A new algorithm for intelligent detection of geohazards incorporating attention mechanism

Abstract

The development of InSAR satellite hardware and data processing technology enables us to rapidly obtain massive high spatial and temporal resolution surface deformation results. The abundant information helps geohazard detection, but also brings challenges for geohazard interpretation. InSAR geohazards intelligent detection methods based on deep learning can greatly improve the efficiency and precision of geohazards interpretation. However, these methods are usually limited by quality of the InSAR result that have various errors, such as decoherence error, topography residue and atmospheric delay. This study proposes a convolutional neural network incorporated an attention mechanism, referred to as DeforNet, which can effectively reduce the influences of the atmospheric delay and topography noises by introducing convolutional block attention module and depth-wise separable convolution. The new method can efficiently and accurately detect small and medium-scale geohazards from InSAR results. The comparison between the DeforNet with FCN, U-net and SegNet, using both synthetic and real samples, show that DeforNet has significant superiority in noise suppression and deformation identification. In the application of the whole Shanxi province, China, the DeforNet detected 1,553 geohazards with the minimum area of 0.21 km2. Our result shows that a strong spatial correlation between the location of geohazards and coal mining in this region. Within the coalfield, the number of identified geohazard accounts for 64.6 % of the total number in Shanxi. We also found that the identification accuracy of DeforNet is affected by the quality of the InSAR results, the scale of the geohazard and the wrap interval. DeforNet can serve to refine the detailed investigation of geohazards and promote the application of InSAR technology in geohazards prevention and mitigation.