目前，基于神经网络已生成多个全球叶面积指数（Leaf Area Index， LAI） 产品，但这些 LAI 产品使用的训练数据集特定于传感器， 标签信息来源于遥感数据或者模拟数据。 现有训练数据集无法直接应用于其他传感器，并且被认为是“真值”的地面实测数据未被用于构建训练数据集。本研究基于迁移学习理论和 Xiao 等[1]利用 MODIS（Moderate Resolution Imaging Spectroradiometer） 地表反射率数据和 MODIS、 CYCLOPES（ Carbon cycle and Change in Land Observational Products from an Ensemble of Satellites） LAI 产品融合值构建的训练数据集提出了两种基于迁移学习的 LAI 反演方法，从异源遥感数据中反演 LAI 值：第一， 从数据层面实现异源数据集之间的知识迁移，提出了基于无监督域适应的 LAI 反演方法。利用迁移成分分析算法（Transfer Component Analysis， TCA）将 MODIS 和 VIIRS（Visible Infrared Imaging Radiometer Suite） 地表反射率数据集映射到同一子空间以最小化两者的概率分布差异。 然后，利用从 MODIS 地表反射率数据集获得的映射数据和 MODIS、CYCLOPES LAI 产品融合值训练广义回归神经网络（General Regression Neural Networks，GRNNs）。最后，为了反演 LAI 值，将从 VIIRS 地表反射率数据集获取的映射数据输入到已训练的 GRNNs 中。第二，从模型层面实现异源数据集之间的知识迁移，提出了基于深度迁移学习的 LAI 反演方法。 利用深度置信网络（Deep Belief Networks， DBN） 构建可迁移网络， 同时使用 LAI 地面实测数据作为标签数据参与反演。 DBN 由受限玻尔兹曼机层（Restricted Boltzmann Machine， RBM）和反向传播层（Back Propagation Neural Network，BPNN）组成。 在本研究中， RBM 层网络被用于提取地表反射率数据的特征属性， BPNN层网络被用于微调 RBM 网络参数。首先，利用 MODIS 地表反射率数据和 MODIS 与CYCLOPES LAI 产品的融合值预训练 DBN 网络。然后,冻结预训练 DBN 网络的所有 RBM层参数，使用 VIIRS 地表反射率数据和地面实测数据微调 BPNN 层网络参数。最终, 使用微调后的 DBN 网络反演 LAI 值。本文采用贝叶斯优化算法确定 DBN 网络的超参数。 为验证反演结果，将不同植被类型站点处的反演结果同地面实测数据进行直接验证，同时利用MODIS 和 VIIRS 全球 LAI 参数产品与反演结果进行对比验证。 结果表明， 对于不同地类，基于无监督域适应方法和基于深度迁移学习方法的反演结果都表现出合理的季节性和较高的反演精度。 基于无监督域适应的 LAI 反演方法可利用已有训练数据集有效地从其他传感器数据中反演 LAI 参数。 基于深度迁移学习的 LAI 反演方法可利用少量地面实测数据提升反演精度, 实现现有训练数据集的再利用，同时减少对训练样本数量的需求。 

Several global leaf area index (LAI) products were generated using neural networks, but the training dataset for the neural networks was sensor-specific, LAI lables are constructed from remote sensing data and the construction of the training dataset was time-consuming. In this paper, two LAI inversion methods based on transfer learning were proposed to retrieve LAI from heterogeneous remote sensing based on the transfer learning theory and the training dataset constructed by Xiao et al[1] using the moderate resolution imaging spectroradiometer (MODIS) surface reflectance data and the LAI fusion values of MODIS and Carbon cycle and Change in Land Observational Products from an Ensemble of Satellites (CYCLOPES) products. First, the knowledge transfer between different datasets is realized from the level of data, and an LAI inversion method based on unsupervised domain adaptation is proposed. A transfer component analysis (TCA) algorithm was first utilized to map the MODIS and Visible Infrared Imaging Radiometer (VIIRS) surface reflectances into the same subspace to reduce the distribution discrepancies between surface reflectances. Then, the embedded data obtained from MODIS surface reflectance dataset, along with the LAI values produced by fusing the MODIS and CYCLOPES products, were employed to train general regression neural networks (GRNNs). Finally, for retrieving the LAI values, the embedded data acquired from VIIRS surface reflectance dataset was inputted into the trained GRNNs. Second, the knowledge transfer between heterogeneous datasets is realized from the level of model, and the LAI inversion method based on deep transfer learning is proposed. A transplantable network is constructed by using Deep Belief Networks (DBN), and the LAI ground measurements are used as label to participate in the inversion process. DBN is composed of Restricted Boltzmann Machine (RBM) and Back Propagation Neural Network (BPNN). In this study, RBM is used to extract the characteristic informations from surface reflectances, and BPNN is used to fine-tune network parameters. The DBN network is pre-trained by the MODIS surface reflectances and the LAI values fused from the MODIS and CYCLOPES products. Then, the parameters of all RBM of the pre-trained DBN network are frozen, and the parameters of BPNN network are fine-tuned by a small sample composed of VIIRS surface reflectances and LAI ground measurements. Finally, the fine-tuned DBN network is used to retrieve LAI. In this paper, Bayesian optimization algorithm is used to determine the hyperparameters of DBN network. In order to estimate the retrieved results, the retrieved results at sites with different vegetation types are directly evaluated with the ground measured data, and the global LAI products of MODIS and VIIRS are compared with the retrieved results. The results indicate that for different vegetation types, the retrieved results based on unsupervised domain adaptation method and deep transfer learning method proposed in this paper show reasonable seasonality and high retrieval accuracy. The LAI retrieval method based on unsupervised domain adaptation can effectively retrieve LAI from heterogeneous sensor data by using the existing training dataset. The LAI retrieval method based on deep transfer learning can improve the inversion accuracy by using a small amount of LAI ground measurements, reuse the existing training dataset and reduce the demand for the number of training samples.