Real-time leaf chlorophyll content (LCC) is critical for managing farm inputs and monitoring crop growth, productivity and quality of the yield. Visible-near infrared spectroscopy is a non-destructive method for the LCC detection, which plays an increasingly substantial role in the high-throughput monitoring in field. Some detection methods use single or fixed bands, which are not sensitive to LCC at each growth stage and obtain low accuracy and robustness. Thus, we aim to improve the robustness of LCC detection models by exploring deep features of hyperspectral data which could be suitable for the dynamic growth effects. In experiments, the hyperspectral data of four growth stages of jointing, tasseling, silking and blister stages were measured in 2020 and 2021, respectively. Firstly, the LCC variation and spectral response at each growth stage were analyzed. Changes existed at different stages in typical vegetation indices (VIs), which included normalized difference vegetation index (NDVI), normalized difference red edge index (NDRE) and red edge position (REP). Secondly, to capture the features sensitive to LCC at each growth stage, a novel method was proposed to explore deep features hidden among the sensitive wavelengths by combining methods of competitive adaptive reweighted sampling (CARS) and long short-term memory (LSTM), which is labeled as CARS-LSTM. Finally, in order to compare the LCC detection performance of typical VIs and our proposed deep features, the partial least squares regression models were established based on NDVI, NDRE, REP, CARS, LSTM and hybrid deep features of CARS-LSTM, respectively. Result showed that REP performed better than NDVI and NDRE and obtained determination coefficient of prediction set (R_P²) and root mean square error of prediction set (RMSEP) with 0.48 and 4.52 mg/L, respectively, which was possibly due to consistence between REP and LCC and the saturation of NDVI. The wavelengths selected by CARS obtained R_P² and RMSEP of 0.71 and 3.32 mg/L, respectively, and achieved better detection results than REP; The R_P² values of each growth stage were in the range of 0.41–0.72 and the RMSEP values were in the range of 1.23–5.40 mg/L. The proposed deep features of CARS-LSTM achieved the best detection results with R_P² of 0.94 and RMSEP of 1.54 mg/L; The R_P² values of each growth stage were in the range of 0.76–0.96 and the RMSEP values were in the range of 0.89–2.52 mg/L. The research demonstrated that the hybrid deep features of CARS-LSTM could capture the complex spectral changes and help improve the relationship between LCC and spectral data. The proposed method can improve the detection accuracy and robustness of LCC to suit the dynamic growth effects and provide guidance for field monitoring and management.

实时叶片叶绿素含量 (LCC) 对于管理农场投入和监测作物生长、生产力和产量质量至关重要。可见-近红外光谱是一种无损检测LCC的方法，在现场高通量监测中发挥着越来越重要的作用。一些检测方法使用单一或固定波段，对每个生长阶段的 LCC 不敏感，获得的准确性和鲁棒性较低。因此，我们旨在通过探索适合动态增长效应的高光谱数据的深层特征来提高 LCC 检测模型的鲁棒性。实验中分别测量了2020年和2021年拔节、抽雄、吐丝和水疱四个生长阶段的高光谱数据。首先，分析了每个生长阶段的 LCC 变化和光谱响应。典型植被指数（VI）在不同阶段存在变化，其中包括归一化差异植被指数（NDVI）、归一化差异红边指数（NDRE）和红边位置（REP）。其次，为了捕获每个生长阶段对 LCC 敏感的特征，提出了一种新方法，通过结合竞争自适应重加权采样（CARS）和长短期记忆（LSTM）的方法来探索隐藏在敏感波长中的深层特征，被标记为 CARS-LSTM。最后，为了比较典型 VI 的 LCC 检测性能和我们提出的深度特征，分别基于 NDVI、NDRE、REP、CARS、LSTM 和 CARS-LSTM 的混合深度特征建立了偏最小二乘回归模型。结果表明REP优于NDVI和NDRE，得到预测集决定系数( R _P ²) 和预测集的均方根误差 ( RMSEP ) 分别为 0.48 和 4.52 mg/L，这可能是由于 REP 和 LCC 之间的一致性以及 NDVI 的饱和度。CARS选择的波长分别获得了0.71和3.32 mg/L的R _P ²和RMSEP，取得了比REP更好的检测效果；各生长阶段的R _P ²值在0.41-0.72范围内，RMSEP值在1.23-5.40 mg/L范围内。提出的CARS-LSTM深度特征检测效果最好，R _P ²为0.94，RMSEP为1.54 mg/L；R __{每个生长阶段的P} ²值在0.76-0.96范围内， RMSEP值在0.89-2.52 mg/L范围内。研究表明，CARS-LSTM 的混合深度特征可以捕获复杂的光谱变化，有助于改善 LCC 与光谱数据之间的关系。该方法可以提高LCC的检测精度和鲁棒性，以适应动态生长效应，为现场监测和管理提供指导。