In response to the identification problem of debris flow gullies in mountainous and canyon areas, a technologically advanced solution, namely the Two-Channel Residual Network (TCRNet) was proposed in this paper. This network uses DEM data and remote sensing data as network inputs and adopts a novel dual-channel architecture to extract spatial and spectral features respectively. To further optimize network performance, the network embeds the ECA mechanism to emphasize image feature information and adds a residual structure improved by global average pooling to output more information within the receptive field. The prediction results are evaluated for accuracy using the confusion matrix, and evaluation metrics such as precision and recall are calculated for model evaluation. Experimental results show that using deep convolutional neural networks to train DEM and remote sensing data of gullies in Nujiang Prefecture can achieve an 80% recognition rate, 0.79 recall rate, and 0.83 precision rate for debris flow gullies, indicating good model performance. In this study, the saved optimal parameters of the model were also used to evaluate the danger level of 672 gullies in Nujiang Prefecture and obtained four danger level zones: high, medium, low, and extremely low, which were visualized using ArcGIS software. These experimental results demonstrate that it is feasible to use deep convolutional neural networks to extract gully image features for rapid identification of debris flow gullies, providing important references.
|