Using GeoAI and machine learning tools for consistent high-resolution land cover mapping based on time-series NAIP...

Area of Science:

• Environmental Science
• Remote Sensing
• Geospatial Artificial Intelligence (GeoAI)

Background:

• High-resolution land-cover maps are crucial for ecological and landscape analysis.
• Scarcity of long-term, high-resolution time-series land-cover products hinders research.
• Temporal heterogeneity in National Agricultural Imagery Program (NAIP) data, due to sensor shifts, complicates consistent multi-year mapping.

Purpose of the Study:

• To develop a GeoAI/Machine Learning (ML) algorithm for producing spatially detailed and temporally coherent 1-m land-cover maps using time-series NAIP imagery.
• To overcome challenges posed by cross-year sensor characteristic shifts in NAIP data without requiring historical training samples.

Main Methods:

• A dual-track adaptive workflow was designed to handle NAIP imagery of varying quality.
• High-quality NAIP imagery (2009-2017) was classified using a foundation model (U-Net/ResNet-34) refined by a Segment Anything Model (SAM).
• Lower-quality imagery (2004-2008) was reconstructed via an NLCD-based spatiotemporal bridging and label back-casting pipeline.

Main Results:

• The algorithm achieved stable, consistent results across years with high overall accuracies (e.g., 0.887 in 2014, 0.733 in 2004).
• High F1-scores were consistently observed for Structure, Water, Wetland, and Cropland classes.
• The algorithm maintained spatial fidelity and temporal consistency despite substantial sensor differences in historical imagery.

Conclusions:

• GeoAI/ML tools and diverse data sources can effectively generate consistent, high-resolution, multi-decade land-cover maps from NAIP imagery.
• The developed approach offers a scalable solution for generating time-series land-cover maps across the USA.
• This supports enhanced long-term land-change analyses and landscape-level planning.