基于DMFM-Net的冬小麦氮营养指数反演研究OA

Real-time and accurate monitoring of the winter wheat nitrogen nutrition index(NNI)is of great significance for optimizing field water and fertilizer management and reducing agricultural non-point source pollution.To address the issues of modal confusion in existing deep networks when processing multi-source UAV remote sensing data,as well as the loss of two-dimensional spatial topological information during temporal modeling,this study proposes a dual-branch multi-source feature fusion memory network(DMFM-Net).Multispectral images and canopy structural data of winter wheat at the tillering,jointing,and booting stages were collected by UAV.An enhanced CBAM-ResNet50 dual-branch backbone was employed to separately extract spectral features and canopy structural features.The spatially preserved feature tensors were then directly input into a convolutional long short-term memory(ConvLSTM)network for temporal modeling.The results show that:(1)Compared with the traditional serial network(CBAM-ResNet-LSTM)and the single-branch network(CBAM-ResNet-ConvLSTM),DMFM-Net effectively overcomes the noise interference of multimodal heterogeneous data and achieves the best inversion performance at all key growth stages.The coefficients of determination(R2)of the model at the tillering,jointing,and booting stages reach 0.76,0.86,and 0.82,respectively,with root mean square errors(RMSE)as low as 0.11,0.10,and 0.14.(2)Ablation experiments further confirm the necessity and scientific validity of the proposed architectural design.The ConvLSTM model lacking a deep spatial encoder shows a significant decrease in R2 of 0.21-0.24.Moreover,compared with the static spatiotemporal feature extraction of 3D-ResNet,DMFM-Net,by virtue of its ConvLSTM module covering a long-range dynamic memory mechanism across the whole growth period,improves the R2 of the inversion models at each stage by 0.07-0.16.(3)The field-scale spatial distribution maps generated by the optimal model accurately reconstruct the local spatial heterogeneity and temporal evolution patterns of winter wheat nitrogen nutritional status under different nitrogen application gradients.In summary,this study provides a new theoretical paradigm and technical support for multimodal time-series dynamic monitoring of crop nutritional status in complex field environments.