、HRNet及Unet模型进行比较评价。结果表明:本方法的召回率为79.13%,与HRNet、PSPNet模型相比分别提升了2.03%、1.52%;平均交并比提升至70.63%,准确度为91.34%;模型参数量相比U-Net模型减少了约20倍;针对石堰梯田空间分布的识别,本方法可以实现速度和效果的均衡,尤其对具有不规则狭长特征的梯田区域具有较好的分割精度和适应性。Terraces,a critical component of sustainable dryland agriculture in Northern China,serve dual purposes in agricultural production and soil-water conservation.They play a crucial role in land use optimization and food security policy implementation.Distinguishing terraced fields from traditional farmland is essential due to their widespread distribution across complex topographies.Additionally,the large scale and dense distribution of dryland terraces present challenges in accurately and efficiently extracting terrace plots that adapt to terrain variations.To address these issues,this study introduces an enhanced model based on DeepLabv3+,termed DeepLabv3+.The research focused on two main aspects.First,we constructed a semantic segmentation dataset for dryland terraces using hyperspectral image data.This data was then clipped with ArcGIS and annotated with LabelMe through visual interpretation.Data augmentation techniques were subsequently applied to enhance the dataset,laying a solid foundation for model training.Second,we propose the DeepLabv3+semantic segmentation network,featuring a lightweight backbone feature network designed to extract terrace features of varying shapes and scales.We also adjusted the expansion rate of the original ASPP to effectively capture larger receptive fields and contextual information.Furthermore,the ECA module is integrated to help the network focus on key small target boundary details.Improved Deeplabv3+has a recall of 79.13%in the terrace image extraction test,which is 2.03%and 1.52%compared with HRNet and PSPNet,respectively;the MIoU is improved to 70.63%,and the accuracy is 91.34%;and the number of model parameters is reduced about 20 times compared with the U-Net model.The improved algorithm not only reduces the dependence on hardware resources for model training under the premise of ensuring 70%accuracy,thus balancing the remote sensing terrace detection accuracy and speed.The improved Deeplabv3+model consistently outperformed the others.Experimental results demonstrate that

关 键 词：图像语义分割 轻量级网络 DeepLabv3+ 高效通道注意力 目标识别 ASSP 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]