Two-stage Segmentation Pipeline for Satellite Parking Lot Paved Area Estimation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Background. Accurately estimating paved parking areas is vital for urban planning and environmental monitoring. However, spectral similarities between parking lot and road asphalt make single-stage segmentation unreliable. Existing methods often depend on manually corrected labels. This study introduces a novel pipeline combining free crowd-sourced annotations for coarse localization with human annotations for detailed surface classification. Objective. The aim was to build a two-stage segmentation pipeline wherein the first stage involves localising with boundaries at a coarse level while the second stage is dedicated to segmenting at a fine-grained level of non-paved regions with utilising different sources of supervision for both stages. Methods. Stage 1 uses a U-Net with EfficientNet-B3 trained on 3019 raw OpenStreetMap polygons from 34 US states for boundary detection. Stage 2 masks the image to the detected boundary and trains a second U-Net on 1144 human-annotated masks to segment non-paved regions. Paved area equals boundary area minus non-paved area. Both stages use Dice loss, evaluated by Binary IoU (IoU) at 512x512. Results. Stage 1 delivers an IoU of 0.7886 even without manual correction of the labels. Stage 2 delivers an IoU of 0.6147 and Dice coefficient of 0.8396 for segmentation of the non-paved regions. The difference between the metrics confirms that errors are mainly at boundaries rather than misclassifications on a large scale. The whole pipeline effectively isolates parking lots from their adjacent roads and is also able to locate major non-paved regions. Conclusions. Compared with Qiam et al. [10], we reduced human annotation by 89% and model size by ~18x. The two-stage design combining free OSM labels for coarse localisation with human annotations for fine-grained classification proves viable for paved area estimation. The paradigm extends to related problems such as building material typing and road surface mapping.