A Tutorial on GeoAI: Designing Footprint Extraction from NAIP Imagery Using U-Net, Grounding DINO, SAM, and Mask R-CNN

In this tutorial, we design a complete GeoAI workflow for extracting building footprints from high-resolution NAIP aerial imagery. We begin by configuring the geospatial deep learning environment, downloading raster imagery and vector labels, and inspecting their spatial properties before generating georeferenced image chips and segmentation masks. We then train a U-Net model with a ResNet-34 encoder, evaluate its learning behavior, and apply sliding-window inference to an unseen scene. Beyond semantic segmentation, we convert predicted masks into cleaned and regularized building polygons, calculate IoU and F1 metrics, explore zero-shot segmentation with Grounding DINO and SAM, and compare the results with a pretrained Mask R-CNN instance segmentation model. We also demonstrate how the same pipeline extends to real-world areas using NAIP imagery from Microsoft Planetary Computer and building labels from Overture Maps.

We configure the environment, install the required GeoAI and deep learning libraries, and verify GPU availability. We define the central configuration parameters for dataset creation, model training, inference, and optional processing stages. We then create the working directory, define reusable execution utilities, and download the NAIP imagery and building footprint labels.

We inspect the raster and vector datasets to understand their coordinate systems, dimensions, statistics, and feature structures. We visualize the building labels over the aerial imagery and generate an interactive map for spatial exploration. We then divide the source imagery into overlapping georeferenced chips and create matching raster masks for model training.

We train a U-Net semantic segmentation model with a ResNet-34 encoder using the prepared image and mask tiles. We configure the training process with validation splitting, early stopping, checkpoint saving, and performance monitoring. We then load the training history, plot the learning curves, and identify the epoch that produces the highest validation IoU.

We apply sliding-window inference to an unseen NAIP scene and generate both prediction and probability rasters. We remove small noisy regions, convert the predicted mask into vector polygons, and regularize the footprint geometries to produce cleaner building boundaries. We also calculate geometric properties and compare the raw polygonized results with the orthogonalized and regularized outputs.

We evaluate the segmentation model by comparing its predictions with rasterized ground-truth building labels. We calculate pixel-level IoU and F1 metrics and visualize the imagery, predictions, and reference masks together. We then apply Grounding DINO and SAM to perform zero-shot building segmentation using text prompts without additional model training.