CoWNet: A correlation weighted network for geological hazard detection

Abstract

The geological hazard of landslide or debris flow is one of the most widespread causes of mortality every year. Segmentation for them has been an effective tool to reduce the number of casualties since its critical contribution to rescue planning. However, existing segmentation methodologies are mainly designed for natural imagery in which most objects have similar shapes and sizes. In contrast to them, landslides and debris flow usually take irregular contour and arbitrary scope, causing most existing methodologies to suffer from low performance. Therefore, it is imperative to design a novel segmentation method that is specific to geological hazard auto-detection. To address these issues, we propose a correlation weighted network (CoWNet) to precisely segment the hazard by concentrating on the negative impact of irregular contour and arbitrary scope. The CoWNet gains profits from two elaborately designed modules, named correlation weighted module (CoW) and feature fusion module (FFM), respectively. Firstly, in contrast to the common preference to blindly expand the receptive field of existing methods, the CoW module chooses to adjust the receptive field based on disaster scope adaptively. The CoW module enables the receptive field of the CoWNet to encompass any scope of hazards. Meanwhile, another module, FFM, is used to relieve the adverse effect brought by the irregular contour problem. The core idea of FFM is to utilize sophisticated semantic context to preliminarily locate the overall disaster and detailed local cues to refine disaster edges. Through the above strategies, the CoWNet could precisely segment any shape and scope of landslides or debris flow. To verify the superiority of the CoWNet, we compare our network with several state-of-art segmentation networks in a high-resolution geological hazard dataset, including landslide and debris flow disasters. The results demonstrate our model outperforms other state-of-the-art models.