GeoAI urban-canopy mapping hits 87.8% LiDAR agreement in Davis

A new preprint from Mohammadreza Narimani, Shreyan Mitra, and Parastoo Farajpoor introduces an integrated optical GeoAI workflow for assessing urban tree canopy, using Davis, California as a testbed. The pipeline starts with 2022 National Agriculture Imagery Program (NAIP) imagery at 0.6 m resolution in RGB and near-infrared bands. DeepForest generates candidate tree crowns, then an NDVI threshold, non-maximum suppression, and a box-prompted Segment Anything Model (ViT-B) refine those into a crown-anchored canopy surface. Within the 25.92 km² municipal boundary and a 100 m grid, the system retained 11,741 candidate crowns and mapped 2.43 km² of canopy, covering 9.37% of the city.

Validation against the 2022 USDA/CAL FIRE LiDAR-assisted canopy product showed 87.8% agreement for mapped pixels and 97.4% for candidate centers, though the optical surface represented only 34.2% of the reference canopy area (IoU 0.288, Dice 0.448). The analysis also found that roughly 49% of candidates sit within 15 meters of a road, and canopy is inversely correlated with Landsat land-surface temperature (Spearman rho = -0.293; partial rho = -0.370 when controlling for built probability). Spatial-lag modeling confirmed neighborhood-level structure. The framework is designed as a transparent, updateable screening layer that complements structural canopy products and municipal inventories, providing reproducible diagnostics for urban heat and planning decisions.