Debris Flow Susceptibility Assessment In Nujiang Prefecture Based On Convolutional Neural Network

Debris flows occur suddenly and pose a great danger to the personal safety of residents and public property.Coupled with the extreme climate change in recent years,frequent heavy precipitation is more likely to induce debris flows.Therefore,it is necessary to conduct wide-range debris flow susceptibility mapping so as to better promote disaster prevention and mitigation.Statistical-based methods and numerical simulation-based methods have been widely used in debris flow susceptibility elevation.However,the statistical based methods inevitably have subjective selection of disaster-causing factors,and the numerical simulation-based methods often require detailed data to achieve better simulation results which are not always available.Thus,both of these methods are not convenient for large-area susceptibility assessment.To advance the assessment of debris flow susceptibility over large areas,this study propose to combine convolutional neural networks(CNNs)with various types of geological data for susceptibility assessment.The main work of the study is as follows:(1)A multi-source debris flow dataset was constructed for Nujiang Prefecture.By using Arc GIS,the digital elevation data,remote sensing data,soil,lithology,vegetation,and rainfall were extracted.The dataset was expanded using image enhancement methods including rotation and flipping to accommodate the training requirements of the neural network.(2)A special CNN model was designed for debris flow susceptibility mapping by using heterogeneous geological data.By referring to residual structure,dense connection,channel shuffle,etc.,different feature extraction modules are built for various types of data.The experiment results showed that the designed network not only has much smaller parameters than existing conventional CNN models,but also has better performance.Compared with other machine learning methods,the CNN model has a better feature capture ability.(3)The debris flow susceptibility assessment was conducted for Nujiang Prefecture.The elevation results are highly consistent with the historical debris flow records and are more reasonable than the previous studies,indicating that the CNN is capable of conducting the broad scale debris flow susceptibility elevation task.Based on the elevation results,the debris flow susceptibility along the Dulong,Lancang and Nu Rivers is relatively high.(4)The mid-level features of the network were visualized.Through the visualization,it is found that the neural network can capture features such as aspect,slope,vegetation cover,etc.These results further validates the feature extraction ability of CNNs and provides interpretability of the model to some extent.