Forage is an essential resource in grassland ecosystems. The efficient identification of forage species facilitates the characterization of germplasm structure, which is vital for breeding superior cultivars and for the intelligent development and remote sensing monitoring of the forage industry. Current identification methods primarily rely on conventional techniques and manual assessment, limiting efficiency and scalability for large-area evaluations. To address these limitations, we integrated unmanned aerial vehicle (UAV) hyperspectral imagery with deep learning to improve forage identification accuracy. We propose a novel Multi-scale Fused Octave Convolution Network (MFO-Net). The MFO-Net suppresses spatial redundancy, captures multi-scale contextual information, and integrates multi-dimensional features. It employs a frequency-splitting strategy to create high and low frequency processing paths. Within this structure, a spectral attention mechanism generates weighting coefficients, enabling dynamic weight adjustment. The model was evaluated against six metrics, including Overall Accuracy (OA), and its effectiveness was validated through comparative experiments with other deep learning models. The experimental results demonstrate that MFO-Net achieved an OA of 99.95%. This approach establishes a new paradigm for high-precision, large-scale forage identification and classification.
Liu et al. (Fri,) studied this question.