关键词: 水稻, 生育期识别, 遥感, 模型, 深度学习, 多源融合

Abstract:

[Significance] ​The efficient and precise identification of rice growth stages through remote sensing technology holds critical significance for varietal breeding optimization and production management enhancement. Remote sensing, characterized by high spatial-temporal resolution and automated monitoring capabilities, provides transformative solutions for large-scale dynamic phenology monitoring, offering essential technical support to address climate change impacts and food security challenges in complex agroecosystems where precise monitoring of growth stage transitions enables yield prediction and stress-resilient cultivation management.​ [Progress] In recent years, the technical system for monitoring rice growth stages has achieved systematic breakthroughs in the perception layer, decision-making layer, and execution layer, forming a technological ecosystem covering the entire chain of "data acquisition-feature analysis-intelligent decision-making-precise operation". At the perception layer, a "space-air-ground" three-dimensional monitoring network has been constructed: high-altitude satellites (Sentinel-2, Landsat) realize regional-scale phenological dynamic tracking through wide-spectrum multi-temporal observations; low-altitude unmanned aerial vehicle (UAVs) equipped with hyperspectral and light detection and ranging (LiDAR) sensors analyze the heterogeneity of canopy three-dimensional structure; near-ground sensor networks real-timely capture leaf-scale photosynthetic efficiency and nitrogen metabolism parameters. Radiometric calibration and temporal interpolation algorithms eliminate the spatio-temporal heterogeneity of multi-source data, forming continuous and stable monitoring capabilities. Innovations in technical methods show three integration trends: first, multimodal data collaboration mechanisms break through the physical characteristic barriers between optical and radar data. Second, deep integration of mechanistic models and data-driven approaches embeds the PROpriétés SPECTrales-Scattering by arbitrary inclined leaves (PROSAIL) radiative transfer model into the long short-term memory (LSTM) network architecture. Third, cross-scale feature analysis technology breaks through by constructing organ-population association models based on dynamic attention mechanisms, realizing multi-granularity mapping between panicle texture features and canopy leaf area index (LAI) fluctuations. The current technical system has completed three-dimensional leaps: From discrete manual observations to full-cycle continuous perception, with monitoring frequency upgraded from weekly to hourly; from empirical threshold-based judgment to mechanism-data hybrid-driven, the cross-regional generalization ability of the model can be significantly improved; from independent link operations to full-chain collaboration of "perception-decision-execution", constructing a digital management closed-loop covering rice sowing to harvest, providing core technical support for smart farm construction. [Conclusions and Prospects] Current technologies face three-tiered challenges: Data heterogeneity, feature limitations and algorithmic constraints. Future research should focus on three aspects: 1) Multi-source data assimilation systems to reconcile spatiotemporal heterogeneity through UAV-assisted satellite calibration and GAN-based cloud-contaminated data reconstruction; 2) Cross-scale physiological-spectral models integrating 3D canopy architecture with adaptive soil-adjusted indices to overcome spectral saturation; 3) Mechanism-data hybrid paradigms embedding thermal-time models into LSTM networks for environmental adaptation, developing lightweight CNNs with multi-scale attention for occlusion-resistant panicle detection, and implementing transfer learning for cross-regional model generalization. The convergence of multi-source remote sensing, intelligent algorithms, and physiological mechanisms will establish a full-cycle dynamic monitoring system based on agricultural big data.

Key words: rice, growth period, remote sensing, models, deep learning, multi-source fusion