An SSM enhanced ViT model for aerial imagery based disaster risk inspection

Highway construction in high-altitude mountain regions of western China faces significant challenges, including high costs and vulnerability to natural disasters such as debris flows, flash floods, barrier lake eruptions and landslides. Since most of these regions are inaccessible to personnel for field investigations due to logistical and geographical constraints, real-time drone imaging and terrain change analysis are crucial for ensuring construction safety. Intelligent inspection drones achieve dynamic monitoring of regional disasters by regularly taking photos of hazardous areas near construction sites and conducting intelligent image analysis, establishing an early warning system of natural disasters. Due to the complex conditions and varying scales of the risk areas in these images, developing robust machine learning algorithms for the semantic segmentation of disaster risk areas and quantifying the changes of the risk states are the keys to this warning system. Vision transformers have shown remarkable success in the tasks of image recognition and segmentation. Recent advancements in state space models, particularly neural ordinary differential equation (NeuralODE) algorithms, address more computational efficiency in long-sequence processing while enhancing image analysis capabilities through their global attractor properties. This paper proposes NeViT, a NeuralODE-enhanced vision transformer model, designed for natural disaster image classification and segmentation. The model integrates a NeuralODE into Swin transformer blocks to establish a dynamic computation framework for feature extraction in aerial imagery. By enabling precise multi-temporal object segmentation, NeViT provides foundational data for subsequent terrain variation analysis, supporting disaster monitoring and risk assessment efforts. Experimental results on two datasets validate the model’s effectiveness, as evidenced by its high accuracy and intersection-over-union (IoU) scores, providing effective algorithmic solutions for risk warning system.