Urban Impervious Surface Age Estimation from Historical Nighttime-Light and Optical Imagery Using an EfficientViT-BiFPN Model

Source abstract

Estimating the age of urban impervious surfaces supports reconstruction of urban expansion, infrastructure screening, and drainage and heat-risk assessment, yet reliable age maps remain difficult to obtain because historical optical coverage is incomplete and nighttime-light sensors are inconsistent. This study proposes an EfficientViT-BiFPN framework that combines annual nighttime-light trajectories, optical spectral indices, and morphology-sensitive texture features at parcel- grid scale. A lightweight EfficientViT encoder captures local and regional context, while a bidirectional feature pyramid preserves weak temporal evidence and fine boundaries associated with construction and redevelopment. A controlled synthetic benchmark representing 38 years of observation and 52,416 labeled grid samples across six urban-zone strata was used to examine accuracy, transfer behavior, ablation response, uncertainty, and computational efficiency. The proposed model achieved 90.8% age-bin accuracy, a mean absolute error of 2.7 years, and a root mean squared error of 4.3 years, with an inference latency of 18.6 ms per 512 by 512 tile. Removing nighttime-light descriptors increased late- stage confusion from 7.9% to 14.0%, whereas removing BiFPN reduced macro F1 from 0.884 to 0.846. These results indicate that compact transformer-pyramid fusion can integrate temporal onset and spatial appearance efficiently under controlled benchmark conditions. Validation against independently dated parcels and infrastructure archives is still required before the estimates can be interpreted as operational construction-age products.