遥感机理与深度学习双驱动的主粮作物叶片叶绿素含量混合反演方法

The accurate estimation of Leaf Chlorophyll Content (LCC) is of considerable significance for crop physiological monitoring and precision agricultural management. However, traditional Vegetation Indices (VIs) based on visible-near infrared canopy reflectance experience notable challenges in LCC retrieval, as follows: (1) the spectral response is highly coupled with target information (LCC) and structural noise due to the confounding effects of canopy structural signals; and (2) canopy structural heterogeneity across different crop types further exacerbates the sensitivity of VIs to structural parameters, significantly limiting the generalization and applicability of inversion models for major food crops, such as rice, wheat, and maize, under diverse growth conditions. To address these issues, this study proposes a hybrid LCC retrieval framework that synergistically integrates remote sensing physical mechanisms with deep learning techniques, aiming to mitigate effectively the confounding influence of canopy structure and enhance model generalizability across diverse crop types. Specifically, a set of Vegetation Index Ratio Features (VIRFS) that exhibit low sensitivity to Leaf Area Index (LAI) is constructed by simulating a wide range of LAI-LCC combinations by using the PROSAIL radiative transfer model. Furthermore, an active learning-based strategy is incorporated into the transfer learning pipeline to optimize the selection of informative samples, enabling efficient model fine-tuning with limited labeled field data. The proposed method is systematically validated on multi-crop, multiregional datasets that are collected from three major agricultural zones in China: Northeast China (maize, rice, soybean), the Huang-Huai-Hai Plain (wheat), and the Yangtze River Basin (rice). The results show the following: (1) the proposed hybrid method, which integrates radiative transfer mechanisms with deep learning, achieves excellent performance in cross-regional LCC retrieval for major staple crops. It exhibits strong stability and generalizability across diverse crop types with R² consistently exceeding 0.69 and root mean square error (RMSE) lower than 4.77 μg/cm^2. (2) Compared with the conventional vegetation index feature set, the newly constructed VIRFS is designed to be less sensitive to LAI, significantly mitigating canopy structural interference under optimal fine-tuning conditions. Across the three major agricultural regions (Northeast China, the Huang-Huai-Hai Plain, and the Yangtze River Basin), VIRFS improves R² by 0.18—0.23 and reduces RMSE by 1.85—2.51 μg/cm² for various staple crops. (3) The transfer learning framework that incorporates active learning enables high-accuracy LCC estimation by using only 30% of locally labeled samples, achieving R² values of 0.69—0.74 and RMSE of 4.98—5.76 μg/cm² across different staple crops. This result represents a notable improvement over random sampling-based transfer learning, with R² gains of 0.02—0.16 and RMSE reductions of 0.05—1.42 μg/cm^2, substantially enhancing model adaptability and inversion efficiency under limited label conditions. In conclusion, the proposed inversion framework, which coupled physical principles with data-driven methods, significantly improves the accuracy and robustness of multi-crop LCC estimation. It provides a universal solution for nondestructive LCC monitoring across diverse crops and regions.