Answers are grounded in this paper and cite their sources.
Accurate estimation of rice Leaf Area Index (LAI) is important for precision agriculture. Although Unmanned Aerial Vehicle (UAV) RGB imagery offers a cost-effective monitoring approach, traditional methods struggle with background interference in sparse canopies and spectral saturation in dense canopies. Meanwhile, standard deep learning models may exhibit a spectral-bias tendency, in which high-frequency structural cues relevant to dense-canopy RGB saturation can be under-represented. To address these limitations, we propose the CNN-Transformer Frequency-spatial LAI Network (CTFL-Net), an end-to-end frequency-spatial complementary architecture. This dual-stream model extracts local textures via a Convolutional Neural Network (CNN) and global semantics via a Transformer branch. Specifically, a Wavelet Local-Global Fusion (WLGF) module is incorporated to enhance wavelet-domain local-global fusion and help retain frequency-related structural cues. Furthermore, a Hierarchical Refinement Multiscale Fusion (HRMF) strategy is employed to align representations across different feature scales. We constructed and publicly released a comprehensive 3-year, 9,500-image Rice LAI UAV (RLU) dataset for model evaluation. Results on this dataset indicate that CTFL-Net reduced background interference and alleviated the RGB saturation tendency in this cross-year evaluation. Across three independent cross-year test sets, the model achieved stable R 2 values ranging from 0.83 to 0.85 and outperformed the evaluated representative traditional feature-engineering pipelines and vision baselines. It also maintained a relatively low computational footprint of 18.0 GFLOPs. This study provides an efficient, scalable deep learning framework for rice LAI estimation, with practical value for precision field management. The source code and dataset are available at https://github.com/narutoxu-0107/CTFL-Net .
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: