To solve the problem that the traditional algorithms, such as Support Vector Machine (SVM) and random forest, cannot make full use of the texture features and optical parameters of satellite images, a method of cloud/snow recognition based on Multi-dimensional multi-grained cascade Forest (M-gcForest) was proposed. Firstly, according to the difference between single-spectral and multi-spectral images, SVM, random forest, Convolution Neural Network (CNN), and gcForest (multi-grained cascade Forest) were selected to recognize cloud and snow on single-spectral satellite images, by quantitatively analyzing the performance of each algorithm on single-spectral images, CNN and M-gcForest were selected for multi-spectral cloud/snow recognition. Finally, improved M-gcForest was used to predict on HJ-1A/1B multi-spectral satellite images. The experimental results show that compared with CNN, the test accuracy of the M-gcForest on the multi-spectral dataset is increased by 0.32%, the training time is reduced by 91.2%, and the testing time is reduced by 53.7%. Therefore, the proposed algorithm has practicability in real-time and accurate snow disaster monitoring tasks.

To solve the problem of gradient vanishing caused by too many layers of Convolutional Neural Network (CNN) in deep reinforcement learning, a deep reinforcement learning method based on weighted densely connected convolutional network was proposed. Firstly, image features were extracted by skip-connection structure in densely connected convolutional network. Secondly, weight coefficients were added into densely connected convolutional neural network, and each layer in a weighted densely connected convolutional network received all the feature maps generated by its previous layers and was initialized the weight in the skip-connection with different value. Finally, the weight of each layer was dynamically adjusted during training to extract features more effectively. Compared with conventional deep reinforcement learning, in GridWorld simulation experiment, the average reward value of the proposed method was increased by 85.67% under the same number of training steps; in FlappyBird simulation experiment, the average reward value was increased by 55.05%. The experimental results show that the proposed method can achieve better performance in game simulation experiments with different difficulty levels.