基于改进DeepLabV3+深度学习模型的冬小麦种植面积提取研究  

Research on Extraction of Winter Wheat Planting Area Based on Improved DeepLabV3+

作　　者：路秋叶 刘法军 丁志国 郭鹏 宫锟霖 LU Qiuye;LIU Fajun;DING Zhiguo;GUO Peng;GONG Kunlin(College of Information Science and Engineering,Shandong Agricultural University,Tai'an 271018,China;The Fifth Geological Brigade,Shandong Geological and Mineral Exploration and Development Bureau,Tai'an 271018,China)

机构地区：[1]山东农业大学信息科学与工程学院,山东泰安271018 [2]山东省地质矿产勘查开发局第五地质大队,山东泰安271018

出　　处：《无线电工程》2023年第11期2564-2572,共9页Radio Engineering

基　　金：山东省自然科学基金(ZR2020MD017,ZR2015DL003)。

摘　　要：冬小麦种植面积的精确提取对粮食估产、农业政策制定、科学研究等具有重要意义。针对原始深度学习语义分割模型DeepLabV3+存在的分割效果差、提取精度低、模型复杂以及训练时间长等问题,采用MobileNetV2主干网络、注意力机制、损失函数、组归一化和条带池化对其网络模型进行改进,以高分二号影像为数据源,利用改进的模型进行冬小麦面积的提取。结果表明,单一改进模块融入对网络模型提升不高,将MobileNetV2主干网络、卷积块的注意力模块(Convolutional Block Attention Module,CBAM)、Focal Loss+Dice Loss函数、组归一化(Group Normalization,GN)和条带池化(Stripe Pooling)模块都有效融入网络模型时,相比未改进的DeepLabV3+网络模型MIoU、MPA、Accuracy分别提高了5.22%、2.43%、2.77%,利用改进模型提取的夏张镇冬小麦面积与政府统计数据相比较,相对误差仅为-7.11%,提高了1.56%;结合外业采样点,夏张镇和徂徕镇的冬小麦面积提取精度分别达到93.91%、92.31%,有力证明了改进算法在冬小麦面积提取模型中的有效性。The accurate extraction of winter wheat planting area is important for grain yield estimation,agricultural policy making,and scientific research.To address the problems of poor segmentation,low extraction accuracy,complex model,and long training time of the original deep learning semantic segmentation model DeepLabV3+,MobileNetV2 backbone network,attention mechanism,loss function,group normalization,and strip pooling are used to improve its network model.Using the Gaofen-2 images as data source,the improved model is used to extract the area of winter wheat.The results show that incorporating a single improved module does not improve the network model much.When the MobileNetV2 backbone network,Convolutional Block Attention Module(CBAM),Focal Loss+Dice Loss function,Group Normalization(GN),and strip pooling module are effectively incorporated into the network model,compared with the unimproved DeepLabV3+network model,the MIoU,MPA and Accuracy are improved by 5.22%,2.43%and 2.77%respectively.Compared with the government statistical data,the relative error of the winter wheat area extracted by the improved model is only-7.11%,which is 1.56%higher than the previous DeepLabV3+algorithm.Combined with the field sampling points,the accuracy of winter wheat area extraction in Xiazhang Town and Culai Town reaches 93.91%and 92.31%,respectively,which strongly proves the effectiveness of the improved algorithm in winter wheat area extraction model.

关 键 词：深度学习 冬小麦 DeepLabV3+模型 高分二号 语义分割 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]