Geographic Prior Guided Subpixel Mapping for Fine-Grained Urban Tree Cover Reconstruction

Abstract

Benefiting from long-term time series and large spatial coverage, Sentinel-2 has been widely used in urban tree cover retrieval. However, the mixed pixel effects in Sentinel-2 imagery make it challenging to accurately identify urban tree covers. To address this problem, subpixel mapping (SPM) is developed to reconstruct a high-resolution urban tree cover from medium-resolution imagery. While deep-learning-based SPM seeks fine-grained patterns solely within medium-resolution feature spaces and spatiotemporal fusion-based SPM leverages additional high-resolution imagery from different times at the same location, both face limitations: the former lacks detailed spatial constraints, and the latter struggles with acquiring geographically aligned imagery. To address these challenges, this study proposes a geographic prior guided SPM (GPSPM) approach for urban tree cover reconstruction. The geographic prior is grounded in the scaling law of geography, a fundamental principle of spatial heterogeneity stating that high-resolution imagery contains far more detailed features (e.g., small tree parcels) than lower resolution imagery. These fine-grained features enhance SPM by providing robust cross-scale spatial prior based on a “teacher-student” domain adaptation training framework. Besides, considering the geometric feature discrepancy and long-tail distribution exists across different geographic scales, cross-scale image mosaicking and resampling strategy are further developed. Experiments on public urban tree cover dataset demonstrate that the proposed method improves the intersection over union (IoU) of urban tree cover by approximately 5% compared to traditional unsupervised SPM and shows significant improvements in spatial detail quality.