高光谱遥感技术在环境监测、应急保障、精细地物提取等方面有着广泛的应用，随着高分五号高光谱数据的正式发布，高光谱遥感技术将发挥更重要的作用。遥感影像分类作为高光谱遥感影像信息处理的重要部分，已成为当前研究重点。本文针对传统多级联森林深度学习中模型复杂、无法利用基分类器差异信息、对类间差异较小的样本无法正确区分等不足，提出了一种改进的多级联森林深度学习模型，在模型框架中，分别采用了随机森林和旋转森林作为基分类器，并引入逻辑回归分类器作为判别器用于训练层扩展。相较于传统的深度神经网络，改进的多级联森林深度网络超参数较少且能够自适应确定训练层，更方便进行模型优化。实验采用了高分五号数据集及两个公开的高光谱数据集(Indian Pines数据集及Pavia University数据集)进行精度评定，同时选择了传统分类器支持向量机、深度置信网等模型作为对比分析。实验结果表明，改进的多级联森林深度学习模型能有效地进行高光谱遥感影像分类，且较传统的分类方法精度有所提升。

With the release of GF-5 hyperspectral data, hyperspectral remote sensing plays an increasingly important role in environmental monitoring, emergency management, and object extraction. Classification is the primary problem of hyperspectral image applications. Some cascade forest models have been proposed to overcome the limitations of traditional deep neural networks, such as requiring excessive training samples and optimization of a large number of hyperparameters. Traditional cascade forest models have several disadvantages, such as (1) high model complexity, (2) homogeneous base classifiers, and (3) inability to discriminate the similar spectrum. In this study, a novel classification approach based on cascade forest is proposed to solve the above drawbacks. The proposed improved cascade forest is an accumulation of layers, and each layer consists of two decision tree forests and a logistic regression classifier. Compared with traditional models, the number of forests in the improved method is reduced from four to two with the same accuracy and efficiency. Meanwhile, the original completely random tree forest is replaced by an efficient rotation forest to improve the diversity. The logistic regression classifier is added to determine the separating hyperplane among similar spectra. The number of layers is determined by the accuracy of validation set. The proposed method is implemented on three hyperspectral datasets (GF-5, Indian Pines, and Pavia University datasets). DBN, SVM, RoF, RF, and original cascade forest are selected as the contrast methods. Experimental results on three different real hyperspectral datasets confirms the superiority of the proposed method, especially on the Indian Pines dataset, which has a similar spectrum. The improved cascade model can combine multiple classification results from different base classifiers to obtain the final results through the logistic regression classifier, and the quadratic discriminant process of the logistic regression classifier can effectively improve the classification accuracy. The impacts of the number of trees on the final results are discussed. The proposed model obtains the best performance with optimal parameters. Although the single training time is long, the insensitivity of model parameters immensely improves the training efficiency. Compared with DNN, the improved cascade forest has the following advantages: (1) adaptive to determine the number of layers on the basis of the classification accuracy, (2) few hyperparameters are required in the improved cascade forest, making it easy to optimize the structure, and (3) each forest is independently trained because the training process do not have backpropagation, thereby accelerating the improved cascade forest by the CPU.