作物理化参数与作物生长条件、氮含量、产量等有着极大的相关性，能够反映植物的衰老过程，是表征作物生长健康状况和作物长势监测的重要指标。遥感观测通过提供准确的作物生物理化参数，如叶面积指数（Leaf Area Index，LAI）和叶片叶绿素含量（Leaf Chlorophyll Content，LCC），为精准农业的发展提供了关键的技术支持。因此，实时准确地估计作物理化参数含量对于监测作物生长健康状况以及优化农业管理具有重要意义。

       本文以黑龙江省齐齐哈尔市农业科学院克山农场马铃薯为研究对象，将辐射传输模型（Radiative Transfer Model，RTM）与作物叶片、冠层ASD光谱以及无人机多光谱影像数据集结合，采用主动学习（Active Learning，AL）及机器学习算法进行作物关键长势参数反演。研究内容及主要结论如下：

    （1）马铃薯叶片尺度LCC反演研究。研究内容主要为基于2022年马铃薯不同生育期地面实测数据，使用马铃薯不同生育期地面实测数据作为标签样本，采用六种主动学习方法从PROSPECT-4模型生成的LUT模拟数据集中选择训练样本，然后使用混合方法对马铃薯LCC进行反演，最后将RTM，高斯过程回归（Gaussian Process Regression，GPR）和AL集成的混合方法与基于成本函数（Cost Function，CF）的查找表（Look-up Table，LUT）反演方法相比较。同时，着重分析了不同的主动学习方法在选择训练数据方面的精度比较以及评估了不同的反演方法的性能。结果表明，与LUT CF方法相比，使用GPR_PROSPECT-AL构建的混合模型具有更高的建模精度，这表明基于RTM可以生成足够大的训练数据集，并用于LCC模型反演，而AL方法有助于优化RTM模拟数据集，提高模型的准确性。

    （2）马铃薯冠层尺度LCC、LAI和CCC反演研究。主要以2022年和2023年两年黑龙江农垦总局齐齐哈尔分局克山农场马铃薯试验为研究对象，采用PROSAIL辐射传输模型构建LUT作为模拟数据集，基于主动学习样本池策略，利用马铃薯不同生育期实测数据作为标签样本集，将PROSAIL模型生成的LUT数据集作为未标注样本集，基于欧式距离基多样性（Euclidean distance-based Diversity，EBD）算法从模拟样本池中选取样本，结合GPR算法进行马铃薯LCC、LAI及CCC参数建模。本文主要评估了主动学习算法在不同标签数据下选择数据的性能以及评估所构建的模型的泛化性。研究结果显示，针对不同反演参量，利用同一套光谱标签数据，达到最优建模精度时筛选的建模样本规模差异较大。对于LCC，需要选择模拟数据集4 %左右的样本，模型验证精度可以达到较高水平；对于LAI，在块茎形成期则需要比其余生育期更多的训练样本（12 %）才可以达到较好的精度。在考虑品种信息时，基于不同标签样本，对于LCC，品种维拉斯需要较其余两个品种多10 %（732/5000）的训练样本才可以达到最优建模精度；对于LAI，在品种垦署时，则需要22 %（1105/5000）训练样本可以取得较好精度；对于CCC，在垦署时较其余两个品种需要15 %（745/5000）的训练样本可以得到较好精度。对马铃薯LCC，LAI和CCC生育期模型及品种模型的分析也显示，不同品种之间模型泛化性差异较大。因此，在进行标签数据选择时，应首先考虑作物生育期的影响，同时，也需要考虑到品种差异对主动学习中训练样本的筛选影响。

    （3）马铃薯无人机冠层反射率LCC和LAI反演研究。主要使用2022年和2023年两年的地面实测冠层反射率数据，旨在探索将先验知识和主动学习结合，并采用混合方法从冠层反射率反演LCC和LAI。首先使用主动学习从模拟数据集中筛选与实测数据较为匹配的训练数据集，然后采用先验知识和主动学习结合的混合方法对马铃薯单生育期和全生育期LCC和LAI构建模型，并评估其混合反演模型性能。研究结果表明，对于LCC，模型验证精度在淀粉积累期表现最好，R²=0.684，RMSE=26.722 µg/cm²，NRMSE=66.980 %，其次为块茎形成期，表现最差的为块茎增长期；块茎形成期和淀粉积累期模型验证精度较未结合先验知识的R²提高了0.077和0.001，而块茎增长期则降低了0.030；同时，全生育期模型验证精度也表现出较好的验证精度，R²=0.396，RMSE=22.704 µg/cm²，NRMSE=43.600 %；对于LAI，模型精度也在淀粉积累期表现最好，R²=0.703，RMSE=0.567 m²/m²，NRMSE=13.584 %，仅次于它的是块茎增长期，表现较差的是块茎形成期，不同生育期模型验证精度较未结合先验知识的R²提高了0.100，0.031和0.014；RMSE分别降低0.061，0.049和0.013 m²/m²。

        Crop physico-chemical parameters are highly correlated with crop growth conditions, nitrogen content and yield. It can reflect the plant senescence process and is an important indicator for characterising the health of crop growth and monitoring crop growth. Remote sensing observations play a key technical support role in the development of precision agriculture by providing accurate crop biochemical parameters, such as Leaf Area Index (LAI) and Leaf Chlorophyll Content (LCC). Therefore, real-time accurate estimation of crop physicochemical parameter content is important for monitoring crop growth and health as well as optimising agricultural management.

