Streamlining High-Resolution NAIP Land-Cover Classification Using Automated Preprocessing and Building Footprint Integration
Presentation Time: Fri, 05/01/2026 - 10:00
Keywords: NAIP imagery, land-cover classification, building footprints, feature engineering, ArcGIS Pro
High-resolution land-cover classification using imagery from the National Agriculture Imagery Program (NAIP) presents challenges due to high spatial resolution and resulting spectral heterogeneity. These conditions produce class confusion, particularly for impervious surfaces such as buildings. This study evaluated classification workflows in ArcGIS Pro across four study areas, representing both 30-centimeter and 60-centimeter spatial resolutions, to examine whether automated preprocessing and building footprint integration improve classification performance. Two classification strategies were compared: a baseline workflow, using Red, Green, Blue, and Near-Infrared bands, and an engineered raster stack incorporating spectral indices and texture measures generated through an automated ArcPy preprocessing tool. Building footprint data was integrated as a post-classification step to improve classification accuracy of impervious surfaces. Classification performance was evaluated using spatially independent validation samples and confusion matrices reporting overall (OA), producer’s (PA), and user’s accuracy (UA). Results varied among sites. At the 30-centimeter sites, the engineered workflow and building footprint integration produced minimal change in OA, indicating that baseline spectral information already captured sufficient class separability. At one 60-centimeter site, the baseline workflow performed poorly, with 19.05% OA, while the engineered workflow increased accuracy to 83.33%. Building footprint integration further improved accuracy to 85.71% and increased impervious PA to 100%. At the second 60-centimeter site, the engineered workflow showed only modest improvement. Across all sites, impervious surfaces remained the most difficult class to detect. These results demonstrate that automated preprocessing and building footprint integration can improve classification robustness under varying image conditions, but their effectiveness depends on site-specific imagery characteristics.
