Evidence-Calibrated Urban Scene Quality Index for High-Resolution Overhead Imagery in Changsha

Abstract

High-resolution overhead imagery captures building arrangement, street structure, vegetation cover, industrial land use, water surfaces, and outdoor activity space in a resolution appropriate for neighborhood analysis. However, semantic classification of such imagery is reduced to an ordinal score by counting favorable criteria, despite the difference in recognition reliability between visual cues and the possibility of moving from one class to another by changing one indicator alone. An Evidence-Calibrated Urban Scene Quality Index (ECUSQI) was devised to convert five semantically identifiable visual indicators into a reliability-based five-point scale. We analyzed a test set of 3038 labeled samples of urban imagery patches extracted from central Changsha (China) among 3874 RGB patches of 250 × 250 pixels with 0.5 m ground sampling distance (GSD). Recognition reliability of 615 test scenes with respect to open building layout, grid-like street structure, vegetation coverage, lack of industrial areas, and presence of activity space was estimated in the training procedure. Jeffreys smoothing of recognition reliability normalizes each indicator increase, the posterior uncertainty component identifies scores based on less reliable semantic information, and the threshold margin term points out classes determined by a less confident inference. The test dataset includes 527 scenes with five correct decisions, 77 with four, 9 with three, and 2 with two, implying an average of 4.836 correct and 0.164 incorrect indicator interpretations per scene. The accuracies of indicators range from 93.17% for buildings to 98.86% for industrial areas. Co-attention reduces the expectation of indicator mistakes by 58.0% and decreases the multi-mistake probability by 8.48% to 1.79%. In the spatial interpretation, ECUSQI is lower in more densely populated districts of older construction and higher in green residential areas with open structure and activity space. Our index serves a concrete measuring purpose, since reliable overhead semantic information can inform fine-grained environmental evaluation, along with its threshold sensitivity.