        In this study, we took potato at Keshan Farm, Academy of Agricultural Sciences, Qiqihar City, Heilongjiang Province, as the research object. The study combined the radiative transfer models (Radiative transfer model, RTM) with crop leaf and canopy ASD spectra as well as unmanned aerial vehicle (UAV) multispectral imagery datasets, and used active learning (AL)  and machine learning algorithms to perform inversion of key crop growth parameters. The research content and main conclusions are as follows:

        (1) Research on leaf-scale chlorophyll content and LAI inversion in potato. The study is mainly based on the ground-measured data of potato at different growth periods in 2022, using the ground-measured data of potato at different growth periods as labeled samples, selecting training samples from the LUT simulation dataset generated by the PROSPECT-4 model by using six active learning methods. Then, the potato LCC was then inverted using a hybrid method, and finally the hybrid method integrating the RTM, Gaussian process regression (GPR) and AL was combined with the look-up table (LUT) inversion method based on Cost function (CF). It also focuses on analyzing the comparative accuracy of different active learning methods in selecting training data as well as evaluating the performance of different inversion methods. The results show that the hybrid model constructed using GPR_PROSPECT-AL has higher modeling accuracy compared to the LUT CF method, which suggests that the RTM-based can generate a sufficiently large training dataset that can be used for LCC model inversion, and the active learning approach helps to optimize the RTM simulation dataset to improve the accuracy of the model.

        (2) Potato canopy-scale chlorophyll content and LAI inversion study. This study used data from potato trials at Keshan Farm, Qiqihar Branch, Heilongjiang Agricultural Reclamation Bureau, in 2022 and 2023. The simulated LUT data set generated by the PROSAIL radiative transfer model was used as an unlabeled sample set. The measured data from the potato field was divided into different labeled sample sets by growth periods and varieties. Then, the informativeness training samples for potato LCC, LAI, and CCC parameters are selected from the simulated unlabeled sample set through the Euclidean distance-based Diversity (EBD) algorithm based on different labeled data sets. The measured data then assess the accuracy and generalization of the models to evaluate the pool-based sampling strategies used in this study. This study focuses on evaluating the performance of active learning algorithms for selecting data with different labeled data and assessing the generalisability of the constructed models. The study results indicated that the selected training sample size varied considerably for different parameters, even using the same labeled data set. For LCC, it needs to select about 4 % of the training samples from the simulated dataset to guarantee a high accuracy level of the model. For LAI, it needs more than (12 %) training samples during the potato tuber formation period to reach a reasonable accuracy. When considering the potato varieties information, for LCC, the potato variety WLS needs 10 % more training samples than the remaining two varieties to achieve optimal modeling accuracy. For LAI, variety Kenshu requires 22 % training samples to achieve better accuracy. For CCC, Kenshu needs 15 % training samples compared to the remaining two varieties to get better accuracy. Therefore, the results of the model generalisability analysis show that the effect of the crop growth period should be considered first when selecting labeled data from the measured data set, and it is also necessary to consider the effect of varietal differences on selecting training samples in active learning.

        (3) Potato UAV canopy reflectance chlorophyll content and LAI inversion study. This study used ground-measured canopy reflectance data for two years, 2022 and 2023. It aims to explore the combination of a priori knowledge and active learning with a hybrid method to invert LCC and LAI from canopy reflectance. First, active learning was used to select training datasets from the simulated dataset that better matched the measured data, and then a hybrid method combining a priori knowledge and active learning was used to construct models for LCC and LAI during single and full potato growth periods, and to evaluate the performance of the hybrid inversion models. The results showed that for LCC, the model validation accuracy performed best at the starch accumulation period with R² = 0.684, RMSE=26.722 µg/cm² and NRMSE=66.980 %, followed only by the starch accumulation period for tuber formation period, and the worst performance was at the tuber growth period. The model validation accuracies for the tuber formation period and the starch accumulation period were improved compared to those of the R² without the incorporation of a priori knowledge by 0.077 and 0.001, while the tuber growth period decreased by 0.030. Meanwhile, the model validation accuracy for the whole-plantation period also showed better accuracy with R²=0.396, RMSE = 22.704 µg/cm² and NRMSE=43.600 %. For LAI, the model accuracy also showed the best performance in the starch accumulation period with R²=0.703, RMSE=0.567 m²/m² and NRMSE=13.584 %. Next to it was the tuber growth period, and the worse performance was in the tuber formation period, where the model validation accuracy of the different growth periods was improved by 0.100, 0.031, and 0.014 compared to the R² that did not incorporate a priori knowledge; RMSE was reduced by 0.061, 0.049 and 0.013 m²/m² respectively.