Search

With the development of artificial intelligence, the most representative deep learning has been applied to various fields, which is greatly influencing human society. However, deep neural networks (DNNs) are still a black-box model, and the process how they make decisions internally is still difficult to understand and control. At the same time, DNNs take up more hardware resources, resulting in high energy consumption. Therefore, it is significant to study the characteristics of deep AI models and deeply understand the interactions between parameters within AI models so as to improve the interpretability of DNNs, optimize their structure and increase their computational efficiency. In this paper, we propose a graph-based interpretability method for deep neural networks (GIMDNN). The running parameters of DNNs are modeled as a graph by using a kernel function or the Graph Transformer Networks (GTN), where the nodes of the graph are obtained by dimensional mapping of the parameters of the DNNs, and the edges are calculated by the Gaussian kernel function. The generated graphs are classified by a graph convolutional network (GCN). The association relationship between the adjacent layers and the running mechanism of DNNs are analyzed, and the importance of the parameters of each layer in the DNNs for the final classification result can be obtained. Convolutional neural networks (CNNs) are one of the most representative models in DNNs. The proposed method is experimentally evaluated on the CNNs. The experimental results show that the proposed method can interpret the associations among the weight parameters as well as the correlation between two adjacent layers. Therefore, the DNNs for special tasks, such as portable applications, edge computing, and so on, can be customized, the number of parameters can be reduced. It is valuable to interpret the operation and principle of CNNs. [ABSTRACT FROM AUTHOR]

The speckle noise found in synthetic aperture radar (SAR) images severely affects the efficiency of image interpretation, retrieval and other applications. Thus, effective methods for despeckling SAR image are required. The traditional methods for SAR image despeckling fail to balance in terms of the relationship between the intensity of speckle noise filtering and the retention of texture details. Deep learning based SAR image despeckling methods have been shown to have the potential to achieve this balance. Therefore, this study proposes a self-attention multi-scale convolution neural network (SAMSCNN) method for SAR image despeckling. The advantage of the SAMSCNN method is that it considers both multi-scale feature extraction and channel attention mechanisms for multi-scale fused features. In the SAMSCNN method, multi-scale features are extracted from SAR images through convolution layers with different depths. These are concatenated; then, and an attention mechanism is introduced to assign different weights to features of different scales, obtaining multi-scale fused features with weights. Finally, the despeckled SAR image is generated through global residual noise reduction and image structure fine-tuning. The despeckling experiments in this study involved a variety of scenes using simulated and real data. The performance of the proposed model was analysed using quantitative and qualitative evaluation methods and compared to probabilistic patch-based (PPB), SAR block-matching 3-D (SAR-BM3D) and SAR-CNN methods. The experimental results show that the method proposed in this paper improves the objective indexes and shows great advantages in visual effects compared to these classical methods. The method proposed in this study can provide key technical support for the practical application of SAR images. [ABSTRACT FROM AUTHOR]

• MLP, GRU, CNN and MSCNN for landslide susceptibility mapping were compared. • CNN combined with multi-scale technique can improve feature utilization. • The joint evaluation method of ROC curve and PR curve for LSM was proposed. Landslide susceptibility mapping (LSM) can be used to determine the spatial probability of landslide occurrence. There are many methods for LSM, including statistical methods, traditional machine learning methods and deep learning methods, etc. However, the difference comparison of these methods has been not perfect, especially the comparison of different neural network models for LSM and their application prospects were rarely studied. In this paper, the classical neural net-work multi-layer perceptron (MLP), convolutional neural network (CNN), gated recurrent unit (GRU) and multi-scale convolutional neural network (MSCNN) four models are selected for comparison. Taking Lanzhou city, Gansu Province, China as an example, eight landslide-related influencing factors and historical landslide and non-landslide locations were selected, and the training set and validation set were divided according to 7:3. Through training the four models, four landslide susceptibility maps were generated. The experimental results were verified and compared by the confusion matrix, Kappa coefficient, F1-score and other statistical indicators. The receiver operating characteristic (ROC) curve and Precision-Recall (PR) curve were plotted to evaluate the classification effect and generalization capability of four models. The results show that the constructed MSCNN is the optimal model, which has the best performance both in the training process and in the mapping results. MSCNN model has the highest value of Recall (99.93%), Kappa (0.96) and F1-score (0.98) in the confusion matrix. In addition, ROC curve and PR curve of MSCNN model maintain the maximum area under curve (AUC) on different data sets. In the comparison, MLP and GRU accept sequence features, while CNN and MSCNN accept neighborhood features. In general, the prediction model considering neighborhood features contains more information in the limited input data and is better than the prediction model considering sequence features in all evaluation indicators. Therefore, we think that the neighborhood features can better represent the landslide occurrence characteristics. In the future model design process for LSM, more attention should be paid to the neighborhood features of landslide influencing factors. [ABSTRACT FROM AUTHOR]