农业机械学报2023，Vol.54Issue(z2)：206-213,8.DOI:10.6041/j.issn.1000-1298.2023.S2.024
基于改进CBAM-DeepLab V3+的苹果种植面积提取
Apple Planting Area Extraction Based on Improved DeepLab V3+
摘要
Abstract
To improve the accuracy of apple cultivation area extraction,a CBAM-DeepLab V3+model based on the fusion of Sentinel-2 and MODIS satellite images was proposed.The main factors affecting the accuracy of apple cultivation area extraction included the quality of remote sensing images and the performance of semantic segmentation models.From the perspective of image quality,a time-series spatiotemporal fusion algorithm called ESTARFM was employed to fuse Sentinel-2 and MODIS remote sensing image data,achieving higher spatial and temporal resolution data.Simultaneously,the training samples were increased from the original 800 to 2 400,providing more abundant sample capacity for the subsequent semantic segmentation model.In terms of optimizing the semantic segmentation model,in order to further improve the accuracy of apple cultivation area extraction,a CBAM attention mechanism based on channel and spatial information was introduced into the DeepLab V3+network,resulting in the development of the CBAM-DeepLab V3+model.Compared with the original DeepLab V3+model,the CBAM-DeepLab V3+model with the addition of CBAM attention mechanism achieved significant breakthroughs in terms of slower fitting speed,less accurate edge target segmentation,inconsistency in segmenting large-scale targets,and existence of holes.These improvements enhanced the training and prediction performance of the model.The original Sentinel-2 images and the spatiotemporal fusion images were used,combined with the datasets of Wanggezhuang Town in Muping District and the apple dataset of Guanshui Town to compare the U-Net,FCN,DeepLab V3+models,and the CBAM-DeepLab V3+model.The research findings indicated that in terms of apple cultivation area extraction,the overall accuracy(MIoU)achieved by the optimized CBAM-DeepLab V3+model was 84.6％,and the accuracy of apple cultivation area extraction reached 90.4％.In comparison,the MIoU of U-Net,FCN,and DeepLab V3+models were 79.2％,75％,and 81.2％,respectively.Additionally,the predicted apple cultivation area of Wanggezhuang Town in Muping District was 3 433.33 hm2,with only 233.33 hm2 deviation compared with the data of 3 666.66 hm2 published in the Yantai City National Economic and Social Development Statistics Report,resulting in a high prediction accuracy of 93.64％.关键词
苹果种植面积提取/时空融合/卷积神经网络/DeepLab V3+/语义分割Key words
apple planting area extraction/spatiotemporal fusion/convolutional neural network/DeepLab V3+/semantic segmentation分类
信息技术与安全科学引用本文复制引用
常晗,郭树欣,张海洋,张瑶..基于改进CBAM-DeepLab V3+的苹果种植面积提取[J].农业机械学报,2023,54(z2):206-213,8.基金项目
中央高校基本科研业务费专项资金项目(2023TC131) （2023TC131）
