Oblique drone panoramic imagery for fine-scale human-induced land-cover change detection

Abstract

High-Resolution Spatiotemporal Monitoring of Seasonal Land-Use and Land-Cover (LULC) Changes (HRST-MSLC) holds broad application prospects. However, existing unmanned aerial vehicle (UAV) change-detection benchmarks are predominantly based on nadir-view, bi-temporal imagery, providing limited lateral structural information and seasonal observations, while data-acquisition costs and multi-temporal registration constrain dataset construction. To address this, this paper introduces the Unmanned Aerial Vehicle Oblique Multi-View Panoramic Change Detection Dataset (UAV-OMVCD) for fine-scale human-induced LULC change monitoring. It contains 864 panoramic observations acquired in four study areas across Nanjing, Hefei, and Tangshan using DJI Matrice 4D/4TD and Matrice 4E platforms. Each panorama is exported at 1448 × 7200 pixels, with nominal nadir-equivalent source-frame ground sampling distances (GSDs) of approximately 3.7–8.5 cm, yielding 2,181 paired 512 × 512 samples. UAV-OMVCD documents seasonal changes in impervious surfaces, engineering machinery, buildings, greenhouses, and piled soil, including structural additions/removals and morphological evolution. Its oblique views capture lateral facades and occluded areas while reducing seasonal pseudo-change interference. Five benchmark models achieved overall accuracies of 0.9429–0.9492, macro-averaged F1-scores of 0.8861–0.9007, and macro-averaged Intersection over Union (IoU) values of 0.8049–0.8266. UAV-OMVCD provides a standardized benchmark for fine-scale human-induced land-cover change detection using oblique drone panoramic imagery.