基于无人机多光谱影像与轻量级深度学习的水稻品种鉴别方法研究OA

The rapid and non-destructive identification of rice varieties is an urgent need for precision agriculture and intelligent breeding management.Traditional field survey methods struggle to meet the demands of large-scale,high-throughput variety monitoring.In this study,multispectral unmanned aerial vehicle(UAV)imagery was utilized to acquire canopy data across five rice varieties,constructing a dataset comprising 11,640 individual plant image patches.A dual-track comparative experiment was conducted in parallel using traditional machine learning methods(Method A)and a lightweight deep learning method(Method B).Method A extracted 250-dimensional statistical features(including six vegetation indices)from entire images and 2×2 grid partitions,which were classified using Random Forest(RF),Support Vector Machine(SVM),and Linear Discriminant Analysis(LDA).Method B employed a Multi-Scale Spectral-Spatial Swin Transformer(MSS-Swin-Lite)lightweight network(with approximately 5.16 million parameters)for end-to-end learning directly from multi-scale image patches.Both methods utilized rigorous spatio-temporal isolation and a leave-one-field-out cross-validation strategy for generalization assessment,incorporating pseudo-replication diagnostics to evaluate model reliability.The proposed MSS-Swin-Lite network achieved a superior image patch-level classification precision of 99.04%,outperforming the three traditional classifiers compared.Among traditional schemes,Random Forest performed best with a classification accuracy of 94.38%.The Red Edge(RE)band carried the most significant varietal discrimination information,achieving a single-band fitting accuracy of 86.80%;the RE+NIR combination reached 92.02%.Pseudo-replication diagnostics revealed an intra-class/inter-class distance ratio of 1.39,indicating that the aforementioned accuracy metrics reflect a combined contribution of inherent varietal differences and field-specific environmental signals.Within the validation framework of"same-field,spatial isolation,"the lightweight deep learning method constructed in this study provides a feasible technical solution for UAV-based remote sensing identification of rice varieties.Band ablation experiments offer quantitative guidance for multispectral sensor band selection,while pseudo-replication diagnostics provide a methodological reference for experimental design in variety identification research,thereby supporting the rapid identification of large-scale rice cultivation.