"作物-秸秆-土壤"图像分割与比例提取方法研究OA

[Objective]This paper presents an in-depth research on the"crop-straw-soil"segmentation and proportion extraction methods to improve the automation level of field straw coverage surveys.[Method]High-definition digital images of crop,straw,and soil were collected in field conditions.U-Net and DeepLabV3+models with backbone networks including VGG16,ResNet50,and Efficient-Net_B0,were designed for"crop-staw-soil"segmentation.Based on the segmentation results,the model with the best performance was selected for proportion extraction.Segmentation accuracy was evaluated using mean intersection over union(MI),Ac,and Re.The accuracy of the proportion extrac-tion was assessed by the coefficient of determination(R2)and root mean square error(RM).[Result]The ResNet50-based U-Net achieved precise segmentation of farmland images with an MI of 80.96%and an accuracy of 90.00%,significantly outperforming the ResNet50-based DeepLabV3+model and other main trunk feature netwok,with an MI of 80.78%and an Ac of 89.92%.The ResNet50-based U-Net achieved the highest accuracy coverage extraction,with an R2 of 0.851-0.979 and an RM of 3.220%-8.554%.[Conclusion]The ResNet50-based U-Net model can accurately extract"crop-straw-soil"coverage information from farmland digital images,providing technical support for dynamically monitoring farming progress and promoting agricultural ecological environmental protection.