STRoadSet, STLandSet, and DHP-Set: Datasets for Post-Disaster Emergency Navigation

This record provides three openly released datasets supporting the study "Generative Disaster Simulation and Deep Heuristic Hierarchical Path Replanning for Post-Disaster Emergency Navigation Using High-Resolution Remote Sensing Imagery," submitted to GIScience & Remote Sensing. All datasets are constructed over a study area centred on the city of Ya'an in the Sichuan–Tibet region of southwestern China. STRoadSet (Sichuan–Tibet Road extraction Set) provides a region-wide road annotation layer. Source imagery is drawn from the Tianditu national geospatial service at zoom level 17, with a spatial resolution of approximately 1.2 m/pixel. Road labels are produced by rasterising the official Tianditu road vector layer into a binary mask. After tiling, the dataset comprises 30,687 image–label pairs of size 256×256. STLandSet (Sichuan–Tibet Land cover Set) provides a land-cover annotation layer registered to the same base imagery as STRoadSet. Labels are derived from the GlobeLand30 land-cover product, following its ten-class taxonomy (seven classes are present in the study area). After tiling, the dataset comprises 1,802 image–label pairs of size 256×256. DHP-Set (Deep Heuristic Perception Set) comprises 45,643 paired samples for training a deep heuristic potential field. Each sample contains a five-channel input tensor and a single-channel ground-truth cost-to-go potential field, both of size 256×256, packaged as compressed .npz files. Ground-truth labels are generated through a two-stage reverse Dijkstra labelling scheme that disentangles static terrain priors from dynamic disaster perturbations. The datasets are released under the Creative Commons Attribution 4.0 (CC-BY-4.0) licence.